Instead of trying to produce a program to simulate the adult mind, why not rather try to produce one which simulates the child’s?

— Alan Turing

The human brain is biology’s greatest achievement. Despite tendencies to fall into heuristic biases, group-think, and occasional irrationality, it remains the most advanced reasoning system of all time (of course, one could argue that some other mammals — like killer whales — rival us in intelligence, but I digress. They didn’t build the pyramids).

The goal of artificial intelligence has always been to model, replicate, and implement the processes of the human brain. From the earliest days of computing until now, engineers, scientists, and researchers have sought to create artificial systems of advanced reasoning, pattern matching, and thought that rival what biology has taken millions of years of evolution to produce.

The first artificial neuron, created by Walter McCulloch and Walter Pitts in 1943, was built on the conviction that simple binary processing units could replicate brain function. But while neural networks have made extraordinary progress, they have strayed from biology, evolving into complex mathematical systems that may have little in common with how real cognition works.

Geoffrey Hinton’s 1986 paper on backpropagation laid the foundation for modern AI, but backpropagation is not how the brain learns. It is a mathematical trick for optimizing statistical models, not a biological process of adaptation and reasoning.

A new class of emerging AI frameworks is beginning to rethink this fundamental approach. Instead of merely optimizing mathematical weights, these frameworks aim to build AI that processes, learns, and adapts more like the human brain itself.

One of the most promising of these frameworks is Neuromorphic Computing. In this deep dive, we’ll explore:

• The fundamental pitfalls of classical artificial neural networks (ANNs)
• How neuromorphic computing proposes a radically different model of intelligence
• Why rethinking AI’s foundations could help us overcome its biggest limitations

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Artificial Neural Networks

While inspired by neurons, artificial neural networks, or ANNs, function fundamentally differently than the human brain. The brain is an event-based, contextually dependent learning machine.

ANNs, in contrast, operate on ever complexly determined static weight models built on dense matric multiplications and predefined learning rules.

Unlike biological systems, ANNs require fixed structures, leveraging vast amounts of computation and data to feed a learning system based on gradient descent, backpropagation, and error minimization.

Core Structure

Artificial neural networks (ANNs) are made up of layers of artificial neurons, each playing a role in processing information.

• Input Layer: Takes in raw data and converts it into numerical vectors.
• Hidden Layers: Contain hundreds to millions of neurons, which process the data by applying weighted sums and activation functions to detect patterns.
• Output Layer: Produces the final result based on the processed information.

Each neuron transforms and passes information forward, allowing the network to gradually refine its understanding of the input.

To learn, ANNs use a repeated system of error minimization to adjust weights and improve prediction accuracy over training cycles.

During training, the network’s final output is compared to the true labels (the correct answers) using loss functions. Through advanced matrix multiplications (where GPUs play a crucial role), the network calculates how much each weight within the hidden layers contributes to the error by computing the gradient loss associated with each neuron.

Think of gradient descent as finding the lowest point in a valley, where said lowest point represents the point at which the sum of the errors across all neurons is minimized.

Repeated over thousands to millions of training cycles, this process produces a highly competent model.

Real World Examples

ANNs dominate the AI market as it stands today, coming in various shapes and sizes — each of which best for specific use cases

• Convolutional Neural Networks (CNNs) use convolutional layers to scan images with filters, extracting specific features like edges, shapes, and textures to make sense of image-based data. These are heavily relied on for computer vision.
• Reinforcement learning-based ANNs use reward-based learning to optimize behavior over time and have emerged as the leading training structure for robotic systems
• Transformer models power the LLM tools we all use, and use self-attention mechanisms to process long sequences of text and other data types.

But…

• CNNs process every pixel of every image independently, even when most contain no useful information. This is a massive waste of computational resources and energy, requiring constant redundant computations across video frames before the model becomes “good.” It’s why we saw Waymo’s self-driving cars with human operators for years before they were ready for deployment.
• Reinforcement learning often requires millions of simulations before meaningful improvements emerge. Even then, generalization and adaptation remain major challenges — robots struggle in environments they weren’t explicitly trained on.(Figure just made a huge breakthrough here, but I’m not 100% sure how they did it. Will do a deep dive at some point in the future).
• Transformers demand billions of parameters, staggering amounts of training data, and immense compute power, making them one of the most resource-intensive AI architectures in existence.

Speaking on Limitations in General

These frameworks are dominant for a reason — when done well, they’re incredibly powerful.

However, their fundamental design is inherently inefficient at scale. As AI is deployed more widely, these inefficiencies become major bottlenecks. Bottlenecks that other solutions may be better suited to address.

The Von Neumann Bottleneck

ANN processors separate memory (RAM) from computation (CPUs and GPUs. Every computation carried out by the system requires data to be fetched from memory, processed, and transferred back to update the model. This shuffling not only results in increased latency, but massive power consumption, leading to significant energy waste..

While GPUs attempt to mitigate this with high-bandwidth memory (HBM), the energy cost remains largely unchanged. GPT-3, a legacy OpenAI model, consumed enough power to run an entire town for a month.

Even on the edge, where power efficiency is a priority, current hardware struggles to balance power demand and memory life. The fundamental constraints of today’s architectures limit AI’s scalability and efficiency, requiring a rethinking of how these systems are built.

No Adaptation Without Retraining

As discussed earlier, neural networks learn through error minimization on ingested data. This adjustment of weights, unfortunately, can only occur during training. Once a model is set loose, it cannot be changed unless retrained.

In real-world use cases, this leads to several potential issues:

• Catastrophic Forgetting: ANNs optimize the network for the most recent task it was trained on. If new data is introduced, the model overwrites all previously learned knowledge. Therefore, the only way to truly “update” a model without destroying it and starting anew is to ingest new data alongside all original data. Obviously, this drastically increases the amount of data required to maintain and improve a model’s performance. Training to perfection becomes a moving target, and every update to the model requires more total data ingestion, making scaling incredibly costly.
• No real-time learning: Forget about on-the-fly adaptation. These models cannot adjust their weights in real-time. It takes retraining. So, if a model encounters data too different from what it was trained on, outright failure becomes a real and significant possibility. A robot trained in one factory might struggle in another, even with minute environmental changes. The only way to address this is through massive pretraining on highly diverse datasets, which leads us back into the VN bottleneck… burning unnecessary compute resources and energy to prepare for real-life unpredictability.

Probabilistic Guessing vs. True Responsiveness

ANNs, regardless of how advanced they may seem, do not understand data. They’re probabilistic models — incredibly good guessers, but guessers all the same.

Why is this an issue?

• Predictions rely on predetermined, precomputed weights rather than real-time adaptation. ANNs don’t update their understanding of data as it’s processed. They apply static, learned patterns instead of dynamically adjusting to new information.
• They cannot logically reason or deduce cause-and-effect relationships. Our brains adapt to new stimuli on the fly, constantly adjusting “weights” and learning how to respond. ANNs, by contrast, are locked in once trained, making them naturally vulnerable to edge cases and unexpected scenarios.
• Small changes in inputs or training biases can lead to catastrophic miscalculations. Because these models lack contextual awareness, even slight errors in training data can cause major failures. (Remember when Gemini delivered a “historically accurate” image of Black Founding Fathers?)

So, the question is: how do we solve this? What new frameworks fundamentally rethink AI’s architecture enough to mitigate these pitfalls?

One of the most discussed emerging alternatives is neuromorphic computing; a radically different approach designed to overcome these limitations.

First, Looking Towards Biology as a Framework

To understand why neuromorphic computing better replicates biological learning and response architectures, we first need to understand how the brain itself functions. Its structure, how it processes information, how it learns and adapts, and why — despite all our advances in artificial intelligence — it remains the most efficient computational engine of all time.

The human brain contains approximately 86 billion neurons, forming an interconnected system that enables us to perceive, learn, and adapt in real-time. Each neuron acts as a processing unit, constantly receiving, filtering, and transmitting information.

Core Structure

Every neuron is composed of four primary components, each playing a distinct role:

• Dendrites: The Input System
  Branch-like extensions that receive electrical and chemical signals from other neurons. Each dendrite contains receptors that determine how the neuron responds to incoming neurotransmitters. The more frequently a specific signal is received, the stronger the connection becomes — a process called synaptic plasticity, which is fundamental to learning and memory formation.
• Soma: The Decision Maker
  Also known as the cell body, the soma is the processing hub of the neuron. It collects and integrates signals from the dendrites, deciding whether the neuron should fire and pass along the message. If the total input surpasses a certain threshold, the neuron activates, sending an electrical pulse down its axon.
• Axon: The Information Highway
  If the soma fires, an electrical impulse travels down the axon, a long, fiber-like structure that transmits signals to other neurons. The axon can stretch from a fraction of a millimeter to over a meter long, allowing rapid long-distance communication within the brain and body.
• Synapses: The Connectors
  At the axon’s end, signals reach the synapse — a tiny gap between neurons where communication happens. Here, axon terminals release neurotransmitters, which bind to receptors on the next neuron’s dendrites, determining whether the signal continues. The strength of these connections evolves over time, forming the basis of learning, adaptation, and memory retention.

Information Processing

Neurons operate through a binary communication system. Unlike artificial neurons, which continuously process and transmit information through weighted matrix calculations, biological neurons have only two states: firing or not firing.

Each neuron in the brain is constantly bombarded with excitatory and inhibitory signals from other neurons through its dendrites. These incoming signals slightly raise or lower the neuron’s internal voltage, dynamically shifting its readiness to fire.

If enough excitatory signals arrive in quick succession and push the neuron’s voltage past its activation threshold, it fires, sending a spike down its axon to communicate with other neurons. If inhibitory signals outweigh excitatory ones, or if the total voltage fails to reach the threshold, the neuron remains inactive, filtering out noise and preventing unnecessary firing.

This threshold-based activation system ensures reliable and efficient communication, preventing minor fluctuations in input from generating random or misleading signals.

Unlike artificial neural networks, where the strength of a signal is determined by numerical weight values, biological neurons encode information through:

• Spike frequency (rate coding): The number of spikes over a given period indicates the intensity of the signal.
• Spike timing (temporal coding): The precise timing of spikes relative to others conveys additional meaning, allowing the brain to process complex patterns efficiently.

How this powers learning

Neurons allow us to learn through three primary methods:

• Hebbian learning: “Neurons that fire together, wire together.” When two neurons fire in close succession, it signals that they’re reacting to similar stimuli and should strengthen their connection. For example, neurons responsible for detecting eyes, noses, and mouths become more tightly linked through repeated exposure — making facial recognition faster and more intuitive.
• Spike timing-dependent plasticity (STDP): If one neuron consistently fires right before another, this is a suggestion of casual relationships, and the connection strengthens. If the order reverses, and the second neuron fires before the first, the connection weakens, preventing false associations.
• Neuromodulation: The brain’s built-in reward system, powered by neurotransmitters like dopamine, serotonin, and acetylcholine. When we complete a task, a dopamine release strengthens the neural pathways responsible for that success. This creates reinforcement loops that help us learn from trial and error — essentially hardwiring “good behavior” into our neural circuits.

So What?

The brain is better. ANNs may have borrowed the term “neural”, but their design isn’t even close to how we know the brain works. Why not build a system that does?

Neuromorphic Computing

Neuromorphic computing leverages an entirely new approach — spiking neural networks — a stark movement away from traditional ANN models.

Unlike ANNs, which rely on dense matrix multiplications, fixed weight assignments, and back-propagation, neuromorphic systems process information through discrete spikes of activity — more closely mirroring how biological neurons communicate. Instead of continuous activation, neurons in an SNN fire only when necessary, and weights are updated locally based on the timing and frequency of these spikes.

The core principles of neuromorphic computing are as follows:

• Spikes: Similar to our brains, neurons in a spiking neural network only fire when a certain threshold is reached, unlike the continuous activation that ANNs leverage.
• Event-Driven Processing: Traditional ANNs process all data constantly, even if no useful information is present. Neuromorphic, by way of the spiking properties, only processes relevant events.
• Local Learning: SNNs leverage Hebbian learning and spike-timing-dependent plasticity to strengthen relationships between neurons, allowing them to learn from sensory inputs as they are ingested.
• In Memory Computing: Unlike ANNs, SNNs do not separate memory and computing. Rather, neuromorphic chips integrate memory directly into their synaptic elements, maintaining weights without any data transfer, subverting the Von Neumann bottleneck.

How Spiking Neural Networks Process Information

Neuromorphic computing is built on a simple but powerful principle: neurons should only process information when necessary. Unlike artificial neural networks (ANNs), which continuously pass numerical values through layers of matrix multiplications, spiking neural networks (SNNs) use discrete, event-driven signals called spikes to transmit information.

But what exactly is a spike? How do they emerge, and what determines when a neuron fires?

In an SNN, neurons don’t just ingest all available data at once like ANNs do. Instead, they respond only to meaningful changes in input, similar to how our own brains focus on new or relevant sensory information rather than processing every single visual detail at all times.

Take vision, for example. When you stare at a still object, your brain doesn’t waste energy continuously reprocessing its features — unless something changes, like a sudden movement or shift in lighting. The same principle applies in an SNN: neurons only fire when the input signal changes in a meaningful way. This makes them far more computationally efficient than traditional AI models, which burn energy on unnecessary calculations.

How Spikes Are Triggered in an SNN

Spikes in an SNN are not constant activations, but discrete electrical pulses that occur only when a neuron reaches its activation threshold.

1. Ingesting External Data
   Every neuromorphic system begins with sensory input — whether it’s a camera detecting light, a microphone processing sound, or another type of data stream. Unlike ANNs, which process entire datasets in massive computational bursts, SNNs continuously monitor their input streams but remain inactive until a meaningful change occurs.
2. Accumulating Charge (Membrane Potential)
   Each neuron in an SNN receives signals from other neurons or direct sensory input. These signals take one of two forms: excitatory signals, which increase the neuron’s charge, and inhibitory signals, which decrease it. Every incoming signal contributes to the neuron’s membrane potential, much like a bucket filling with water. If the bucket fills past a certain point, it overflows — causing the neuron to fire.
3. Reaching the Activation Threshold & Firing a Spike
   If enough excitatory signals arrive within a short time frame, the neuron’s charge crosses the activation threshold, triggering a spike. This spike is then transmitted to connected neurons, influencing their behavior. If the neuron doesn’t receive enough stimulation, it remains silent — conserving power and avoiding unnecessary computation.
4. Reset and Refractory Period
   After firing, the neuron resets, entering a brief refractory period where it cannot fire again immediately. This prevents overactivity and noise, ensuring that only truly important signals propagate through the network.

A Real-World Example: Event-Based Vision

To illustrate how SNNs handle information differently from traditional AI, consider neuromorphic vision sensors — a technology already being deployed in robotics, self-driving cars, and industrial automation.

A standard camera captures and processes every frame in full, whether or not the scene has changed. This means enormous amounts of redundant data are processed, leading to unnecessary computation and energy waste.

A neuromorphic camera, on the other hand, only activates pixels that detect change — if nothing moves, nothing is processed. This event-driven approach saves orders of magnitude in power and computation while maintaining real-time responsiveness.

The key takeaway: SNNs process data only when it matters, making them far more efficient than traditional deep learning models, which indiscriminately process every input at all times.

How Spiking Neural Networks Learn

For an AI system to be truly intelligent, it needs to learn from experience. Traditional artificial neural networks achieve this through backpropagation, a brute-force method that fine-tunes every connection in the network based on error calculations across massive datasets. But the human brain doesn’t learn this way, so why should SNNs?

Instead, they learn by dynamically strengthening or weakening connections between neurons based on their activity patterns, just like biological brains. The result? Real-time adaptation without the need for expensive retraining.

1. Synaptic Plasticity: Learning by Strengthening & Weakening Connections

At the core of SNN learning is synaptic plasticity — the idea that the strength of a connection between two neurons is not fixed but instead evolves over time.

Whenever a neuron fires a spike, it sends that signal to connected neurons. But what determines whether those connections grow stronger or weaker?

• If two neurons fire close together in time, their connection strengthens — indicating that the relationship between the two signals is meaningful.
• If one fires without the other, their connection weakens — suggesting that the link may not be relevant.

This self-organizing process allows SNNs to adjust in real-time based on the data they receive, rather than relying on pre-trained weights. This is the core principle behind Hebbian Learning, commonly summarized as:

“Neurons that fire together, wire together.”

Example: Imagine a neuromorphic robotic arm learning to grasp different objects. Each time the arm successfully picks up an object, the connections between the sensory inputs and motor outputs strengthen, making future grasps smoother and more accurate. Over time, the system refines its actions through experience — without ever needing a full retraining cycle.

2. Spike-Timing-Dependent Plasticity (STDP): Learning Through Timing

SNNs don’t just learn which neurons fire together, but also the precise timing of those firings — a process known as Spike-Timing-Dependent Plasticity (STDP).

• If a neuron fires just before another, it suggests a causal relationship, and the connection strengthens.
• If a neuron fires after another, it suggests the relationship is irrelevant, and the connection weakens.

Example: Consider speech recognition. When hearing the word “hello,” neurons associated with processing the “h” sound will fire slightly before those processing the “e”, and so on. Over time, the system builds stronger connections between these sequential spikes, making it more adept at recognizing spoken words — just as humans do when learning a new language.

3. Reinforcement Learning via Neuromodulation

In biological brains, learning isn’t just about association — it’s also about rewards. Our brains release dopamine when we succeed at a task, reinforcing the pathways that led to that success.

SNNs can replicate this mechanism through neuromodulation, where a simulated “reward signal” influences which pathways strengthen or weaken.

Example: A neuromorphic drone navigating a complex environment. Every time it successfully avoids an obstacle, a reinforcement signal strengthens the neural pathways responsible for that movement. This allows the system to actively improve performance in real-time — without an engineer manually adjusting its weights.

Why This Matters: The Advantage Over Traditional ANNs

SNN learning offers major advantages over conventional deep learning models:

1. No Expensive Retraining — Traditional AI models require entirely new datasets and high-powered GPUs to learn new tasks. SNNs adapt instantly as they interact with their environment.
2. Real-Time Learning — Traditional models can only make predictions based on pre-learned patterns. SNNs continuously update their knowledge, making them ideal for dynamic environments.
3. Energy Efficiency — Instead of constantly recalculating every weight in a massive neural network, SNNs adjust only the relevant connections, dramatically reducing compute power.

How Neuromorphic Chips Are Designed: The Hardware That Enables Learning

Understanding how spiking neural networks process information and adapt in real-time is only half the picture. Unlike traditional AI models, which are largely software-based and optimized for GPU or TPU acceleration, neuromorphic computing requires an entirely different hardware architecture — one that is built from the ground up to function more like a brain.

Neuromorphic chips are not simply faster versions of existing processors. They do not rely on brute-force matrix multiplications, nor do they depend on separate memory and compute units like traditional von Neumann architectures. Instead, they integrate memory, computation, and learning into a single, energy-efficient system — allowing AI to run on ultra-low power while continuously adapting to new inputs.

This shift is profound. Rather than simply simulating brain-like processes in software, neuromorphic chips physically replicate the structure and function of biological neural circuits — bringing AI one step closer to true autonomy and efficiency.

How Neuromorphic Chips Store and Process Information

Neuromorphic chips don’t have DRAM, VRAM, or traditional storage hierarchies. Instead, they store information directly within their synaptic elements — creating a system where memory and computation occur in the same location.

This structure allows neuromorphic systems to process data as it arrives rather than batching information into discrete training cycles like traditional AI. It also enables real-time responsiveness without the need for external updates, making it possible for AI to continuously refine its behavior without being retrained from scratch.

Take Intel’s Loihi 2, for example — a state-of-the-art neuromorphic processor designed to replicate biological learning mechanisms. Loihi 2 integrates 1 million artificial neurons and 120 million synapses on a single chip, with each synapse storing its own weight value. This means that rather than pulling weights from memory every time it needs to process information, the chip adjusts and recalls them dynamically, just as the human brain does.

This structure allows neuromorphic chips to function autonomously, without requiring constant software intervention. While a GPU-based AI model needs external retraining to adapt, a neuromorphic system can learn on the fly, adjusting to new inputs in real time — without requiring petabytes of training data or millions of pre-labeled examples.

On-Chip Learning: A Self-Improving System

One of the most radical differences between neuromorphic chips and traditional AI hardware is their ability to learn directly on the chip itself.

Most modern AI models are trained offline on massive datasets, optimized using backpropagation, and then deployed for inference on separate hardware. This means that once an AI model is trained, it is frozen — it cannot adapt to new situations without being retrained from scratch.

Neuromorphic chips eliminate this limitation by embedding learning mechanisms directly into the hardware. Instead of relying on global weight updates through backpropagation, they use local, biologically inspired learning rules — allowing them to adapt continuously without external training.

This is possible thanks to specialized circuit designs that mimic synaptic plasticity, enabling connections between neurons to strengthen or weaken dynamically based on incoming spikes. When a neuromorphic chip detects a useful pattern, it physically modifies its synaptic weights — creating a form of hardware-embedded memory that evolves with experience.

This has profound implications for AI systems that need to operate in unpredictable environments. A neuromorphic-powered robotic arm could learn and refine its grip over time, adjusting to different objects and surfaces without needing new training data. A neuromorphic drone could navigate unknown terrain by dynamically updating its perception model on the fly, rather than relying on pre-programmed rules.

In short, neuromorphic computing enables AI to operate more like a self-learning organism rather than a static program — bringing artificial intelligence closer to human-like adaptability than ever before.

Analog vs. Digital Neuromorphic Chips

Not all neuromorphic hardware is built the same way. There are two major approaches to designing these chips:

Digital Neuromorphic Chips (Intel Loihi, IBM TrueNorth)
Most commercially available neuromorphic chips, like Loihi 2, use digital circuits to simulate spiking activity. These chips are scalable, robust, and more compatible with existing AI systems, making them a practical choice for early adoption.

Analog Neuromorphic Chips (BrainScaleS, SynSense DynapCNN)
Some research labs and specialized companies are exploring analog neuromorphic chips, which use real electrical charge and resistance to model synaptic behavior. These chips consume even less power than their digital counterparts but are harder to manufacture at scale due to inherent variability in analog components.

Both approaches aim for the same goal: hardware that learns, adapts, and processes information in a brain-like way — but the trade-offs between scalability, energy efficiency, and manufacturability will determine which design dominates the future of neuromorphic computing.

Challenges and Opportunities for Neuromorphic Computing

While promising, neuromorphic computing faces significant hurdles before it can achieve widespread adoption. Here’s where the field stands today:

Challenges

🔹 Hardware Scalability — Manufacturing neuromorphic chips at scale remains complex, particularly for analog designs.
🔹 Programming Complexity — Most AI engineers are trained in deep learning, not neuromorphic architectures — new tools and frameworks are needed.
🔹 Lack of Standardization — Unlike GPUs and CPUs, there is no dominant neuromorphic hardware platform yet, leading to fragmentation.
🔹 Industry Adoption Lag — Most AI applications are optimized for traditional ANN architectures, requiring significant rethinking of AI workflows.

Opportunities

✅ AI at the Edge — As demand for edge AI grows, neuromorphic computing is uniquely positioned to power next-gen wearables, IoT, and mobile devices.
✅ Sustainability — The AI industry faces increasing scrutiny over energy use — neuromorphic computing offers a path to high-performance AI at low power.
✅ Next-Gen Robotics — Real-time learning without retraining is a game-changer for autonomous systems, opening the door to AI that truly adapts.
✅ Brain-Like Intelligence — Neuromorphic AI is not just about efficiency — it brings AI one step closer to human-like cognition, learning, and adaptability.

The Road Ahead: A New Paradigm for AI

For decades, artificial intelligence has followed a singular path: bigger models, more data, more compute. That approach has brought us remarkable breakthroughs, from computer vision to large language models. But it’s also running into hard limits — skyrocketing power consumption, rigid learning mechanisms, and a fundamental inability to adapt in real-time.

Neuromorphic computing offers an alternative. Not just an incremental improvement, but a fundamentally different way of thinking about intelligence. Instead of optimizing brute-force computation, it focuses on efficiency, adaptability, and real-time processing — qualities that make biological intelligence so powerful.

If neuromorphic AI reaches its full potential, we could see:
✅ AI that learns like humans — No more frozen models. Systems that evolve and refine themselves without retraining.
✅ Massive reductions in AI power consumption — AI that runs on a fraction of the energy required today, making intelligence sustainable.
✅ Truly autonomous systems — Robotics, self-driving cars, and AI-powered assistants that adapt to their environments in real time.
✅ Seamless human-AI integration — Brain-machine interfaces and neuroprosthetics that think and respond dynamically, rather than relying on static algorithms.

We’re still in the early days. While neuromorphic chips exist, they are nowhere near the level of commercial GPUs in terms of software support and integration. But the shift has started.

Conclusion: The Future of AI Is Brain-Like, Not Brute Force Based

Artificial intelligence was always meant to be modeled after the brain. But in our race for progress, we took a detour — building static, data-hungry models instead of systems that can truly adapt, learn, and reason.

Neuromorphic computing corrects course. Not by abandoning everything we’ve built, but by evolving it — moving AI from rigid architectures into flexible, dynamic intelligence that processes the world the way we do.

It won’t replace traditional AI overnight. But as intelligence moves from the cloud to the edge, from static predictions to real-time reasoning, and from power-hungry clusters to efficient, embedded systems, the demand for neuromorphic computing will only grow.

The AI race won’t be won by those who build the biggest models — but by those who build the smartest ones.

Startups to Watch

BrainChip’s (Acquired) Akida platform is a neuromorphic computing solution designed for ultra-low-power, event-driven AI processing. Unlike traditional AI accelerators, Akida employs spiking neural networks (SNNs) and on-chip learning to enable real-time adaptation without cloud connectivity. Its architecture is optimized for edge applications, such as smart sensors, robotics, and cybersecurity, delivering AI capabilities with minimal power consumption.

iniVation (Acquired) develops neuromorphic vision systems that leverage event-based cameras inspired by biological retinas. Unlike traditional cameras, which process full image frames at fixed intervals, iniVation’s Dynamic Vision Sensors (DVS) only capture changes in a scene, drastically reducing data load, power consumption, and latency. This makes their technology well-suited for robotics, autonomous vehicles, and real-time machine vision applications where efficiency and speed are critical.

Syntiant (Series D, $311.4M raised in funding) designs ultra-low-power AI processors optimized for always-on, edge-based inference, using deep learning and neuromorphic principles to enable efficient speech, audio, and sensor processing. Their hardware accelerates neural network computations with minimal energy consumption, making it ideal for voice activation, wearable devices, and battery-powered AI applications.

Prophesee (Series C, $148M in funding) develops neuromorphic vision sensors inspired by the human eye, leveraging event-based processing to enable ultra-fast, low-power visual perception. Unlike traditional cameras that capture full frames at a fixed rate, Prophesee’s event-driven sensors only react to changes in a scene, dramatically reducing power consumption and computational load for applications in robotics, industrial automation, and AR/VR.

Synsense (Series B, $86.7M in funding) develops ultra-low-power neuromorphic AI chips and event-based vision processing solutions, designed to bring real-time, energy-efficient intelligence to edge devices. By integrating spiking neural networks (SNNs) into compact, high-performance silicon, SynSense enables applications such as always-on smart sensing, autonomous robotics, and ultra-low-power speech and vision processing.

Rain AI (Series A, $67M in funding) develops neuromorphic processing hardware designed to mimic the efficiency and adaptability of the human brain. Their approach leverages analog computation and in-memory processing to enable low-power, high-performance AI inference, particularly for edge applications.

AIStorm (Series B, $29.2M in funding) develops AI-in-Sensor™ technology, which processes data directly at the sensor level using event-driven, neuromorphic-inspired architectures. By eliminating the need for external processing and reducing latency, AIStorm’s approach enhances energy efficiency and real-time decision-making for edge AI applications like smart devices, automotive, and industrial automation.

Innatera (Series A, $27.2M in funding) develops neuromorphic processing units (NPUs) that use spiking neural networks (SNNs) to deliver ultra-low-power AI acceleration for edge applications. Their chips mimic the brain’s event-driven computation, enabling real-time pattern recognition and decision-making in energy-constrained environments such as robotics, wearables, and smart sensors.

Polyn (Series A, $26.M in funding) develops neuromorphic analog signal processing (NASP) chips, designed to perform AI inference directly at the sensor level with ultra-low power consumption. By integrating analog neuromorphic computing into compact edge devices, POLYN enables real-time, event-driven AI for applications like healthcare monitoring, industrial IoT, and always-on sensing.

Aspinity (Series B, $22.4M in funding) develops neuromorphic analog processing chips designed for ultra-low-power edge AI applications. Their Reconfigurable Analog Modular Processor (RAMP™) mimics brain-like event-driven processing by performing analog signal analysis at the sensor level, reducing the need for continuous high-power digital computation. This enables efficient always-on sensing for applications such as voice detection, industrial monitoring, and battery-powered IoT devices.

Opteran (Seed, $15.4M in funding) develops neuromorphic-based autonomy solutions inspired by insect brains, focusing on lightweight, efficient perception and decision-making systems. Their natural intelligence approach enables real-time, low-power navigation and object recognition without the need for large datasets or extensive training, making it ideal for robotics, drones, and autonomous systems.

Cognifiber (Series A, $13.5M in funding) develops fiber-optic-based neuromorphic computing solutions that integrate photonic processing directly into optical fibers, enabling ultra-fast, low-power AI inference. By leveraging light-based computation, their approach aims to overcome traditional electronic bottlenecks, offering real-time AI processing for edge applications without the energy overhead of conventional semiconductor-based chips.

Femtosense (Series A, $13.4M in funding) relevant to neuromorphic computing. The company develops ultra-low-power spiking neural network (SNN) processors designed for real-time AI inference at the edge. Their SparseNet™ architecture processes data efficiently using event-driven computation, significantly reducing power consumption while enabling on-device intelligence for applications like audio processing, wearable devices, and smart sensors.

Natural Intelligence Systems (Seed, $13M in funding) provides a neuromorphic computing system that employs pattern-based machine learning, enabling rapid learning from small datasets while ensuring explainability through methods like unsupervised learning, anomaly detection, and incremental training. This technology aims to deliver actionable insights and improve the interpretability of AI predictions, even when dealing with imperfect data.

Cortical Labs (Seed, $12.7M in funding), is pioneering “DishBrain,” a hybrid neuromorphic system that merges biological neurons with silicon chips to create adaptive, self-learning processors. By leveraging living neurons to process information and respond to stimuli, the company is exploring a radical new approach to neuromorphic computing, blending synthetic intelligence with biological intelligence for real-time learning and decision-making.

Gemesys (Pre-seed, $12.43M in funding) is developing analog-based neuromorphic processors designed to mimic the functionality of biological neurons and synapses. Their approach leverages memristive and event-driven computing to enable real-time, adaptive learning with ultra-low power consumption, eliminating the need for traditional backpropagation.

Mentium Technologies (Pre-seed, $8M in funding) develop low-power, brain-inspired AI acceleration hardware optimized for edge computing applications. Their technology focuses on energy-efficient processing by integrating event-driven computation and local memory storage, similar to how neuromorphic architectures operate. While they don’t explicitly mention spiking neural networks (SNNs), their emphasis on efficient, real-time AI processing aligns with key neuromorphic principles.

CyberSwarm (Seed, $4.5M in funding) has developed a neuromorphic computing engine that mimics brain functions, specifically designed for low-power, edge-computing applications in IoT devices, robots, and vehicles. This technology employs memristors to create hardware artificial neurons, facilitating local execution of machine learning algorithms while enhancing data security and enabling brain-computer interface capabilities.

Neuronspike (Pre-seed, $174k in funding) develops neuromorphic processors that leverage spiking neural networks (SNNs) for ultra-low-power, event-driven AI processing. Their hardware mimics the way biological neurons communicate, enabling real-time learning and inference with significantly reduced energy consumption. NSPIKE’s technology is particularly suited for edge AI applications, where power efficiency and adaptive intelligence are critical.

Neurobus (Unfunded) specializes in neuromorphic computing solutions tailored for space applications, offering real-time detection, tracking, and identification capabilities with low power consumption. Their technology enhances space situational awareness and autonomous in-orbit services by processing information in a manner inspired by biological neural systems.

Analog AI (Unfunded) develops neuromorphic computing hardware that leverages analog in-memory processing to enable real-time learning and ultra-low power AI. Their architecture integrates computation and memory, mimicking biological neural networks to improve efficiency and adaptability in AI applications.

Existential AI (Unfunded) is developing neuromorphic-inspired AI solutions designed for real-time adaptability and efficient decision-making. Their approach aims to integrate event-driven processing and biologically inspired learning into AI systems, enhancing situational awareness, energy efficiency, and on-the-fly adaptation across various industries, beginning with healthcare.

Twistient (Unfunded) is developing neuromorphic-inspired AI solutions that leverage event-driven processing and low-power computation to enable real-time adaptability and edge AI applications. Their technology focuses on mimicking biological intelligence to improve efficiency and responsiveness in autonomous systems.

NimbleAI (Unfunded) is developing neuromorphic vision systems that integrate event-driven sensing and processing, inspired by biological vision. By combining spiking neural networks (SNNs) with low-power, edge-friendly hardware, NimbleAI aims to deliver highly efficient, real-time perception for applications such as robotics, autonomous systems, and augmented reality. Their approach reduces energy consumption while enabling dynamic, context-aware vision processing.

Thanks for reading.

Updates to this article will be made with every subsequent funding round. Stay tuned.

I put out a deep-dive like this one every week. If you enjoyed, follow me here.

If you’re interested in contributing to Garden Research as a writer, email me at pjlabarbera11@gmail.com.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Brain Inspired AI: A Deep Dive on Neuromorphic Computing was originally published in Garden Research on Medium, where people are continuing the conversation by highlighting and responding to this story.

Open-source software, by many measures, rules the tech world.

• 96% of proprietary codebases are built on top of open-source components
• 90% of cloud infrastructure runs on open-source software like Linux, Kubernetes, or PostgreSQL
• Trillion-dollar industries, and massive corporations like Meta, Google, and OpenAI have profited heavily off PyTorch and TensorFlow

Yet, only 12% of open-source maintainers, and even fewer contributors, receive financial support or compensation for their work.

That’s the model, and by many metrics, it works fine.

But what happens when it doesn’t? The 2021 Log4Shell vulnerability — a security flaw stemming from an open-source model — affected millions of systems simultaneously and can be attributed to an overworked, burnt-out, under-compensated maintainer left responsible for keeping an essential open-source tool secure.

Merit Systems, who recently raised a $10M seed round led by a16z Crypto and Blockchain Capital, is attempting to introduce fair market dynamics into the open-source ecosystem.

Can they succeed where others have failed?

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Key Gap Served

Open-source refers to software whose source code is freely available for anyone to view, modify, and distribute. It’s an example of collaborative innovation in its purest form. On any given project, thousands of devs can come together to create a product asynchronously and without traditional central oversight.

The first example of the open-source model can be traced back to 1983 when Richard Stallman launched the GNU Project, a free alternative to the Unix operating system. Stallman’s goal was to ensure software could be freely used, modified, and shared without restrictions.

These tools laid the foundation for many modern-day mainstays, like Linux, to evolve into the backbone most modern day systems.

AI, cloud computing, and nearly every major industry depend on open-source software. Many of the technologies we take for granted today were built on top of open-source software. Primary examples below:

• Linux: Open-source operating system — Powers 96% of cloud servers, supercomputers, and Android devices
• Kubernetes: Automates cloud application deployment — Used by AWS, Google Cloud, and Microsoft Azure for cloud computing
• PostgreSQL: Open-source database system. Used by Apple, Instagram, Netflix, and financial institutions
• PyTorch & TensorFlow: AI and machine learning frameworks. Powers AI models from OpenAI, Google DeepMind, and Tesla Autopilot
• OpenSSL: Secure internet encryption (TLS/SSL) — Used in banks, healthcare, and government security systems
• Apache Kafka: Real-time data streaming platform. Critical for Uber, LinkedIn, and stock market trading
• WordPress: Website publishing platform — Powers over 40% of all websites globally
• Android OS: Open-source mobile operating system — Runs on over 3 billion smartphones worldwide
• Node.js: JavaScript runtime for web applications — Used by Netflix, eBay, and LinkedIn for backend services

How does it work?

Generally speaking, open-source participants can be broken into three classes: maintainers, contributors, and users.

• Maintainers: Oversee the project, review pull (change) requests, merge code, and ensure stability and security of the codebase
• Contributors: Submit code, fix bugs, add features, and improve documentation
• Users: Use open-source software as a base-layer for their software (reminder: 96% of proprietary software contains open-source code)

GitHub and its Impact

GitHub, released in 2008, made open-source collaboration widespread, accessible, and seamless. Before its release, open-source developers relied on fragmented environments to find and participate in projects. GitHub centralized it all.

With over 100 million active contributing developers and 420 million repositories as of 2023, GitHub has become the backbone of modern-day open-source development. Virtually every major open-source project is hosted there, from Linux and Kubernetes to PyTorch and OpenSSL. It has become the default hub for individuals, startups, and Fortune 500 companies alike, integrating seamlessly into global software supply chains.

GitHub’s role in shaping the modern open-source economy cannot be overstated.

It revolutionized how code is written, shared, and improved, introducing powerful collaboration tools like:

• Pull requests & version control: Developers can submit changes while maintainers ensure quality.
• Forking & cloning: Anyone can experiment without modifying the original project.
• Issue tracking & discussions: Open-source projects are managed like professional software products.
• CI/CD pipelines & automation: Developers can test and deploy software directly within GitHub.

Still, regardless of how efficient GitHub has made open-source building, a fundamental gap remains; one that may be philosophical in nature, which we’ll touch on, but a gap nonetheless. The global tech economy owes billions of dollars in revenue to open-source, but the majority of contributors never see a dime for their contributions.

Gap #1: The Open-Source Market Failure: A $500B+ Industry with no Revenue Model

If I were Richard Stallman, I’d be pissed. His “free software” revolution was right in intention — to liberate users and developers alike from corporate oligarchy, but in practice, it evolved into a marketplace that enables trillion-dollar companies to benefit from an unpaid system of labor.

Let me be very clear. This is not a malicious act on the company or user’s part. Corporates are not being malevolent by not paying because there is no incentive to pay.

The open-source model is free to use by nature. There are no licensing requirements, no built-in structures like customer support to encourage payment, and no enforcement mechanism disallowing users from benefiting from a project they didn’t contribute to.

It’s a free-rider economy. Maintainers are expected to do the job they signed up for — to secure and support the critical code they oversee — and it isn’t any specific users’ job to make sure they’re thanked for it.

Here’s where it gets slightly philosophical in nature. Many developers are happy not being paid. That’s not what open-source is for, and devs are comfortable benefiting from the ecosystem in other ways:

Skill Development & Portfolio Building

• Developers improve their coding skills by contributing to real-world projects.
• Open-source work serves as a public portfolio, increasing job opportunities. (Ex. Many developers land jobs at top tech firms by showcasing GitHub contributions)

Reputation & Industry Credibility

• Top contributors gain recognition and influence in developer communities.
• Maintainers become sought-after experts, often leading to job offers, speaking opportunities, and consulting work. (Ex. The creator of Python, Guido van Rossum, was hired by Microsoft partly due to his open-source reputation)

Passion for Software Freedom & Collaboration

• Many contributors believe software should be open, accessible, and community-driven.
• Some view proprietary software as restrictive, preferring decentralized, collaborative models. (Ex. The GNU Project and Linux grew from the belief that software should be freely shared).

If this system works, then great — everyone gets what they want. Companies get best-in-class software for free, and developers gain skills, credibility, and the satisfaction of contributing to something bigger than themselves. But what happens when the balance tips? What happens when unpaid maintainers can’t keep up with bug fixes? When a security vulnerability arises, or when a developer simply burns out and walks away?

These questions have already been answered. It happened.

In 2021, the Log4Shell vulnerability exposed millions of systems worldwide because a critical open-source tool was maintained by a single unpaid developer. When open-source software goes unsupported, the consequences extend far beyond developer frustration.

Companies assume that someone else will take care of the projects they depend on, but without structured funding, these projects risk abandonment, stagnation, or exploitation. And if that happens, businesses will be forced to pay the price retroactively — either through costly security breaches, infrastructure failures, or being left without viable alternatives. The current model works until it doesn’t, and right now, no one is on the hook to fix it.

Past Attempts at Fixing a Broken System

There have been several unsuccessful attempts at solving the funding issue, and one wildly successful effort to solve the issues that arise from the broken model.

GitHub sponsors, Open Collective, and Gitcoin are all examples of what happens when companies try to introduce funding mechanisms but do nothing to change the underlying incentive structure. These platforms assumed that if developers had an easy way to receive financial support, companies and users would step up. They didn’t. GitHub Sponsors relied on voluntary donations that rarely scaled. Open Collective made fundraising easier but failed to create recurring revenue. Gitcoin’s funding model fluctuated with crypto markets, making it unreliable. Each assumed companies would fund open-source out of goodwill, but none created a reason for enterprises to pay. Each failed.

The only company to successfully monetize open-source at scale was Red Hat, but they didn’t fund maintainers. They didn’t distribute wealth to contributors, and they didn’t sell open-source software. Instead, they sold a safety and support layer over the top of open-source. Enterprises paid Red Hat for a managed version of Linux that guaranteed security patches, compliance, and long-term support. Red Hat created something companies were willing to pay for, and they did — leading to a $34 billion acquisition by IBM in 2019. But maintainers saw none of that money.

That’s not to say companies never fund open-source. Some corporations fund projects they rely on, like Google backing contributors for Kubernetes and TensorFlow. Organizations like The Linux Foundation and Apache Software Foundation receive corporate sponsorships to maintain key projects. But funding is selective and uneven — companies support what benefits them while leaving thousands of essential dependencies unfunded.

So while some open-source projects get corporate money, most do not. The free-rider problem remains: there is no universal, structured funding model to compensate maintainers. Red Hat proved open-source can generate billions — but not in a way that fixes the root issue.

Gap #2: Corporate Inefficiencies and Shifting the Incentive Model

Like the open-source funding and incentive model, corporate innovation is also fundamentally broken. It’s hard to build software internally, rendering corporates with a few different options:

• Buy proprietary software that is tough to modify and often comes with vendor lock-in.
• Build on top of open-source code, which might not be perfectly suited for their needs and comes with security risks.
• Hire internal developers, which is costly, slow, and inefficient for non-software companies that lack deep engineering expertise.

None of these options are ideal. Just as freelance platforms like Upwork and Toptal allow companies to contract developers for short-term work, an open-source bounty marketplace could offer enterprises a faster, more flexible way to fund software development.

In addition, startups stand to gain just as much from a shifted incentive model as large enterprises. Many successful companies — MongoDB, HashiCorp, and Airbyte — began as open-source projects before commercializing, leveraging developer adoption as an organic growth strategy. However, without a structured funding mechanism, startups today have no direct way to pay the developers to maintain the open-source infrastructure they rely on. Implementing a successful bounty-based marketplace for development could lend startups access to top talent from around the world, allowing their products to be built in the open by an incentivized, decentralized group of developers.

We’ll get into this in more depth when discussing potential business models.

First, let’s introduce Merit Systems.

Founding Story

Merit Systems has a founding team with experience spanning software engineering, venture capital, and open-source:

• Sam Ragsdale (Founder, CEO): Former software engineer at Google and Investment Engineer at Andreessen Horowitz, where he was a member of the infrastructure deal team and open source engineering team
• Mason Hall (Head of Strategy): Former software engineer at BloXroute Labs and Partner on the Crypto Investments team at Andreessen Horowitz
• Ryan Sproule (Founder, CTO): Former software engineer at AWS and Principal Research Engineer at Blockchain Capital LLC, a crypto-dedicated venture fund

Their mission is to enable the open-source ecosystem to function more like a market, a version of “Open-Source Capitalism”, wherein both developers have the means to be compensated and companies are incentivized to pay for their work.

The question remains: Can an efficient, dual-sided market approach fix current inefficiencies?

Market-driven solutions have worked for SaaS, AI, and fintech, but open-source has historically resisted financialization. Will maintainers embrace this model, or will it be seen as an attempt to commercialize what was meant to be free? Most importantly, can Merit convince corporations to participate, or will this fall into the same trap as past failed funding attempts?

Product and Technology

“Open source software is eating the world. It empowers every major company — and supports trillions of dollars of value every year. It’s a lever on the world — the way to make more with less. What if we could give it superpowers?

We believe the path forward is through Open Source Capitalism — where everyday market forces, not uncertain subsidies, drive invention. A world where builders can earn a fair share and move fluidly between projects, coordinating at an unprecedented scale.

How do we get there? Merit automates ownership in open source, adding attribution so funds flow directly to builders. By programmatically tying capital to impact, it spurs innovation, fosters a dynamic market, and seeks to attract top talent looking for maximum agency and impact.”

This was sourced directly from a post by Mason Hall, Head of Strategy, when he announced Merit Systems’ Seed round. Technical details on product, approach, and technology have not been made entirely clear, but this section will attempt to both detail what has been shared and hypothesize about what is less clear.

Based on the available information, Merit Systems is building a software layer meant to be deployed on top of existing platforms like GitHub to create an automated ownership, payment, and attribution system for open-source projects.

This will allow for detailed tracking of contributions and programmatic distribution of funds to developers based on their impact.

How I can see this working in practice:

Attribution:

• GitHub & GitLab Integration → Merit attaches to repositories, logging and verifying contributions in real-time.
• Weighted Contribution Scoring → Determines which contributions deserve more funding (e.g., fixing critical security flaws vs. minor code tweaks).
• Smart Attribution System → Prevents disputes over who should get paid and how much.

Payment:

• Smart Contract Payments → Funds are automatically distributed when contributions meet specific conditions.
• Developer Wallets or Payment Accounts → Contributors register for payouts, ensuring verified compensation.
• Payment Pools → Companies allocate funds for desired edits or builds on the projects they depend on or create a pool to start a custom project they need

Key Considerations and Theories

Below, I’ve listed out some questions I had during my research that the Merit Systems team has not given public answers to, plus some of my theories on how they might be approaching it.

Please note that none of the information below is confirmed, and is purely conjecture based on what I know.

How will Merit Track Contributions Without Disrupting Existing Workflows?

Developers won’t adopt a system that adds frictions to the processes they’re used to. Merit has suggested their product is an attribution layer built into GitHub and other platforms, meaning it must work seamlessly within the bounds of these established frameworks.

How They Might Solve This:

• Version Control Integration: I think it’s likely Merit will track contributions directly through GitHub’s API, analyzing commit histories, pull requests, and issue resolutions
• Automated Attribution Tagging: Rather than requiring developers to manually log work, which would disrupt flows, Merit may be attaching metadata to commits to identify contributions programmatically

These systems would enable passive tracking of work without leaning on user input or logging. This is likely what is meant when they refer to “automated attribution tracking”.

Some remaining questions:

• Will this be an automated layer across all GitHub repositories, or will maintainers need to opt in?

Will Merit Rely Solely on GitHub Data?

If Merit only racks GitHub activity, they might miss valuable contributions happening off platform. That said, creating a proprietary tracking systems introduces added complexity and may affect usability/adoption.

How They Might Solve This:

• GitHub as The Primary Data Source: The path of least resistance is to pull most commit data, issue resolutions, and contribution history from GitHub, as it is the most used platform.
• Expand to Multi-Platform Tracking: To broaden out their capabilities, they may also integrate or have plans to integrate with platforms like GitLab, Bitbucket, or self-hosted repositories

Some remaining questions:

• How will it handle projects that are partially developed on GitHub but rely on other collaboration tools (e.g., Slack, Jira)?

How is “Value” Determined?

Not all contributions to a project are equal. Patching a major security bug is more important than making UI prettier. So, how will impact be accurately measured? If this is not done well, developers may spend time on “easier” but less impactful contributions to optimize for payouts.

How They Might Solve This:

• Weighted Contribution Scoring: Merit could assign different weights to contributions based on difficulty, criticality, or impact.
• Hybrid Model: Algorithm + Human Review: AI or algorithms could do the first pass of ranking, while project maintainers or stakeholders validate funding decisions.
• Crowdsourced Governance: If Merit incorporates decentralized elements, funding allocations could be partially community-driven.

Will Merit Rely on Blockchain?

The founders have blockchain backgrounds, and a16z Crypto led their round, so it’s plausible Merit incorporates blockchain fundamentals into their product.

What They May be Building:

• Hybrid System: Fiat + Blockchain: Merit could enable both traditional payments and crypto-based incentives, giving users flexibility.
• On-Chain Contribution Logging: Blockchain could be used to permanently record contributions, ensuring transparency and accountability.
• Decentralized Payments Through Smart Contracts: If Merit uses blockchain, payments could be automated without needing a central payment processor.

Similarly, Will Merit Rely on Crypto, Fiat, or a Hybrid System for Payments?

Crypto is emerging as a valid corporate tool through stablecoins, but still has some hurdles to gap. Transparency will be key in a system like this, but Merit needs to ensure compliance is not an issue.

How They Might Solve This:

• Dual Payment Model: Merit could enable crypto-based micropayments for automation, while larger corporate transactions occur in fiat.
• Stablecoins for Payment Stability: If Merit does use crypto, it may adopt stablecoins to eliminate market volatility.
• Direct Bank Transfers for Enterprise Users: To encourage adoption, Merit may offer fiat-based payment rails for corporate funders.

In Any Case, Who Sets Pricing?

Markets need pricing mechanisms. Merit will need to establish a system to determine the market rates for dev work.

How They Might Solve This:

• Bounty-Based Pricing: Companies set their own bounties, and developers choose which to complete based on payout.
• Algorithmic Compensation: A system could analyze contribution difficulty, project complexity, and historical payouts to suggest fair compensation.
• Maintainer-Led Rate Setting: Project maintainers might have the authority to set minimum payout requirements.
• Reputation-Weighted Pay Scaling: Experienced contributors with a strong track record could command higher payouts.

Business Model

Now to address the most critical, fundamental challenge. How will Merit succeed where many have failed?

Companies have never had any reason to pay for open-source, and truly, that’s never what open-source was meant for. What’s powerful enough to change that?

Once again, the following is purely conjecture based on the limited amount of public information. This is the system that makes the most sense to me based on what I know, and may not be representative of what Merit is actually planning to do

What Do Corporations Lack, and What’s Missing from Open-Source?

Corporations have an innovation problem. Buying licensed software is easy, but off-the-shelf solutions rarely fit their exact needs. The alternative — hiring internal developers to build custom tools — is slow, expensive, and often derailed by corporate bureaucracy. Innovation inside large companies frequently stalls, not because they lack resources, but because they lack efficient ways to build what they actually need.

Meanwhile, open-source doesn’t build for any one company — it builds for broad, practical use cases. Instead of catering to specific enterprise demands, open-source projects focus on creating tools that solve common problems across industries. That means corporations can use open-source, but they can’t shape it to their exact specifications without heavy internal development.

What if Merit was building a system that let companies — from startups to enterprises — directly leverage the open-source ecosystem to create tools tailored to their needs? Not only would this accelerate innovation, but those tools could then proliferate across the entire tech ecosystem, benefiting any company facing the same challenges. Instead of being locked into proprietary software or hampered by slow-moving internal teams, companies could fund open-source development that serves their exact use case — while strengthening the ecosystem as a whole.

This introduces a suite of new incentives that are non-existent in the market currently:

• Faster, More Custom Innovation: Instead of waiting for internal teams to build solutions, enterprises can fund open-source projects to get tailored, cutting-edge tools faster.
• Access to Elite Developers: Companies struggle to hire top software engineers — Merit could give them a way to contract world-class open-source talent on-demand.
• Security & Stability: Funding open-source development ensures critical dependencies remain stable and secure, reducing future risks like Log4Shell.
• Cost Savings Over Proprietary Software: Instead of licensing expensive proprietary software, companies could fund the development of custom open-source tools that suit their exact needs.

So, how would that model work in practice?

I’m envisioning some sort of bounty system. Companies could post bounties for specific needs, with a pool of funds, and developers can use an open search across bounties to work on projects that interest them.

Devs get to work on cool issues and get paid for it, and companies get a more custom tailored version of open-source while benefiting from accelerated time to solution and innovation.

AGAIN… no idea if this is actually what they’re doing, but it makes the most sense to me…

I could see this enticing a suite of different company types as well:

• Enterprises: Need custom-built, open-source solutions without slow-moving internal teams. A bounty system lets them fund specific features, security fixes, and integrations on demand.
• Traditional Startups: Struggle to hire top engineers but need to move fast. Bounties allow them to pay for targeted development work without committing to full-time hires.
• Startups Building Open-Source as a Business Model: Many modern startups build their core products as open-source (e.g., MongoDB, HashiCorp, Airbyte) and monetize through enterprise services, hosting, or premium features. A bounty system could help them fund development, engage top contributors, and build a stronger ecosystem around their product.
• VCs & Investors: Increasingly funding open-source-based startups, but often lack control over whether critical dependencies will be actively maintained. Bounty funding could help de-risk their portfolio companies by ensuring foundational open-source tools remain stable and well-supported.
• Corporate R&D Labs & Innovation Teams: Large enterprises looking to experiment with new technologies could use bounties to test ideas, prototype solutions, and validate concepts faster than hiring internal teams.

How will Merit Make Money?

Again, unclear based on public knowledge, but could be a mix of:

• Transaction Fees on Bounties → Merit takes a small percentage of every successfully completed bounty, similar to how Upwork or Fiverr charge fees on freelance work.
• Enterprise Subscription Plans → Large companies pay for priority access, advanced security assurances, or premium analytics on open-source dependencies.
• Developer Premium Accounts → Contributors could pay for visibility boosts, faster payouts, or access to exclusive high-paying bounties.
• Partnerships with Open-Source Foundations → Merit could work with The Linux Foundation, Apache Software Foundation, or CNCF to integrate funding mechanisms into widely used open-source projects.
• Corporate Sponsorships & Grants → Companies reliant on open-source could pay into a general funding pool that Merit distributes to critical projects.
• Data & Insights Services → Enterprises could pay for detailed usage analytics, dependency tracking, and ecosystem intelligence to manage risks associated with open-source reliance.

Competitive Positioning

Previous attempts at fixing open-source funding have largely failed. Donation-based models like GitHub Sponsors and Open Collective never scaled — they assumed companies would fund open-source out of goodwill, but created no actual incentive to pay. Corporate funding remains selective and uneven, with companies only supporting projects that directly benefit them while leaving thousands of critical dependencies unfunded.

Merit is taking a different approach. Rather than relying on donations, they’re building what appears to be a marketplace where companies can directly fund development work they need. The closest comparison might be HackerOne, which proved companies will pay significant money for security fixes. Merit is betting this model can extend beyond security to feature development, integrations, and maintenance.

The key insight is incentive alignment. Instead of asking companies to donate to open-source broadly, Merit could enable them to fund specific development that serves their exact needs. If executed well, this could transform open-source from an unpaid labor model into a true marketplace for top engineering talent, with market dynamics driving both innovation and compensation.

Challenges and Opportunities

Key Challenges

• Enterprise Adoption — Companies have historically avoided paying for open-source. Will the marketplace model drive consistent funding, even with direct benefits?
• Pricing Mechanics — Establishing fair bounty values is complex. Risk of either undervaluing developers or creating opportunities to game the system.
• Community Response — Open-source developers might resist a market-driven model, viewing it as contrary to open-source principles.
• Quality Control — Must prevent low-quality contributions, duplicate work, and bounty manipulation in a paid marketplace.

Market Opportunities

• First Scalable Marketplace — No one has successfully built a structured marketplace for open-source development. Merit could unlock an entirely new model for funding and building open-source software.
• Clear Market Need — Both startups and enterprises struggle with innovation bottlenecks. If Merit makes custom open-source development faster and cheaper than traditional engineering teams, adoption could accelerate.
• Security Incentives — As supply chain risks increase, enterprises might view Merit as a critical security investment, ensuring open-source dependencies remain stable and maintained.
• Sustainability Model — Could create the first financially sustainable model for open-source maintainers and contributors.
• Industry Impact — Success could fundamentally reshape how open-source development is funded across tech.

Conclusion

Obviously, there’s a lot of intracacies that are unclear about what the guys at Merit Systems are building. That said, based on this analysis, I think the potential here is really fascinating. Done well, this could fundamentally alter the open-source model (a $500B market opportunity…) for good.

Thanks for reading.

Updates to this article will be made with every subsequent funding round. Stay tuned.

I put out a deep-dive like this one every week. If you enjoyed, follow me here.

If you’re interested in contributing to Garden Research as a writer, email me at pjlabarbera11@gmail.com.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Marketization of Open-Source was originally published in Garden Research on Medium, where people are continuing the conversation by highlighting and responding to this story.

A predominant majority of software being pushed today, from both startups and enterprises alike, has been at least partially generated by AI.

In 2023 alone, 92% of polled U.S. developers reported using AI-powered coding tools like GitHub Copilot in their workflows. At Google, AI is responsible (as reported by Sundar in Oct. 2024) for generating 25% of all new code.

It’s only natural. The speed and efficiency with which AI systems can generate code enables devs to be incredibly productive, and code is being written and pushed at a higher pace than ever before.

But as AI code floods the market, it begs the question: Who’s verifying the code? Who’s making sure it is working as intended? For years, software development has been an intensely human-driven process, with engineers meticulously reviewing, optimizing, and debugging every line of code. Now, with AI writing millions of lines of code, the chance for technical debt, security vulnerabilities, and misaligned logic is accumulating at an unprecedented rate.

AI might be able to write code, but does it truly understand it?

Almost a year ago to the date, Merly AI raised ~$7M from Greycroft and IBM Ventures to address this issue directly. This article outlines the broader challenges facing the AI-driven coding market while discussing how and why Merly may be uniquely positioned to be a disruptor in a sea of nearly identical approaches.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Key Gap Served

As discussed above, AI-generated coding has proliferated deep into general software development workflows. Naturally, the black-box nature of the AI decision-making process, especially prevalent in LLMs, has introduced new risks. With little to no direct oversight of how the code is generated, the chance of bugs, security vulnerabilities, and technical debt surfacing has risen.

A report from the Center for Security and Emerging Technologies (CSET), released in Nov. 2024, outlines three broad categories of risk associated with AI code generation models:

• Models generating insecure code.
• Models themselves being vulnerable to attack and manipulation
• Downstream cybersecurity impacts, such as feedback loops in training future AI systems

What’s more concerning is this report shows a fundamental lack of understanding or awareness of these risks. While 76% of software developers polled believe that AI-generated code is more secure than human-written code, the same survey found that more than half of organizations encounter security issues or bugs with AI-generated code “sometimes or frequently.”

As a result, much of the work of the software developer has shifted from writing code to trying to understand it to mitigate risks. According to a recent report, 70% of their time “understanding code rather than writing it”.

There’s a reason this bottleneck exists, and it comes down to the fundamental nature of LLM-based code generation.

Why Existing Solutions are Fundamentally Inefficient

OpenAI’s GPT models, Llama, Gemini, and many of the AI-coding platforms like Poolside and Cosine are LLMs. Their output is based on the prediction of the most likely sequence of tokens based on training data, meaning that no matter how intelligent they seem, they operate as probabilistic pattern-matching engines rather than true reasoning systems.

Training and Architecture

In the simplest of terms, LLMs are expert guessers. They don’t “reason” in the traditional sense and they aren’t deterministic.

These models are built using transformer-based neural networks that excel at processing large amounts of sequential data. Input training data (including code, words, etc.) is broken into tokens (small chunks of words, symbols, or code), which are converted into vector representations so that the model can understand relationships between words and concepts. In this sense, similar token types (like “for” and “while”, or “add” and “sum”), have similar vector representations.

Transformer layers of the model allow it to determine which parts of the input sequence are most relevant based on vector relationships, and feedforward layers help refine relationships as time goes on. For example, when predicting the next token in def calculate_tax(income), the model will give more attention to terms like “return”, “income”, or “tax_rate”.

To become proficient at coding, the model is trained on massive datasets of human-written code. If, in the training data, the function return income * tax_rate comes up often, the model will learn that “*” often follows “Income” in tax-related calculations. It doesn’t understand taxes, or why it’s calculated that way, it just associates those tokens together based on frequency and probability.

Prompt Processing and Code Generation

So, when a developer prompts any of these LLMs to write code, the following steps occur:

• Prompt encoding: The model converts the prompt into tokens and compares it to its training data
• Pattern Matching: The model identifies patterns from similar tokens seen before
• Token-by-Token Generation: The model generates one token at a time, based on probability distributions for what each individual token is most likely to be followed by. The most probable token is selected, and this process repeats until the output is complete. This is why LLMs can sometimes generate different outputs for the same prompt — they’re sampling from probability distributions, not retrieving static answers.

The issue this creates

LLMs aren’t “thinking” through the logic of a program, and they definitely don’t understand the intent behind the code.

They’re masters of replication and association, but the priority is placed on probability, not accuracy. If an incorrect coding pattern is common in training data, the mistakes will proliferate in the LLMs response. For example, if the tax rate in 80% of the cases in the training data is 5%, the LLM might incorrectly assume that all tax calculations use a flat 5% rate, which is not the case. Other issues include overfitting to training data and inherent struggle with edge cases and context that are not well represented in training data.

Because LLMs lack true intent-based reasoning, they introduce significant risks in production environments:

Security Vulnerabilities

• AI-generated code frequently contains insecure implementations.
• A study by Stanford University found that LLM-assisted programmers were more likely to introduce security vulnerabilities compared to those who wrote code manually.
• Example: GitHub Copilot has been found to suggest vulnerable code in cases like SQL injection risks, missing authentication steps, and improper input validation.

Technical Debt & Maintainability Issues

• Since AI-generated code is optimized for short-term functionality rather than long-term maintainability, developers may accumulate significant tech debt by relying on it.
• Poorly structured AI-generated code can be difficult to refactor and debug.

Unpredictable & Inconsistent Outputs

• Because LLMs randomly sample from probability distributions, they do not always return the same answer to the same prompt.
• This lack of determinism makes AI-generated code less reliable in production settings.

The resulting market gap

Because nearly 100% of AI code generation is based on LLM models, there is a significant need in the market for a tool that can objectively, and deterministically assess the credibility and accuracy of the code, and the potential risks present.

Enter Merly.

Founding Story

Merly is founded by Dr. Justin Gottschlich, developer of the “Three Pillars of Machine Programming” framework that defines how AI can fundamentally change software development (successfully). The framework found below outlines how machines can autonomously understand, create, and evolve code over time. This is the foundation of Merly.

Pillar One: Intention

One of the biggest issues within AI-assisted software development is the translation of human intent to code. Writing software is structured, but human ideas are ambiguous. For a machine to properly generate code, it has to bridge the gap between what the prompter “says”, what the prompter “wants”, and what specific logic needed is to execute that goal. Current AI-generated code often misinterprets intent, putting a burden on developers to precisely articulate intent, something even the most experienced programmer will struggle to do.

Pillar Two: Invention

Once the intent is understood, the next step is to write the code. Unfortunately, this isn’t as simple as it sounds. It’s not just about producing any code, it has to be the right code for the problem at hand. If solving that problem requires a novel approach, LLMs creating code based on historical examples will struggle to produce novel problem-solving approaches. What’s more — even if the code does “work”, it may not be optimal.

Pillar Three: Adaptation

Software isn’t static. Codebases evolve, requirements change, and security vulnerabilities naturally emerge. Today’s models focus on writing new code, not maintaining and improving existing code. This makes sense when you consider the fundamental token-by-token probability-based approach of modern LLM models. To make a truly proficient software dev, it must not only generate code but refine and optimize it as the software evolves.

Dr. Gottschlich recognizes that AI-generated code is inevitable, but unaddressed problems exist in the market:

• Current models aren’t very good at writing “great” code
• There’s no solution in the market to check if the code is “bad”

Merly was founded to address this core issue. Their near-term goal is to ensure AI-generated software is reliable, scalable, and high quality.

We’ll discuss long term ramifications of successful achievement of the near-term goal later.

Product and Technology

At its core, Merly is a privacy-first, deterministic AI that has been built to analyze code for quality, effectiveness, and correctness.

Their first model, Mentor, allows users to upload their repository into Merly for quality scoring, anomaly detection, and improvement. Merly’s features include:

• Mentor Score: Measures software quality over time.
• Mentor Compare: Benchmarks similar repositories.
• Mentor Insights: Detects anomalies and inefficiencies.
• Mentor Watch: Alerts for significant code quality shifts.
• Mentor Auto: Integrates with CI/CD pipelines for continuous monitoring.

Key Differentiators

Model

Merly is not an LLM, it’s a logic-based, deterministic AI reasoning engine trained specifically for code analysis. It does not work by predicting code, but by understanding and validating code structure and intent through an advanced, multi-tiered abstraction model. This means that the exact same output will be generated for specific inputs. No “guessing”.

Training

Unlike other AI-generated coding solutions in the market, Merly does not train on human-written code. Rather, they translate human-written code into an abstracted representation that removes language-specific idiosyncrasies and narrows in on the core logic and structure of the code.

Training data is sourced from open sources like GitHub and Stack Overflow, where both code and the documentation, bug reports, and discussions that surround the code are ingested, allowing Merly to relate human intent represented in the surrounding information to the end code logic.

By using this mass amount of data, Merly can analyze how similar logic is structured across thousands of repositories — millions of lines of code — to understand:

• The common structure of well-written code
• How code evolves throughout its lifetime
• How user intent, instructions, and goals are correlated to machine logic

Energy and Compute Efficiency

Most LLM-based coding assistants require high-dimensional matrix calculations across billions of parameters, involving complex operations that require GPUs to run properly. Merly can run on a single CPU.

Merly achieves this in two ways:

• Training: As outlined above, Merly does not analyze raw source code directly. Instead, the code is transformed into an intermediate representation prior to training. This means that instead of processing every individual token, it’s training on a reduced logic structure, subverting the need to process every token individually. This vastly reduces compute requirements for data processing.
• Compute: LLMs work by predicting probabilities for every next token, requiring potentially billions of matrix multiplications. These operations require GPU hardware for execution. In contrast, Merly runs a deterministic reasoning engine. It’s not guessing or generating code, just systematically checking underlying logic and quality, requiring much less computational power.

Privacy by Design

The lack of computational power needed to run Merly allows it to be run on-prem, and on-device without special hardware. In addition, translation to machine language in the intermediary processing steps secures private data from being ingested.

Continuous Learning through Human in the Loop

Every recommendation Merly makes to update, fix, or improve code is run by a developer. This ensures that AI recommendations align with the developer’s intent and that the nuances of a project’s architecture, dependencies, and long-term goals remain intact.

The model learns from every interaction, improving its reasoning over time for the specific projects it’s working on, and enabling a better understanding of the specific coding styles, preferences, and needs of the devs across the entire lifecycle of the codebase.

Back to the Pillars

How does this approach compare and contrast with LLMs as it relates to Intention, Invention, and Adaptation?

Intention

Restating the problem: AI-generated code often misaligns with developer intent, leading to bugs, inefficiencies, and security vulnerabilities.

How Merly Solves It:

✅ Intent Training and Validation: Merly focuses on underlying code logic and structure within the context of its surrounding discussions, bug fixes, documentation, etc.

✅ Human-in-the-Loop AI Feedback Loop: Every recommendation is reviewed by a developer, refining Merly’s understanding of how developers structure code for specific use cases.

✅ Privacy-First Learning: Unlike LLMs that require massive datasets and cloud processing, Merly’s on-device learning model allows it to improve continuously without compromising source code security.

Invention

Restating the problem: LLM-based AI code assistants (e.g., Copilot) guess code patterns but do not understand software logic or best practices. There is no current AI model that can generate code while also verifying its correctness in real time.

How Merly Solves It:

✅ Deterministic Code Reasoning: Instead of generating code probabilistically, Merly analyzes and scores code based on logic, security, and efficiency.

✅ Model Improvement Through Use: As more developers interact with Merly, its ability to predict better solutions increases, creating a self-improving model that enhances every recommendation.

✅ Future Expansion into AI Code Generation: Merly’s deterministic approach to code validation and optimization makes it well-positioned to eventually move into AI-powered code generation. Unlike existing models that generate code first and validate later, Merly could evolve into a system that generates correct code from the start, reducing errors and security vulnerabilities at the point of creation.

Adaptation

Restating the problem: Codebases degrade over time, accumulating technical debt, security risks, and inefficiencies. Traditional code review tools focus on static analysis rather than continuously monitoring code quality over its entire lifecycle.

How Merly Solves It:

✅ Long-Term Code Monitoring: Unlike traditional static analysis tools that only assess code at a single point in time, Merly continuously evaluates how codebases evolve over months and years.

✅ Proactive Security and Performance Updates: Detects aging, inefficient, or vulnerable code before it becomes a major problem. Identifies code drift and silent bugs that developers might otherwise miss.

✅ Integration with CI/CD Pipelines: Merly’s Mentor Auto feature integrates into continuous deployment environments, ensuring that every new commit is validated in real-time.

✅ Custom Adaptation Per Codebase: Merly remembers previous developer decisions and refines its recommendations based on historical patterns from each specific repository.

Business Model

Revenue Streams

• Enterprise SaaS licensing for large organizations.
• Startups relying on AI code-gen, distribution through VCs to portfolio companies
• M&A due diligence processes relying on manual code review

Cost Advantages

• No need for expensive GPU hardware — runs efficiently on standard CPUs.
• Reduces software vulnerabilities, minimizing security-related expenses for enterprises.

Customer Value Proposition

• Helps systematically validate code at scale.
• Reduces tech debt, security risks, and debugging overhead.
• Aligns AI-written code with developer intent.

More Use, More Data, and Better Model

• The human-in-the-loop feedback process ensures that Merly continuously improves with more usage.
• Each interaction refines its understanding of developer intent and best practices, creating a compounding improvement effect.
• Unlike LLM-based tools that rely solely on static training data, Merly’s adaptive learning system makes it more accurate over time for specific teams.

Competitive Positioning and Future Outlook

Merly has positioned itself as a necessary tool for teams leveraging AI-based software development platforms. It’s the best product in the market I’ve seen tackling verifying, debugging, de-risking, and maintaining code. As such, with the right GTM strategy, Merly should become a significant add-on to software dev workflows.

While AI code-generation startups are not their direct competitors yet, I’m confident in my thesis that this is just the first step for Merly. By supplanting themselves into the market and continuously training Merly to be a better, deterministic coding engine, I think they’re primed to take on other startups like Poolside, which is currently valued at $3B.

Unlike current AI coding tools, which generate code and then require post-hoc debugging, Merly could create pre-validated, high-quality code from the start, eliminating unnecessary debugging cycles. If successful, this would position Merly as the first AI-driven system that both generates and verifies software with deterministic precision, surpassing traditional AI coding assistants entirely.

Challenges and Opportunities for Merly

Challenges

Scaling to larger codebases: As software projects grow in complexity, Merly must ensure that its performance remains efficient even in massive repositories containing millions of lines of code. AI-driven reasoning at this scale presents computational challenges, requiring optimization techniques that maintain speed without sacrificing accuracy. Unlike LLMs, which are computationally expensive, Merly’s deterministic reasoning model must continue operating efficiently as adoption scales across large enterprise environments.

Convincing enterprises to adopt AI validation tools: Many organizations remain skeptical of AI’s role in software validation, as existing AI-powered coding tools have introduced security risks, hallucinations, and technical debt. Enterprises accustomed to manual code reviews and traditional static analysis tools may be hesitant to trust an AI-driven validation system, especially when dealing with mission-critical applications. Overcoming this skepticism requires demonstrating Merly’s tangible benefits, including enhanced software quality, reduced debugging time, and long-term maintainability improvements.

Regulatory and compliance concerns: Certain industries — such as finance, healthcare, and government — require strict compliance measures when using AI-driven tools for software development. Merly must navigate legal and regulatory frameworks that dictate how AI systems interact with sensitive codebases. Ensuring that on-prem deployment, auditability, and data privacy protections meet industry-specific compliance standards will be essential for enterprise adoption in regulated sectors.

Opportunities

First-mover advantage in AI code reasoning: Unlike existing AI coding tools that focus on code generation, Merly is pioneering an entirely new category: AI-driven code reasoning. No other product systematically evaluates AI-generated code for correctness, security, and intent at scale. Establishing itself as the first enterprise-ready AI validation layer provides Merly with a strong competitive moat as AI-generated software becomes more widespread.

Growing demand for privacy-first AI solutions: Enterprises are increasingly seeking AI tools that can be deployed on-prem or within private cloud environments to avoid data security risks. Unlike GitHub Copilot and other LLM-based coding assistants that require cloud-based inference, Merly’s fully local processing offers a key differentiator for security-conscious organizations. As regulatory pressure around AI privacy increases, Merly’s privacy-first design will position it as a preferred solution in enterprise and government environments.

Alignment with macro trends: As AI-generated code becomes the norm, the demand for AI-driven validation tools will surge. Enterprises will need trusted systems to verify AI-generated software, ensuring that it remains functional, secure, and maintainable over time. Merly is strategically positioned to become an essential infrastructure tool for AI-driven software development, much like CI/CD tools became indispensable for DevOps workflows.

Potential future expansion into AI code generation: While Merly currently focuses on AI-driven code validation, its underlying research suggests a future expansion into AI-powered software generation. Unlike LLM-based tools that generate code without built-in verification, Merly could eventually introduce pre-validated AI-generated code — ensuring correctness before execution. This would provide a significant competitive advantage over existing AI coding assistants, making Merly both a validation and generation system in one.

Thanks for reading.

Updates to this article will be made with every subsequent funding round. Stay tuned.

I put out a deep-dive like this one every week. If you enjoyed, follow me here.

If you’re interested in contributing to Garden Research as a writer, email me at pjlabarbera11@gmail.com.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Achieving True Machine Programming was originally published in Garden Research on Medium, where people are continuing the conversation by highlighting and responding to this story.

Globally, transportation accounts for approximately 7 billion metric tons of CO₂. Of this, freight, or shipping goods, merchandise, and cargo from one location to another, contributes 2.1 billion metric tons, or 29 percent.

Most of this contribution can be attributed to trucking (76%), maritime shipping (14%), and aviation (8%) — areas where most current decarbonization efforts are focused.

In November, Philadelphia-based Voltify raised $2.5 million in funding from theDOCK, E44 Ventures, and J-Impact to take a different approach — turning their attention away from traditional target markets for decarbonization and towards rail freight.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Key Gap Served

Rail freight handles 28% of freight ton-miles in the U.S., moving approximately 1.7B tons of goods annually. Ideal for long-haul routes, rail is critical for the movement of bulk commodities like grain, chemicals, coal, and manufacturing materials.

The majority of the remaining 72% of transport is carried out by trucking. Trucking dominates short and medium-haul freight due to its flexibility and ability to provide door-to-door service, something that is not achievable using rail infrastructure.

When looking at total emissions, there is a fundamental and potentially surprising imbalance in the makeup of total emissions. Though trucking only accounts for 68% of all ton-miles, it contributes an 85% share of freight emissions (around 400 metric tons of carbon). On the other hand, rail, handling 28% of ton-miles, only accounts for 2% of freight emissions (around 25 million metric tons of carbon).

This is an efficiency issue. On a per ton-mile basis, rail emits only 20 grams of CO₂, compared to 80 grams for trucks. As such, decarbonization efforts have, by and large, focused on the trucking industry. Intuitively, this makes the most sense. Attack the largest market. Find a solution to the biggest problem. This has left the rail industry largely untouched and undiscussed by emerging energy startups and major players alike.

An Alternative Approach

Reports from the International Energy Agency have highlighted that switching from road to rail where possible could significantly reduce total emissions. A 10% shift in freight-ton miles from trucks to rail in the U.S. could save approximately 16M tons of CO₂ on an annual basis.

This is entirely possible. Currently, medium-haul and long-haul freight (250–500+ miles) accounts for an estimated 45–55% of total trucking ton-miles. As opposed to short-haul freight, which trucks are necessary for, these longer trips could theoretically be moved over to rail for immediate reductions in emissions.

Why this hasn’t already happened

The transition from truck to rail faces several critical bottlenecks. Current infrastructure isn’t built for efficient intermodal transfers — the systems needed to move goods between rail and truck require significant optimization and capital investment. In addition, just-in-time logistics companies, which form the backbone of modern supply chains, favor trucking for its flexibility and reliability.

Rail’s legacy infrastructure poses another challenge. Built primarily for bulk commodities like coal and grain, the existing rail network struggles with diverse cargo movement. Limited intermodal facilities and capacity constraints in high-traffic areas create operational friction. Service reliability compounds this — weather delays, maintenance requirements, and passenger train prioritization often make rail significantly slower than trucking on crucial routes.

While rail offers lower costs per ton-mile for long distances, the economics get complicated when accounting for intermodal handling and infrastructure requirements. Add to this the trucking industry’s deep integration into existing supply chains and substantial political influence, and you get a system highly resistant to change.

Potential Future State

Despite the challenges discussed, several major policy moves could accelerate the shift to rail. The Bipartisan Infrastructure Law has allocated $66 billion for rail modernization, targeting safety, capacity, and connectivity. States are driving change too — California’s Sustainable Freight Action Plan aims to transition cargo from roads to cleaner alternatives like rail.

The global picture reinforces this trend. The EU targets a 50% increase in rail freight by 2030. China and India are building dedicated freight corridors to handle growing cargo volumes.

Implementation remains the key challenge. Success depends on overcoming entrenched logistical systems, economic incentives, and political resistance. If these hurdles clear, rail could emerge as a compelling alternative for medium and long-haul freight, addressing both emissions concerns and infrastructure strain.

Voltify

In my opinion, and based on my analysis, Voltify’s bet is that rail will become a more critical piece of our freight infrastructure in the future. As a result, their focus is on becoming a critical piece of the next-gen energy infrastructure of the market they believe is set to benefit the most from changing policies, enabling them to grow as the market does.

If they’re successful and actively cement themselves as a trusted provider of green solutions to the rail market now, they’ll benefit more later.

In the rest of this deep dive, we’ll discuss their approach, the feasibility of the technology, and the challenges they’ll have to overcome to achieve this.

Founding Story

Voltify is founded by Daphna Langer and Alon Kessel, serial entrepreneurs whose past companies have a combined valuation of over $1B.

The company’s values, as listed on its website, are as follows:

Sustainability: We are committed to driving sustainability by reducing emissions, conserving resources, and fostering a cleaner, greener future for rail transportation.

Innovation: We embrace innovation as the cornerstone of our mission, pushing the boundaries of technology to create cutting-edge solutions for sustainable rail travel.

Collaboration: Collaboration is at the heart of our success. We work hand-in-hand with rail companies, communities, and partners to achieve mutual goals.

Resilience: We build resilience into our solutions, ensuring they can withstand challenges and disruptions, while also adapting to industry changes.

Customer-Centric Approach: We prioritize our customers’ unique needs, striving to provide tailored solutions that add value to their operations.

Financial Responsibility: Our approach is rooted in financial responsibility, offering sustainable solutions that make economic sense for our clients.

Safety: Safety is paramount in our operations, with a focus on maintaining secure and reliable rail transportation.

Ethical Practices: We uphold the highest ethical standards, practicing transparency, integrity, and strict adherence to all laws and regulations.

Their stated mission is to offer an end-to-end, zero-emission alternative to rail freight operators at diesel parity, meaning there should be no additional cost for adoption.

Product and Technology

Voltify’s approach is centered around key efforts across both technological infrastructure for energy-efficient propulsion, plus the infrastructure necessary to maintain it.

Their core product is the “VoltCar”, a battery-equipped freight car that can be plugged into existing rail infrastructure to propel the train. These batteries are designed to supply electric power to locomotives, effectively replacing the need for diesel fuel and thereby significantly reducing emissions. This approach allows for the utilization of existing locomotive fleets without necessitating extensive modifications.

Key features and functionalities of the VoltCar include:

Battery Composition: VoltCars utilize sodium-ion batteries, which are known for their cost-effectiveness and abundance compared to lithium-ion alternatives. Sodium-ion technology offers a promising balance between performance and sustainability, making it suitable for large-scale applications like freight rail.

Dynamic Charging: The system is designed to charge the VoltCars dynamically while they are en route and during yard operations. This fast-charging capability ensures that the batteries are replenished without necessitating prolonged stops, thereby maintaining operational efficiency.

Dual-Energy Propulsion: Voltify’s solution includes a dual-energy propulsion system that allows locomotives to seamlessly transition back to diesel mode if necessary. This flexibility ensures that operations can continue uninterrupted in scenarios where electric power may be insufficient or unavailable.

Regenerative Braking: The VoltCars are capable of charging continuously using regenerative braking. This means that energy typically lost during braking is captured and stored in the batteries, enhancing overall energy efficiency.

AI-Powered Energy Management: An AI algorithm forecasts data and analyzes real-time information to maintain energy consistency without compromising capacity. This intelligent system optimizes energy consumption, ensuring reliable operation and reducing maintenance costs.

To supplement the dynamic charging of the VoltCars, Voltify will build and establish microgrids along freight lines. Powered by solar farms and grid-independent, these systems will enable stationary, in-yard charging of the VoltCars, while also providing extra energy to primary grids while not in use.

Business Model

Voltify’s selling point to rail operators is simple. They won’t pay anything more than they already do using diesel, and all infrastructure development and integration of technology will be handled by Voltify.

This removes the “green-premium” that is a common barrier that blocks the transition to zero-emission technology.

With that said, it’s not immediately clear to me how Voltify plans on achieving cost parity.

Currently, rail operators pay approximately $0.09/kWh, the price of energy Voltify will have to match, while also maintaining room for profit.

To implement their technology, Voltify will have to take on significant expenditures spanning solar farm implementation, battery storage, and microgrid control systems. A theoretical cost breakdown (not exact — very “back of the napkin”) can be found below, with the assumption that all infrastructure is installed in railyards.

Note: The following cost analysis is based on industry benchmarks and my own calculations. Voltify has released limited public information about its exact approach and business model. While this analysis aims to understand the scale of investment required, actual implementation costs and Voltify’s specific strategy may differ significantly:

Solar Farms

Solar farms are the foundation of Voltify’s renewable energy model, providing consistent power generation for charging infrastructure and microgrids. Their capacity is tailored to the energy demands of the rail yard, with sizing dependent on the volume of VoltCars serviced daily and the regional solar irradiance.

Size:

• Medium yards: 5–10 MW capacity.
• Large yards: 20–30 MW capacity.

Cost per MW: $1.5–2 million.

Total Solar Farm Cost:

• Medium yard: $10–20 million.
• Large yard: $40–60 million.

Battery Storage

Battery storage ensures uninterrupted power delivery by balancing the inherent intermittency of solar energy. These systems store excess energy generated during peak sunlight hours and discharge it to meet demand when generation drops, enabling Voltify’s off-grid model to function reliably.

Capacity:

• Medium yard: ~10 MWh.
• Large yard: ~20–50 MWh.

Cost per kWh: $400–600.

Total Battery Cost:

• Medium yard: $6 million.
• Large yard: $12–30 million.

Charging Infrastructure

Charging infrastructure includes high-capacity fast chargers, transformers, and associated installations. These systems are essential for maintaining operational efficiency, allowing VoltCars to charge quickly without disrupting freight schedules.

Number of Chargers:

• Medium yard: 20–30 chargers.
• Large yard: 50–75 chargers.

Cost Per Charger: $200,000–300,000.

Transformers and Power Lines: $500,000–1,000,000 per yard.

Installation Costs: ~$1 million per yard.

Total Charging Infrastructure Costs:

• Medium yard: $7.5–11 million.
• Large yard: $17–25.5 million.

Microgrid Control Systems

Microgrid control systems integrate the solar, battery, and charging components into a seamless network, ensuring energy flows are optimized and interruptions minimized. These systems are critical to managing the complexity of Voltify’s off-grid infrastructure.

Cost: ~10% of total yard infrastructure costs.

Total Control System Cost:

• Medium yard infrastructure: Solar ($10–20M) + Batteries ($6M) + Charging ($7.5–11M) = $23.5–37M.
• Microgrid system: $2.35–3.7 million.
• Large yard infrastructure: Solar ($40–60M) + Batteries ($12–30M) + Charging ($17–25.5M) = $69–115.5M.
• Microgrid system: $6.9–11.55 million.

All of this adds up. Undertaking the responsibility of installing new infrastructure is expensive. This raises some inherent concerns.

A) It’s not clear what their approach is to funding this effort.

B) To achieve diesel parity and produce an eventual profit, they’ll likely have to generate energy at around $0.03/kWh, the extreme low end of solar farm generation.

Theorized Cost Mitigation Approach

It seems like Voltify’s model has to hinge on the ability to drastically lower capital expenses and operating costs.

To make this feasible, leaning on future developments of tech in the long term, and leaning on short term government incentive programs and strategic partnerships will likely play a pivotal role.

Here’s how Voltify may be approaching this:

Solar Energy Cost Trends

The cost of solar power has seen a dramatic decline over the past decade due to advancements in technology, increased manufacturing scale, and government incentives:

• Utility-Scale Solar Costs: The average cost of utility-scale solar power in 2023 was $0.03–0.06/kWh (National Renewable Energy Laboratory, NREL). This represents an 80% drop since 2010, when costs exceeded $0.15/kWh. Projections suggest further cost reductions of 15–25% by 2030, driven by more efficient photovoltaic cells, lower material costs, and innovations in solar farm deployment.
• Solar Farm Capital Costs: Current installation costs are $1.5–2 million per MW, down from $3 million/MW in 2015. The introduction of emerging technology like bifacial solar panels and robotic cleaning systems could reduce costs further by increasing efficiency and lowering maintenance expenses.

Battery Storage Cost Trends

Battery storage costs are a critical factor for Voltify’s microgrid systems, ensuring consistent energy availability for rail operations:

• Sodium-Ion Batteries: Sodium-ion technology, reportedly used by Voltify, is emerging as a cost-effective alternative to lithium-ion. These batteries are less expensive due to the abundance of sodium compared to lithium and have a lower environmental impact during manufacturing. While commercial deployment is still scaling, sodium-ion battery costs are expected to stabilize at $50–80/kWh by 2030.
• Utility-Scale Storage: Current utility-scale battery systems cost $400–600/kWh installed, but this is expected to fall to $200–300/kWh by 2030, further improving project economics.

Government Incentive Programs

Government initiatives like the Inflation Reduction Act (IRA) and various state-level programs provide substantial opportunities to reduce Voltify’s upfront costs and improve financial viability:

• Investment Tax Credits (ITC): The IRA offers a 30% ITC for renewable energy projects, including solar farms and battery storage. This could significantly reduce infrastructure costs. Example: For a 20 MW solar farm costing $30 million, ITCs could save ~$9 million upfront.
• Production Tax Credits (PTC): Voltify could earn ~$0.025 per kWh of solar energy produced. Example: A 20 MW system generating 35,040 MWh annually could receive ~$875,000 per year, offsetting operational expenses.
• Battery Storage Incentives: With batteries now qualifying for ITCs, Voltify could save ~30% on battery costs. For a 10 MWh system costing $6 million, this equates to a $1.8 million reduction.
• Carbon Credits: Voltify could monetize emissions reductions through carbon markets, earning $20–40 per ton of CO₂ saved. Example: Saving 100,000 tons of CO₂ per yard annually could generate $2–4 million in revenue per year.
• Freight Infrastructure Grants: Programs like BUILD (Better Utilizing Investments to Leverage Development) and RAISE (Rebuilding American Infrastructure with Sustainability and Equity) could provide direct funding for infrastructure projects.

Strategic Public-Private Partnerships (PPPs)

Voltify could mitigate costs further by partnering with government agencies and rail operators to share risks and benefits:

• Cost Sharing with Rail Operators: Rail companies could co-invest in microgrid development, leveraging the long-term savings from reduced maintenance and carbon credits.
• Subsidized Financing: Accessing low-interest loans through the Department of Energy’s Loan Programs Office (LPO) or similar entities could lower the financial burden of upfront costs.

Revenue Opportunities from Surplus Energy

Voltify’s solar farms, installed in rail yards, could generate excess energy during non-peak freight operations. Selling this surplus energy to the grid creates an additional revenue stream:

• Revenue Potential: For a 20 MW solar farm generating 35,040 MWh/year, selling 25% of surplus energy at $0.04/kWh could yield ~$350,000 annually per yard.

Across 1,500 rail yards, this could generate $525 million/year.

Wrap Up

In sum, this could result in the following total reductions in cost:

• Solar Costs: Declining costs could reduce Voltify’s per MW installation costs by 20–25%
• Battery Storage Costs: Sodium-ion adoption and ITC incentives could lower battery costs by 50–80%
• Government Programs: Incentives like ITCs, PTCs, and carbon credits could offset 30–40%
• Revenue from Surplus Energy: Selling excess solar power could offset long-term costs by generating $350,000 per yard annually, amounting to $525 million per year across 1,500 yards.

I want to add a disclaimer here. Voltify has released very limited public information on their progress or plans, so all of the above is pure, analysis-based theory. That said, based on it, there are a few more considerations I want to make.

Limited Funding: Voltify has raised only $2.5 million to date, a modest amount compared to the capital-intensive nature of its proposed infrastructure. Even with future reductions in cost and incentive programs, the scale of development — spanning solar farms, battery storage, microgrid systems, and dynamic charging infrastructure — requires substantial investment well beyond their current funding.

Resource Constraints: Supporting a team of approximately 20 employees, Voltify’s current operational capacity is likely focused on research, development, and early-stage business planning rather than large-scale implementation.

Funding Requirements for Scaling: To transition from concept to reality, Voltify will need to secure:

• A significantly larger funding round from venture capital, strategic investors, or infrastructure-focused funds.
• Debt instruments, such as green bonds or loans, to finance large-scale infrastructure.
• Government grants or subsidies, which could offset a substantial portion of upfront costs if aligned with federal and state decarbonization initiatives.

Competitive Landscape

The race to decarbonize freight transport is accelerating across multiple fronts. While rail offers inherent efficiency advantages, competing technologies and approaches are evolving rapidly. Let’s examine each potential threat to Voltify’s market position.

Battery-Electric and Hydrogen Trucking

The trucking industry is investing heavily in electrification. Companies like Tesla, Volvo, and Daimler are pushing battery-electric trucks beyond previous limitations, achieving ranges of 300–500 miles with improving economics. Total ownership costs are projected to reach diesel parity by 2030, supported by $7.5 billion in federal funding for charging infrastructure.

Hydrogen fuel cell trucks, led by Toyota and Hyundai, present another compelling alternative. Their longer range and faster refueling could make them ideal for long-haul routes where battery-electric trucks struggle. While both technologies face significant infrastructure challenges — particularly in rural areas — their continued advancement could eventually challenge rail’s cost advantages for long-distance freight.

Alternative Rail Decarbonization Methods

Within rail itself, several competing approaches warrant attention:

1. Renewable diesel offers immediate emissions reductions of 50–90% without major infrastructure overhaul. Its compatibility with existing locomotives makes it attractive for rapid deployment, but limited supply and high costs restrict its scalability.
2. Hydrogen fuel cell trains, already operational in parts of Europe, demonstrate zero-emission capability for non-electrified lines. However, the economics remain challenging — infrastructure costs and locomotive conversion expenses create significant barriers to widespread adoption.
3. Traditional catenary electrification has proven successful globally but faces unique challenges in the U.S. The massive cost of retrofitting existing freight corridors makes this solution less viable in the American market.

Emerging Disruptions

Looking further ahead, autonomous trucking could fundamentally reshape freight economics. Companies like Waymo and Aurora are working to eliminate driver costs, which currently represent 40–50% of trucking expenses. The ability to operate continuously without human limitations could significantly narrow the efficiency gap between rail and road for long-haul routes.

More speculative technologies like Virgin Hyperloop’s 600 mph system represent potential long-term disruption. While currently cost-prohibitive and unproven at scale, these innovations could eventually compete for high-value, time-sensitive cargo.

Policy and Integration Advantages

Government initiatives could provide tailwinds for rail adoption. The EU’s aggressive Green Freight Strategy, targeting a 30% shift from road to rail for long-distance freight by 2030, demonstrates growing policy support for rail transport. Similar U.S. initiatives focusing on intermodal infrastructure development suggest increasing recognition of rail’s role in sustainable freight.

Complementary innovations in logistics are making rail more competitive. Companies like J.B. Hunt are optimizing intermodal transfers, while AI platforms reduce coordination friction and idle time. These advancements help address traditional pain points in rail freight adoption.

Critical Success Factors

For Voltify to succeed in this dynamic landscape, several factors appear critical:

• Rapid deployment to secure market share before competing technologies mature
• Seamless integration with existing rail operations and intermodal systems
• Cost advantages that persist even as competing technologies improve
• Ability to leverage policy support while maintaining commercial viability

Challenge and Opportunities

Challenges

High Capital Requirements

The capital needed for Voltify’s vision is staggering (unless I’m completely wrong in my assumptions). Each rail yard implementation demands tens of millions of dollars for solar farms, battery storage, and microgrid systems. Their current $2.5M in funding doesn’t cover even a pilot deployment. Scaling nationwide would require potentially hundreds of billions through venture capital, debt financing, and government subsidies.

Cost Parity and Profitability

The economics are challenging. Achieving diesel parity at $0.09/kWh requires generating power at $0.02–0.03/kWh — a difficult target even with advanced solar technology. Energy price fluctuations and potential yard underutilization threaten profitability while trying to cover massive infrastructure costs.

Implementation Challenges

Physical deployment faces significant hurdles. Land near rail yards is often limited or expensive, and local permitting processes could create substantial delays. Their reliance on dynamic charging and regenerative braking may struggle to meet energy demands in high-traffic corridors.

Evolving Technological Threats

Competition is emerging from multiple directions. Electric and hydrogen trucks could eventually negate rail’s efficiency advantages by offering zero-emission alternatives with greater flexibility. Meanwhile, hydrogen rail solutions might provide a simpler path to decarbonization, requiring fewer systemic changes.

Market and Regulatory Barriers

The fragmented nature of U.S. freight rail complicates adoption. With multiple operators like Union Pacific, BNSF, and CSX each having different priorities, coordinated implementation becomes complex. While government support exists today, policy incentives remain subject to political shifts.

Integration Across a Complex Network

Integrating across 140,000 miles of rail infrastructure presents enormous challenges. The variety in terrain, traffic density, and existing infrastructure conditions means significant customization would be required, further complicating scalability.

Opportunities

Market Timing and Urgency

The momentum behind decarbonization is accelerating, with rail becoming increasingly central to emissions reduction strategies. Voltify’s position as a first mover in zero-emission rail freight is well-timed. The Inflation Reduction Act provides substantial tax credits and funding aligned directly with their infrastructure needs.

Cost-Reduction Trends in Renewable Energy

The economics continue to improve. Solar and battery costs have dropped over 80% since 2010, with further declines expected. Voltify’s reported use of sodium-ion batteries could provide additional cost advantages over traditional lithium-ion technology, improving their path to profitability.

Complementary Modal Shifts

Policy support for rail transport is growing. The EU’s Green Freight Strategy targeting 30% shift from road to rail signals broader momentum. U.S. Department of Transportation initiatives supporting intermodal facility development suggest increasing recognition of rail’s role in sustainable freight.

Scalability Through Rail Yards

The focus on rail yards as deployment sites simplifies implementation compared to network-wide solutions. These fixed locations with established infrastructure provide natural nodes for energy generation and distribution, creating a clearer path to scale.

Revenue from Surplus Energy

With over 1,500 rail yards in the U.S., surplus energy generation during non-peak freight operations could create significant additional revenue through grid sales. This secondary income stream strengthens the overall business model.

Synergies with Rail Operators

Voltify’s infrastructure development and maintenance service model provides compelling value for operators looking to reduce costs and emissions without capital expenditure. The operational savings from reduced maintenance and fuel expenses create natural adoption incentives.

Conclusion

I’m writing about Voltify because of their vision. It’s compelling, and they’re doing something I haven’t seen anyone else think of or work towards. That said, success hinges on incredible execution. Scaling from concept to deployment and implementation requires navigating immense capital requirements, complex logistics, and an evolving competitive landscape, while also relying on a variety of shifting factors in future markets.

The team has proven technically capable, and market timing appears favorable with the growing momentum behind rail decarbonization. However, the narrow path to success demands flawless execution across multiple dimensions — from securing massive funding to coordinating with fragmented rail operators to hitting aggressive cost targets.

That said, this is the kind of ambitious infrastructure play needed to address climate challenges. Whether they can overcome the substantial hurdles ahead remains to be seen, but they’re asking the right questions and tackling the problem at the necessary scale.

I’ll continue to update this article as rounds are raised and progress is made.

Thanks for reading.

Updates to this article will be made with every subsequent funding round. Stay tuned.

I put out a deep-dive like this one every week. If you enjoyed, follow me here.

If you’re interested in contributing to Garden Research as a writer, email me at pjlabarbera11@gmail.com.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

How to Decarbonize Rail Freight was originally published in Garden Research on Medium, where people are continuing the conversation by highlighting and responding to this story.

For weekly updates and musings on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

“China’s military is Netflix; the U.S. military is Blockbuster. China is Amazon; the U.S. is Barnes & Noble. China is Tesla; the U.S. is General Motors. People thought the automotive industry could not be disrupted and that Ford, GM and Toyota had preordained rights to be the leaders for the next thousand years. Automotive executives mocked and balked at the idea an outsider could do what they do. They spent billions on research and development; they allocated billions in resources and thousands of people to their products and capabilities. Yet here we are today, after less than a decade: Tesla is a $1T company; it had dethroned the incumbents”

This quote can be found on the home page of autonomy and defense company Shield AI’s website. It’s inflammatory. It’s supposed to be.

Shield AI, along with a slew of other startups and investors, is addressing a need in the market that’s becoming increasingly relevant: technology that protects America’s national interests, deters conflict, reinforces American industrialization, and positions us to continue being a highly competitive influence on the geopolitical international stage.

Recently, it’s been revealed that Palantir is in advanced talks to participate in an upcoming funding round of ShieldAI for $200M at a valuation of $5B. Today, we’ll be taking a closer look at this partnership, Shield AI’s offering, and what this example, plus other funding data, tells us about the emerging focus on “American Dynamism”.

Shield AI

Shield AI was founded in 2015 by former Navy Seal Brandon Tseng, his serial entrepreneur brother Ryan Tseng, and perception/navigation engineering wiz Andrew Reiter.

The company specializes in unmanned, autonomous systems for Defense applications. Its current clientele list includes the Ukrainian and United States governments, the U.S. Special Operations Command, the USAF, the USMC, the US Navy, and several allied international militaries.

Product Suite

Shield AI’s foundational product is HiveMind, a fully autonomous aircraft pilot that can read and react to changing battlefield conditions in real-time — without relying on GPS, marked waypoints, or external comms tools. Their systems are trained in synthetic, physics-modeled environments combined with human-machine tactics and behaviors, leveraging a suite of proprietary sensors and cameras to create a reliable world-view and map that informs the aircraft on how to react, read, and respond to changing environments.

The significance of HiveMind’s GPS-independent operation cannot be overstated. In modern warfare, GPS denial is a common tactic — Russia has actively demonstrated this in Ukraine. By operating without GPS dependency, Shield AI’s systems maintain effectiveness in contested environments where traditional autonomous systems would fail.

HiveMind is offered as an SDK, allowing companies to rapidly implement autonomy into their hardware by providing reference architectures, autonomy libraries, and simulation/testing/analysis tools. Using this SDK, and in partnership with DARPA’s ACE program, ShieldAI “took over” an F-16 and X-62 Vista earlier this year, enabling it to carry out fully autonomous dogfighting and mission-specific capabilities. (Spoiler: it beat the human 5–0)…

In another, quite frankly remarkable feat, ShieldAI successfully went from project beginning to live flight on a Kratos Firejet in just 120 days, exemplifying their speed of implementation and proficiency in platform flexibility.

In addition to the software, ShieldAI is manufacturing several proprietary vehicles with HiveMind pre-built in, including:

V-Bat

A VTOL drone revolving around a patented ducted fan design that increases thrust by over 80% at an equivalent engine power of competitors. The V-Bat is edge-enabled, running off NVIDIA Orin GPUs, and will be used to create swarms of unmanned aircraft to carry out specific missions like “find, tag, kill”, and “missile sponge” missions.

Nova-2

A small-scale quadcopter drone used for tactical indoor room clearing, surveillance, and reconnaissance. It was named as the “most mission capable indoor drone” in the world built for indoor room clearing by joint US and Israeli forces.

Palantir’s Involvement

In December, Shield AI announced that they’d be expanding work with Palantir Technologies, a leading provider of AI systems, to develop and deliver large-scale command and control of autonomous uncrewed systems.

ShieldAI will receive prioritized access to Palantir’s operation Warp Speed, their proprietary manufacturing OS that provides an adaptable platform providing the speed, flexibility, and security required for the modern manufacturer.

Along with Shield AI, Palantir is providing Warp Speed to several companies working to bolster American defense and manufacturing efforts, including Anduril Industries, L3Harris, and Panasonic Energy of North America. Currently, ShieldAI is using warp speed to improve the speed of iteration and manufacturing of the V-Bat, for which they’ve received record demand.

Now, as of last week, Palantir is in advanced discussions to invest in Shield AI’s upcoming $200M round.

Why This is Important

Here’s my take on it. Palantir’s activities within Warp Speed, and this investment, can be judged as an example of where the overall market may be headed — not just in Defense, but in Energy, Infrastructure, Manufacturing, and other critical areas important to national security and public safety.

The investment in ShieldAI is part of a broader, focused, intentional investment strategy for Palantir. The company has disclosed holdings in various companies, with a recent portfolio value calculated to be $9,719,645. Top holdings include:

• Surf Air Mobility Inc. — a company redefining American air travel and electrification of transport
• Rubicon Technologies, Inc. — a company redefining American waste management
• MSP RECOVERY INC — a company redefining American Healthcare payment recovery

What do these have in common?

They represent a shift towards American manufacturing, energy, defense, and critical infrastructure development.

It’s not surprising, in my opinion. The global threat of conflict is rising, and there is a growing narrative that increasing tech enablement across critical industries relevant to American protectionism may be necessary to combat adversarial entities.

Supporting Data

This is a graph of total capital deployed into American drone companies by year.

What do you notice?

Let’s dive deeper. Of the approximately 1200 American startups that fall into this category, 272 (~23%) operated at the intersection of drone tech and Defense.

Look at this spike in deal count for those companies.

Interesting…

What about companies working to redefine American manufacturing through tech-enabled solutions?

Hmmm.

There’s a trend shaping here, and you don’t need to take my word for it. On A16z’s 2025 “requests” for startups, one of the few mentioned categories “American Dynamism” (see below). You can read more in the hyperlink.

What does this suggest?

The shift we’re seeing isn’t just about defense technology. It’s about rebuilding American industrial capability, national security, and infrastructure from first principles.

Shield AI and Palantir’s partnership represents one example of a much broader trend: Silicon Valley is turning its attention to foundational infrastructure.

While the last decade of startup funding may have focused on consumer apps and enterprise SaaS, this new wave is taking on the hard problems. Defense tech, domestic manufacturing, energy independence, supply chain resilience, the space economy… areas that were previously left to aging incumbents or were thought of as “too risky” are now attracting both startup talent and venture dollars.

The implications are significant. Silicon Valley’s next chapter might not be defined by social networks or food delivery apps but by companies rebuilding American industrial and technological superiority. In this context, a $5B valuation for Shield AI isn’t just a bet on autonomous systems — it’s an early indicator of where smart money sees the future of American technology heading.

The market data tells the story. VC investment in American Dynamism is accelerating. Shield AI might be today’s headline, but they’re just the beginning.

For a final example of market traction, here’s a list of several early-stage companies operating within Manufacturing or Defense that raised Angel/Pre-seed/Seed/Early Stage rounds in 2024.

Defense

• Delta Black: Hybrid VTOL drone with 28+ hour endurance
• General Hypersonics: On-demand, reusable hypersonic launch systems for rapid payload delivery
• Guardian RF: AI-powered systems for real-time detection and tracking of drones and their operators
• Swarmer: AI-powered autonomy solutions for coordinated UAV operations
• DarkHive: Rapidly deployable, short-range tactical drone designed for indoor and outdoor threat assessment
• ZeroMark: AI-enabled auto-targeting fire control systems
• Primordial Labs: Natural language interface enabling intuitive control of uncrewed systems
• Firestorm Labs: Modular, 3D-printed unmanned aerial systems (UAS) designed for rapid production and mission adaptability
• HavocAI: End-to-end maritime autonomy solutions
• Gambit: Platform-agnostic AI solutions for scalable, autonomous unmanned systems

Manufacturing/American (Re) Industrialization

• AIM: Retrofits existing heavy equipment to operate autonomously
• Eagle Electronics: American manufacturer specializing in cellular modules
• SirenOpt: Revolutionizes manufacturing by enabling real-time, non-destructive materials analysis
• Deterrence: Autonomous manufacturing for critical energetic defense components
• Bedrock: Develops and manufactures active materials for sodium-ion batteries
• Fabri: Accelerates manufacturing with rapid, on-demand casting production
• Fortius Metals: Enhances manufacturing with stronger, lightweight metal fabrication.

Thanks for reading.

I put out content like this every week. If you enjoyed, follow me here.

If you’re interested in contributing to Garden Research as a writer, email me at pjlabarbera11@gmail.com.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Palantir, Shield AI, and “American Dynamism” was originally published in Garden Research on Medium, where people are continuing the conversation by highlighting and responding to this story.

In the months leading up to October 2024, a high-stakes game of Startup Bachelorette was playing out in Silicon Valley. Over 200 venture capitalists were scrambling, competing for the attention and a coveted spot on the cap table of one promising company.

What could garner the attention of so many VCs at once?

The company in question is Lumen Orbit. At the end of the process, in a round led by NFX with participation from FUSE, Soma Capital, and the scout funds of Andreesen Horowitz and Sequoia, they raised $11M to address one of the most critical issues affecting the scaling potential of AI — the limitations of terrestrial data centers.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Key Gap Served

Data Centers

Data centers are critical to the growth and future of Artificial Intelligence, serving as the backbone of modern AI infrastructure. They house the essential hardware and networking systems required for the immense computing power, data processing, storage, and connectivity demanded by today’s AI models for both training and inference.

The scaling of the AI market directly correlates with the expansion of the data center market. Over the past five years, the number of leased data centers has grown 17-fold. To sustain the rapid pace of AI innovation and adoption, it’s estimated that the U.S. will require an additional 50 to 60 GW of data center capacity within the next five years — effectively doubling current capacity.

To put this into perspective: a typical large enterprise data center consumes 20 to 50 MW. Achieving 50 to 60 GW of new capacity would necessitate the construction of 2,500 to 3,000 new data centers.

Scaling to this level presents one primary challenge: energy demand.

Energy Issues

Traditional data centers are voracious consumers of energy, requiring staggering amounts of power to keep their servers, cooling systems, and networking infrastructure running. AI-specific data centers, designed with denser server racks and specialized hardware like energy-intensive GPUs, amplify this demand even further, pushing energy consumption to new heights.

Some stats to help frame this:

• Capacity and Energy Consumption: As discussed above, data center capacity is expected to double over the next five years. This growth will be accompanied by a substantial increase in electricity consumption, with data center electricity use projected to grow by an average of 23% annually between 2023 and 2028. AI-specific data center energy usage is expected to grow even faster, at an average of 43% annually over the same period.
• Current and Future Consumption: Currently, data centers account for 3–4% of total U.S. power consumption. This figure is anticipated to reach 11–12% over the next five years, fueled largely by the expanding AI market. By 2030, electricity demand for U.S. data centers is projected to grow by approximately 400 TWh, representing a compound annual growth rate (CAGR) of 23%.
• Financial Ramifications: The critical power required to support key components such as GPU and CPU servers, storage systems, cooling, and networking switches is expected to nearly double by 2026, reaching 96 GW globally. AI operations alone could account for over 40% of this demand. Meeting this growth will require substantial financial investment, with U.S. utilities needing to allocate approximately $50 billion for new generation capacity. Additionally, data centers will drive an estimated increase of 3.3 billion cubic feet per day in natural gas demand by 2030, necessitating expanded pipeline infrastructure.

It’s important to note that the projected surge in data center energy demand coincides with unprecedented growth in energy consumption from other sectors. The electrification of transportation, the expansion of renewable energy infrastructure, and the rising energy needs of modern manufacturing are already placing significant pressure on energy markets.

To handle this demand, there needs to be a massive, massive increase in energy production capabilities. Unfortunately, there are critical, possibly insurmountable issues with current approaches.

Current Approaches and Limitations

To ensure grid reliability and minimize disruptions, data centers often rely on a mix of grid-based electricity and locally produced energy. Many data centers incorporate on-site generation sources, such as natural gas generators or renewable energy systems like solar and wind, to provide backup power or supplement grid energy during peak demand periods. Let’s look at each energy source critically.

Natural Gas

Natural gas faces several critical challenges as a reliable energy source for data centers. Infrastructure expansion presents a significant hurdle — building out pipelines and distribution networks is both expensive and politically complex. While cleaner than coal, natural gas still contributes substantially to greenhouse gas emissions, creating tension with corporate sustainability goals. The volatility of natural gas prices, driven by geopolitical factors and seasonal demands, makes long-term energy planning particularly challenging for data center operators.

Grid Dependency

Traditional power grids are increasingly strained by the growing energy demands of data centers. Many grid systems simply weren’t designed to handle this level of power consumption, leading to increased risks of blackouts and brownouts. The push toward grid decarbonization, while necessary, has introduced additional complexities. Higher reliance on intermittent renewable energy sources has made it more difficult to maintain stable power delivery. Modernizing grid infrastructure to meet future demands will require massive investment, extensive coordination, and significant time — resources that may not be available quickly enough to meet accelerating data center growth.

Solar and Battery Energy Storage

Solar power, while promising, faces fundamental limitations when applied to data center operations. The inherent intermittency of solar generation — the simple fact that solar panels can’t produce power at night or during heavy cloud cover — makes it unreliable as a standalone solution. Land requirements pose another significant challenge, particularly in densely populated areas where securing sufficient space for large-scale solar farms near data centers is often impossible. Weather variability further complicates matters, making it difficult to guarantee consistent energy supply for critical data center operations.

The integration of battery storage systems, while necessary to address solar’s intermittency, introduces its own set of challenges. Current battery technology, primarily based on lithium-ion systems, remains prohibitively expensive at the scale required for data centers. These systems face significant degradation over time, necessitating regular replacement and driving up long-term operational costs. The production of batteries relies heavily on critical minerals like lithium, cobalt, and nickel, all of which face supply chain constraints and environmental concerns. The physical space required for large-scale battery systems compounds the land use issues already present with solar installations.

Nuclear

Nuclear power presents a complex value proposition for data center energy needs. While capable of providing substantial base-load power, nuclear facilities require enormous upfront capital investment and face extensive regulatory hurdles. Public resistance and safety concerns further complicate deployment timelines. Though excellent for consistent power output, nuclear plants struggle with load flexibility, potentially limiting their effectiveness for data centers with variable demand patterns. The scale of nuclear operations, while impressive, may not align well with the distributed nature of data center deployments.

Wind Energy

Wind power, similar to solar, struggles with reliability and consistency issues. Its dependence on weather conditions makes it an unreliable primary power source for data centers. The remote location of most wind farms necessitates significant investment in transmission infrastructure to deliver power to data center locations. Public opposition to wind turbines, driven by concerns about noise and visual impact, can significantly delay or prevent deployment in many areas.

Battery Storage as a Standalone Solution

Attempting to use battery storage systems independently of renewable generation presents its own challenges. The capital requirements for large-scale battery installations remain prohibitive, even without accounting for associated solar or wind infrastructure. Energy efficiency losses during storage and retrieval reduce overall system effectiveness. Current battery technology’s limited discharge duration — typically just a few hours — makes it insufficient for sustained data center operations without additional power sources.

What’s next?

A breakthrough is needed to overcome the looming energy crisis for data centers. Hundreds of companies are racing to solve critical challenges in battery technology, solar generation, nuclear power, and beyond. Many are well-funded and may bring significant advancements. Yet, even with the most optimistic projections, we are likely to reach a critical inflection point — a moment when terrestrial-based data centers may no longer be viable for sustaining AI’s exponential growth.

There’s only one company actively addressing this: Lumen Orbit.

Founding Story

The founders of Lumen Orbit are pursuing what, upon reflection, seems like a natural solution to the growing energy crisis: look to space.

Where is power abundant and consistently available? Space. Above Earth, the limitations of terrestrial solar energy — day/night cycles, weather disruptions, and energy loss through the atmosphere — vanish. Additionally, the immense cooling demands of data centers are resolved with passive cooling, utilizing the naturally cold vacuum of space.

Addressing these two challenges alone drastically reduces operational costs and significantly enhances scalability. Early calculations from Lumen Orbit indicate that solar power in space can achieve energy yields approximately 40% higher than terrestrial solar farms at peak performance. When accounting for consistent exposure to sunlight, this translates to generating over five times more energy than Earth-based systems.

Of course, it’s not as I’m making it sound. Fortunately, I’m not the founder. These guys are:

Ezra Feilden, PhD: With a background in aerospace engineering and data systems, Ezra brings critical expertise in building and deploying complex systems in the harsh environment of space. His experience with deployable structures, honed at Airbus Defence & Space and Oxford Space Systems, directly applies to the large-scale platforms Lumen Orbit envisions.

Adi Oltean: As a former Principal Software Engineer at SpaceX, Adi’s work on the Starlink network provides invaluable insight into managing large-scale data transmission in space. His expertise is essential for Lumen Orbit to efficiently handle the massive data flows generated by AI applications.

Philip Johnston: As CEO, Philip is tackling two of the fundamental challenges of space-based data centers: creating scalable, adaptable structures and ensuring efficient heat dissipation in the vacuum of space. His experience with satellite projects for national space agencies while at McKinsey & Co. equips him with a broad understanding of the space industry’s complexities, and he is a 2x founder with an exit under his belt (Opontia).

The challenges are undoubtedly significant, but as Lumen Orbit’s whitepaper states:

“While the challenges to these spacecraft will be substantial, working from first principles, Lumen Orbit has developed a range of concept designs and has not found any insurmountable obstacles. With new, reusable, cost-effective heavy-lift launch vehicles set to enter service, combined with the proliferation of in-orbit networking, the timing for this opportunity is ideal.”

Technology and Product

Principles

Lumen is bringing data centers to space. Their approach to design is based on five key principles:

Modularity: The data centers will be designed around independent modules that can be docked or undocked as needed, allowing for flexibility in upgrading hardware, replacing faulty systems and components, and scaling capacity with containers that have different computing abilities.

Maintainability: The ease of replacing parts without causing significant disruption to operations is a key consideration. This will ensure the longevity of the data center and reduce downtime.

Minimization of Moving Parts and Failure Points: The use of connectors, actuators, latches, and other moving parts should be minimized. In addition, Lumen Orbit is designing each container to have one single universal port to combine power, network, and cooling plug-ins.

Design Resiliency: The above naturally results in the elimination of single points of failure. Any failure of the system should be a result of what Lumen calls “graceful degradation of performance”, or a decrease in performance slowly over time.

Incremental Scalability: The modularity of the system will result in the ability to scale the number of containers in the system from one to N. Profitability should be achieved from the first container, and CapEx increases should be consistent as scaling occurs.

Technology

Solar Arrays and Cell Selection

Lumen’s systems will harness solar power by way of large-scale solar arrays, approximately 4km x 4km in size. Ensuring performance is optimized, the effects of UV, reduced efficiency at high temperatures, and ionizing radiation must be mitigated by careful cell selection. Lumen will use thin-film silicon solar cells for several key reasons:

• The cells use silicon wafers less than 25 micrometers in thickness, resulting in a power density of over 1000 watts per kg. This significantly reduces the weight of the solar arrays even at large scale — a critical factor in reducing launch costs tied to total mass
• Thin-film solar cells maintain high efficiency over their lifetime, even when exposed to intense radiation. The cells have a unique ability to heal from radiation damage at moderate temperatures, allowing for long-term operation in space.
• The thin and flexible nature of the cells allows them to be folded and rolled into compact configurations during launch. Once in orbit, they can be rolled out and expanded.

Cooling Infrastructure

The vacuum of space prevents conventional cooling methods from working properly. Luckily, space is naturally very, very cold (-270 degrees Celsius). Lumen leverages this to act as a natural pathway for passive cooling, eliminating the need for expensive, energy-intensive chillers or liquid cooling systems used by terrestrial data centers.

Instead, the space data centers will use lightweight radiators with large surface areas to dissipate the massive heat load generated by computing modules, radiating it outwards into space.

To transfer the thermal load from the modules to the radiators, Lumen will use cooling loops, using two-phase systems where practical to reduce mass flow requirements and reduce pumping losses. In addition to this, Lumen proposes using one of two methods to cool compute modules to maintain operating temperature while achieving high power densities:

• Direct-to-Chip Liquid Cooling: Delivers coolant directly to the heat-generating components
• Two-Phase Immersion Cooling: Submerges computing hardware in a dielectric liquid that undergoes a phase change from liquid to vapor as it absorbs the heat generated by computing modules. This allows for even higher heat dissipation while achieving greater space efficiency.

Physical Architecture

The physical architecture of Lumen Orbit’s data centers is based on a central spine with radiating solar arrays and radiators. This design provides a stable platform for the compute modules while ensuring efficient energy generation and heat dissipation.

• Central Spine: The spine serves as the backbone of the data center, providing structural support and housing the main power distribution and network infrastructure.
• Radiating Solar Arrays: The solar arrays extend outward from the spine, maximizing exposure to sunlight and ensuring continuous energy generation.
• Radiating Radiators: The radiators are also positioned to radiate outward from the spine, effectively dissipating the heat generated by the computing modules.

This design allows for the compact storage and efficient deployment of large structures, crucial for space-based applications. High-voltage DC (HVDC) lines are used to transfer power from the solar arrays to the compute modules, minimizing transmission losses.

Network Architecture

Lumen Orbit’s network architecture will be designed to meet the specific demands of AI workloads requiring low latency and high bandwidth for training and communication.

To achieve low latency, containers will be deployed close together in a tightly coupled daisy-chain configuration, aiming to achieve sub-millisecond latency between containers, compared to typical inter-data center latencies of 5–10 milliseconds on Earth. This reduction in latency is crucial for AI training workloads, where model parallel training requires constant communication between processing nodes.

To support bisection bandwidth for training large AI models, the infrastructure will contain multiple layers of directories and switches appropriately sized for the workloads the data center is designed to execute.

Connectivity and Data Transfer

Lumen will use a dual approach to data transfer and connectivity, catering to both immediate and bulk data needs.

For real-time data transmission, the data centers will utilize laser-based communication links with existing satellite constellations such as Starlink, Kuiper, or Kepler. For transferring larger volumes of data, Lumen envisions using data shuttles — small modules that can dock into the system to transfer petabytes, or even exabytes, of data in a single trip.

Product Considerations/Strategy

Scalability

The design of these space-based data centers results in significant increases in scalability as compared to terrestrial options. Unlike terrestrial data centers, which are constrained by physical space and planning regulations, orbital data centers can be scaled almost indefinitely.

This enables the construction of data centers with power generation well into the gigawatt range, an accomplishment necessary to achieve the massive leaps in data center infrastructure needed for AI development. In addition, unlike terrestrial data centers, which are limited by 2D land area, orbital data centers can scale in three dimensions. This allows for a more tightly coupled cluster with lower latency, a critical factor for AI training.

Orbit Selection

To achieve the necessary power generation while mitigating potential challenges associated with space debris and access to sunlight, Lumen Orbit will deploy its data centers in low-Earth, sun-synchronous orbit. This orbit follows the dawn-dusk line, allowing for uninterrupted sunlight, and uninterrupted power generation. In addition, the SSO offers a good balance between solar energy capture and radiation exposure.

Business Model

Lumen Orbit aims to provide space-based data center solutions tailored for high-performance computing, particularly AI workloads. By leveraging the unique environment of space, they offer continuous solar energy, efficient passive cooling, and virtually unlimited scalability. This positions them to meet the growing demands of AI applications more sustainably and cost-effectively than traditional land-based facilities.

Launch Plans and the Impact of Decreasing Launch Costs

The company’s deployment strategy is closely aligned with advancements in reusable heavy-lift launch vehicles, which are reducing the cost of accessing space. By capitalizing on these cost-effective launch options, Lumen Orbit plans to deploy data center modules that are competitive with terrestrial alternatives. Key benefits include:

• Reduced Deployment Costs: With launch costs targeting $5 million per 40 MW data center module, Lumen Orbit can offer highly competitive services.
• Increased Payload Capacity: Improved launch vehicles allow for larger, more powerful modules, enhancing scalability and efficiency.
• Faster Deployment: Frequent launch schedules enable rapid infrastructure deployment to meet market demands.

The decreasing cost of launching infrastructure is critical for Lumen Orbit’s long-term success, allowing the company to focus on its core mission: delivering a superior data center solution without being hindered by prohibitive launch expenses.

Financial Model and Cost Comparisons

Lumen Orbit’s approach offers several financial advantages over terrestrial data centers:

• Lower Energy Costs: Continuous solar power in space results in an estimated energy cost of ~$0.002/kWh — 22 times lower than average energy prices in the US, UK, and Japan.
• Reduced Cooling Expenses: Passive cooling in the vacuum of space eliminates the need for energy-intensive chillers.
• No Water Usage: Unlike land-based data centers that rely on water for cooling, space-based centers offer both cost savings and environmental benefits.

These factors lead to a significantly lower total cost of ownership, especially for large-scale, energy-intensive AI workloads. A cost comparison for a 40 MW cluster over 10 years shows a clear financial advantage for space-based solutions.

Competitive Analysis

Lumen Orbit’s vision for space-based data centers positions it uniquely within the data center landscape.

1. Terrestrial Data Centers: Lumen Orbit eliminates the limitations of land-based centers, such as land scarcity, escalating energy costs, and grid dependency. While terrestrial centers offer established infrastructure and lower latency for local users, they face mounting challenges in supporting AI workloads. Lumen Orbit’s scalability, reduced energy costs, and innovative design tackle these issues head-on.
2. Underwater Data Centers: Initiatives like Microsoft’s Project Natick and Network Ocean leverage underwater environments for cooling but remain tied to grid power or localized renewable sources. Lumen Orbit’s continuous access to solar energy in space provides a distinct edge in power generation and scalability, free from location-based constraints.
3. Technological Innovations for Terrestrial Data Centers: Advances in liquid cooling, energy-efficient hardware, and AI-driven energy optimization offer incremental improvements but fail to address the core limitations of Earth-based infrastructure. Lumen Orbit provides a fundamental shift, leveraging the unique benefits of space for sustainable, large-scale computing.
4. Decentralization and Edge Computing: While decentralized models and edge computing reduce latency and improve local processing, they lack the capacity for large-scale AI workloads. Lumen Orbit complements these approaches by offering centralized, energy-efficient solutions for computationally intensive tasks, creating a hybrid model for future infrastructure needs.

Challenges and Opportunities

Challenges

• Collision Avoidance and Space Traffic Management: As space becomes more congested, ensuring the safety of orbital data centers and avoiding collisions with debris or other spacecraft is paramount. This requires sophisticated tracking systems, adherence to space debris mitigation guidelines, and potentially collaboration with international space agencies.
• Feasibility Testing and Data Transmission to Earth: While Lumen Orbit presents a strong case for the feasibility of their concept, thorough testing and demonstration are needed to validate the technology and its long-term reliability in the harsh environment of space. Efficient and high-bandwidth data transmission to and from Earth is crucial for the viability of space-based data centers.
• Maintaining a Stable Workforce: Given the specialized nature of the technology, attracting and retaining skilled engineers and technicians to design, build, and operate these orbital data centers could be a challenge.
• Cybersecurity: Protecting sensitive data stored and processed in space from cyberattacks requires robust security measures and protocols.
• Regulatory and Legal Framework: The evolving regulatory landscape for space-based activities and data security presents challenges in navigating legal complexities and ensuring compliance.

Opportunities

• First-Mover Advantage: Lumen Orbit is positioned to be a first mover in the space-based data center market, potentially capturing a significant share of the growing demand for AI-driven, energy-intensive computing.
• Untapped Space: Space offers a virtually unlimited environment for expansion, providing ample room for scaling data center operations to meet future demands.
• Lower Energy Costs and Sustainability: Access to continuous solar power and efficient cooling technologies in space translates to lower energy costs and a reduced environmental footprint compared to terrestrial data centers.
• Enhanced Scalability: The modular design and vastness of space enable Lumen Orbit to scale its data centers to unprecedented levels, accommodating the growing needs of AI and other high-performance computing applications.
• Potential for New Markets: Space-based data centers could open up new markets and applications, such as supporting space exploration, research, and commercial activities in orbit and beyond.
• Geopolitical Advantages: Establishing a presence in space with critical infrastructure like data centers could provide strategic advantages and contribute to national space programs and technological leadership.

Lumen Orbit’s success will depend on its ability to navigate the challenges and capitalize on the opportunities presented by this emerging market. By effectively addressing concerns, demonstrating the feasibility and reliability of its technology, and continuously innovating, Lumen Orbit can position itself as a leader in the future of data center solutions.

Conclusion

I love companies that are doing something nobody else has even thought of. Combine that with a massive, massive market, and I’m excited about Lumen. Considering the number of investors clamoring to participate, I’m not alone.

What do you think?

Thanks for reading.

Updates to this article will be made with every subsequent funding round. Stay tuned.

I put out a deep-dive like this one every week. If you enjoyed, follow me here.

If you’re interested in contributing to Garden Research as a writer, email me at pjlabarbera11@gmail.com.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

The Final Frontier… of Data Centers was originally published in Garden Research on Medium, where people are continuing the conversation by highlighting and responding to this story.

Investing in OpenAI at a $150B valuation is ludicrous.

To defend my stance, let’s examine why this valuation isn’t just aggressive — it’s fundamentally misaligned with market realities and internal signals.

Market Context and Valuation

I’m not saying OpenAI is a bad company. I’m not saying their technology isn’t awesome. It is.

They are, undeniably, one of the most transformative companies in recent history.

As first movers in GenAI, their products have revolutionized how consumers and enterprises leverage artificial intelligence. They’ve reached a reported $3.4B in ARR, up substantially from $2B at the end of 2023, and even more impressive considering their ARR mid-way through 2023 was $1B.

However, technology prowess and growth aside, every investment must be evaluated based on potential returns. OpenAI is looking to raise $6–7B at a $150B valuation. Any way you look at it, this represents a massive departure from industry standards:

• Mid-stage AI companies typically see revenue multiples of 10–30x on the high end, 5–15x on the conservative end
• Round-size-to-valuation multiples typically range from 2–5x
• OpenAI’s proposed metrics: 44x revenue and 22x round size

At $150B, OpenAI would become the third most valuable private company in history, behind only SpaceX ($180B) and ByteDance ($220B). This premium is largely driven by their mission to develop artificial general intelligence (AGI). While achieving AGI could fundamentally reshape the global economy, betting on such an outcome to achieve desired returns is speculation at best.

The Return Challenge

VC investors typically target 3–5x returns on late-stage investments. For OpenAI to deliver for this new batch of investors, they’d need to IPO at a minimum of $450B — and that’s assuming optimal conditions, without accounting for dilution, pre-IPO rounds, lock-up periods, stock price fluctuations, or fees.

For context, the highest-valued IPOs in history are Saudi Aramco ($1.88T), Alibaba Group ($231B), and Meta ($104B). While a $450B exit is theoretically possible, it relies on several ambitious assumptions:

• Continued significant revenue growth
• Formulation of a reasonable path to profitability
• Continued technological advancement
• Sufficient capital to achieve the above

Meta provides an interesting comparison. Their market value rose approximately $500B after announcing their AI strategy and Llama investment. Realistically, the portion truly attributable to their AI capabilities is likely $200–300B. Using this comparison, OpenAI’s technology, while more capable than Llama but not by a significant margin, might be valued similarly by public markets.

Moreover, public markets have historically punished excessive spending without clear paths to profitability. Meta’s stock dropped 14% overnight after revealing their AI spending plans. Meanwhile, Sam Altman recently stated he didn’t care if OpenAI lost $50B yearly while pursuing AGI. With projected 2024 losses of $5B despite $3.4B in ARR, this stance could significantly impact public market receptivity.

Technical Parity and Competition

The $100B+ valuation gap between OpenAI and Anthropic becomes harder to justify when examining capabilities. Anthropic’s Claude 3.5 Sonnet performs nearly identically to GPT-4o across most tasks, with some compelling advantages:

• 200,000 token context window vs 128k
• Lower costs ($0.003 vs $0.01 per 1k input tokens)
• Achieved with fewer resources in ⅓ of the time

While GPT-4o1 currently leads in capability, history suggests this gap will be temporary. The enterprise story is similar — Anthropic offers a 500,000 token context window compared to OpenAI’s 128k, at lower operating costs. Should OpenAI be valued at $100B more than a competitor that could potentially overtake them?

Internal Stability Concerns

Perhaps most concerning are the internal signals. Consider this pattern of departures:

Founding Team:

• Elon Musk (Co-Founder) — Left in 2018 over company direction
• Dario Amodei (VP Research) — Left in 2020 to start Anthropic
• John Schulman (Co-Founder) — Left for Anthropic in 2023
• Greg Brockman (Co-Founder) — Currently on extended leave
• Mira Murati (CTO) — Departed last week

Safety Team:

• Ilya Sutskever and Jan Leike left post-2023 leadership crisis
• Supalignment team disbanded
• Daniel Kokotajlo departed citing safety concerns

If team members with significant equity weren’t confident enough to bet on the company’s future, why should outside investors?

The Path Forward

The case against investing in OpenAI at $150B isn’t about questioning their innovation — it’s about fundamental business metrics and organizational stability. With unprecedented valuation multiples, mounting competition from more efficient competitors, and concerning patterns of key departures, the investment thesis requires too many optimistic assumptions to justify the risk.

For venture-scale returns from a $150B entry point, you’re not just betting on OpenAI maintaining market leadership — you’re betting on them achieving technological breakthroughs while simultaneously solving significant organizational and operational challenges. That’s a speculative bet, not an investment thesis.

I could be wrong. The folks deploying money into this round have much better track records than I. With that said, I can’t seem to make the justifications needed to rationalize this with the publicly available information. Maybe the folks at SoftBank know something we don’t…

This past March, Taalas emerged from stealth to develop the next evolution in AI hardware architecture. With backing from Quiet Capital and semiconductor legend turned investor Pierre Lamond, Taalas hopes to create the first-ever “hardcoded” AI chip.

Key Gap Served

Silicon Computing

Let’s start at the beginning. All modern computing is built on the intentional manipulation of binary data. Every program, application, and piece of software you use — every line of code and even the words you’re reading right now — ultimately reduces to binary: 1s and 0s.

At the heart of this process are transistors, tiny electrical switches that control the flow of electricity in a circuit. When electricity flows, it represents a “1”; when it doesn’t, it represents a “0.” Billions of these transistors are densely packed onto the silicon chips inside computers, forming the foundation of all modern computation.

More complex calculations are performed using logic gates: arrangements of transistors that execute basic Boolean operations like AND, OR, NOT, NAND, and XOR. These gates process binary inputs to produce binary outputs based on predefined logical rules. Processors, such as CPUs and GPUs, integrate billions of transistors and millions of logic gates to perform intricate operations, enabling the execution of a vast range of tasks.

To perform these operations, transistors use a “fetch-decode-execute” (FDE) cycle.

• Fetch: The processor retrieves an instruction from memory (stored as binary code).
• Decode: The control unit translates the instruction into signals that the processor’s components understand.
• Execute: The instruction is carried out by components like the Arithmetic Logic Unit (ALU), a specialized part of the processor designed to perform basic arithmetic (e.g., addition, subtraction) and logical operations (e.g., comparisons).

This process is governed by the Instruction Set Architecture (ISA) of processors. Acting as the interface between software programs and the hardware that executes them, the ISA translates software into binary instructions that the processor can interpret. Each processor’s ISA determines how it fetches, decodes, and executes instructions, making it a foundational component of modern computing.

To execute these instructions, processors rely on memory systems that serve as a bridge between processing units and stored data. Instructions dictated by a software program are stored in primary memory systems (RAM), allowing the processor to access them during the FDE cycle. While effective, this setup introduces inefficiencies: the constant movement of data between memory and the processor adds latency and energy overhead, posing significant challenges for high-performance computing tasks.

CPUs and GPUs

Central processing units (CPUs) were introduced, and remain the most common processor for all general-purpose computing. The “brain” of computers, CPUs are built from a small number of powerful central processing units, or cores, optimized for sequential execution, or the processing of instructions one at a time. This sequential processing makes CPUs highly proficient at performing a variety of tasks, as many computing workflows involve sequential dependencies, where each step relies on the completion of the previous one. For example, CPUs are the most proficient processor for:

• Running an Operating System: Tasks like managing resources, allocating memory, and scheduling processes must follow a strict order. CPUs ensure these operations execute reliably and efficiently.
• Application Logic: Software programs often include conditional logic (e.g., “if-then-else” statements), loops, and procedural code that must be executed sequentially to produce correct results.
• Data Processing Pipelines: When steps in a computation depend on the results of previous steps (e.g., parsing a file, performing calculations, and storing results), CPUs handle the dependencies seamlessly.

What CPUs are not well-suited for are tasks requiring parallelization, or the processing of multiple steps simultaneously. To fill this gap, Graphics Processing Units (GPUs) were introduced. Originally designed for graphics rendering — where massive amounts of visual data must be processed simultaneously to render pixels, shade, and apply textures — GPUs are built with thousands of smaller cores designed to execute instructions independently and in parallel. GPUs leverage a Single Instruction, Multiple Threads (SIMT) execution model, where the same operation is applied across multiple data points simultaneously, making them highly efficient for parallel workloads.

GPUs and AI

GPU’s proficiency in parallel processing has made them indispensable for training and running artificial intelligence models. Modern deep learning architectures, such as neural networks, rely heavily on complex mathematical operations to process and learn from data. Matrix multiplications, convolutions, and activation functions form the foundation of these computations, enabling networks to transform input data, extract patterns, identify relationships, and ultimately optimize the weights that dictate the model’s output.

Modern large models rely on massive clusters of GPUs — individual GPUs interconnected to handle large AI workloads — held in data centers. These GPU clusters serve two distinct purposes: training models (which requires enormous computational resources) and running trained models for inference. While these arrays of GPUs have become the foundation that most emerging AI models are built on, as evidenced by NVIDIA’s unbelievable bull run, their current architecture results in constraints that pose challenges to the scalability and efficiency of AI systems, particularly for model deployment and inference.

• Dependence on External Memory Systems: GPUs rely on high-bandwidth memory to store the models’ inner workings — weights, parameters, activations, and intermediate results. The transfer of data between the GPU cores and external memory systems creates bottlenecks, increasing both latency and energy consumption, which only grow worse as model size and complexity increase.
• Simulation of Intelligence: GPUs are software-driven. All models are executed as software layers on top of GPU hardware, which requires the interpretation and translation of high-level abstractions into binary instructions. In other words, every operation carried out by an AI model must first be translated into machine executable commands. This process introduces limitations relating to computational efficiency and increases processing overhead.
• Energy Consumption: The power demands of GPU clusters and the data centers housing them are enormous. High-performance GPUs, such as NVIDIA’s A100, consume up to 400 watts each under maximum load. Training advanced AI models can require thousands of GPUs, resulting in total energy consumption measured in megawatts. For instance, training a single large model like GPT-3 reportedly consumed 1,287 megawatt-hours (MWh) of electricity — the equivalent of powering over 120 U.S. homes for a year. Cooling systems required to manage the heat generated further compound these costs, adding 30–50% more energy consumption to already staggering numbers. These inefficiencies represent not only a financial burden but also a significant environmental challenge.
• Scalability: As AI models continue to grow in size and complexity, demand for GPU clusters grows in tandem. Building and maintaining these clusters necessitates significant capital expenditures, not only for the GPUs themselves but also for the networking infrastructure needed to connect them to each other. Communication delays and bandwidth limitations between GPUs further exacerbate these scalability challenges, particularly when models require frequent synchronization during training.
• General Purpose Design of GPUs: While versatile and powerful, GPUs were adapted to handle AI workloads — they weren’t purpose-built for them. While they are currently the best available option, they are far from perfect. GPUs are not tailored to the unique demands of specific neural network architectures, such as transformers or convolutional networks. This general-purpose design results in wasted energy and computational resources, as GPUs execute a range of operations they are not specifically optimized for.

Founding Story

Recognizing that the exponential growth of AI models may soon outpace the capabilities of general-purpose hardware like GPUs, Taalas was founded to reimagine how models are deployed at scale.

The founding team accumulates decades of combined experience across semiconductor design and AI hardware. CEO Ljubisa Bajic, the former founder of Tenstorrent, is joined by his former employees and early engineers Drago Ignjatovic and Lejla Bajic. Together, the trio was involved in designing and implementing history-defining advancements in CPU, GPU, and AI processor design and architecture.

With their deep understanding of hardware-software optimization and silicon engineering, they are pioneering the development of chips that bypass the limitations of general-purpose processors, delivering a future where AI computation is faster, more efficient, and more sustainable.

As stated by Pierre Lamond:

“Making chips is a difficult and risky endeavor in which success requires experience, ingenuity, and persistence. Taalas’ founders have produced numerous cutting-edge chips and systems. Their track record in the industry is second to none.”

Product and Technology

Taalas is creating an automated flow for embedding AI models directly into silicon chips — creating “hardcore models” that don’t rely on external memory or software. They claim these chips result in a 1000x in efficiency as compared to current methods.

Their platform, known as the “Taalas Foundry”, facilitates the transformation of trained AI native software models into dedicated hardware, embedding the model’s pre-established architecture and weights directly into the silicon. This approach specifically targets the efficiency of model deployment and inference, rather than the initial training process, which would still require traditional GPU infrastructure.

To understand the significance of Taalas’s hardcoding approach, consider a traditional AI model deployment: every time the model runs, its architecture and weights must be loaded from memory into the processor, translated from software instructions into hardware operations, and executed through multiple layers of abstraction. Taalas eliminates these steps by physically embedding the model’s architecture and pre-trained weights into the chip’s circuitry. This is analogous to the difference between running a program from a disk versus having its functionality directly built into a calculator — the dedicated hardware approach is inherently more efficient.

While specific technical details remain sparse, the general methodology behind Taalas’s approach likely involves several key steps:

1. Model Development: AI models are developed using PyTorch, as suggested by references on their public-facing website. PyTorch provides a flexible framework for creating and training neural networks.
2. Hardware Translation: The trained AI model is translated into a hardware description language (HDL), such as VHDL or Verilog. This step maps the model’s operations — like matrix multiplications, convolutions, and activations — onto hardware constructs such as logic gates, arithmetic units, and memory blocks.
3. Chip Design Synthesization: The HDL representation is synthesized into a gate-level design, defining the specific arrangement of transistors and interconnections. This design is then used to create the physical layout of the chip.
4. Silicon Fabrication: The finalized chip layout is sent to a semiconductor fabrication facility (foundry) to manufacture the silicon-based AI chip.

Early reports suggest that Taalas’s processors are tailored to specific use-case requirements, with pre-defined architectures optimized for particular AI tasks. For instance, one chip might be purpose-built for machine vision, leveraging a Convolutional Neural Network (CNN), while another could be designed for natural language processing, embedding a custom Large Language Model (LLM) based on transformer architecture. This level of specialization positions Taalas to deliver highly efficient hardware solutions for a wide range of AI applications.

One significant challenge with this approach is flexibility. Hardwiring a model into a chip raises questions about how its weights can be adjusted after deployment — a critical capability for adapting to new data, optimizing performance, or correcting biases. While details remain limited, Taalas mentions that their chips support fine-tuning, which suggests including programmable elements or reconfigurable logic within the chip to enable such adaptability.

Another comes down to cost. Developing a custom processor can take years and hundreds of millions of dollars in development costs. Taalas plans to address this by creating a proprietary automated engineering workflow that they claim accelerates and reduces the cost of this process.

In addition, Taalas’s site claims that using their technology, apps deployed on their chips may be “written in human language”, suggesting they’re building a programming interface that would allow human written instructions to be abstracted into hardware design without any prior expertise.

This all-encompassing approach fundamentally rethinks how AI computation is performed, and addresses many of the inefficiencies previously outlined with the GPU model. By embedding models directly into the hardware, “hardcore models” eliminate reliance on external memory systems, reducing latency and reliance on infrastructure carrying high energy consumption capital expenditure. In addition, they sidestep the inefficiencies of software-driven execution, avoiding the processing overhead that comes with simulating intelligence on general-purpose hardware like GPUs. Their chips are purpose-built for specific AI workloads, maximizing performance and reducing waste in ways that general-purpose designs can’t compete with.

Matt Humphrey, a partner at Quiet Capital, sees the benefits as threefold:

“…Dramatically resetting the cost structure of AI today, viably enabling the next 10–100x growth in model size, and efficiently running powerful models locally on any consumer device. This is perhaps the most important mission in computing today for the future scalability of AI.”

The Market

The global Artificial Intelligence (AI) market, valued at approximately $515 billion in 2023, is projected to grow to over $2.7 trillion by 2032, with a CAGR of 20.4%. This growth is driven by increasing adoption across industries, advancements in AI technologies, and the proliferation of data that enhances AI capabilities. Additionally, the rise of cloud-based AI services and significant investments in AI research are democratizing access and accelerating its integration into business processes, solidifying AI’s role as a cornerstone of future innovation. By addressing a dire need relevant to the scalability, efficiency, and cost-effectiveness of AI, Taalas looks to take a piece of this expanding pie.

Competitive Landscape

Taalas operates in a competitive hardware landscape, with key rivals spanning several categories.

• GPU manufacturers like NVIDIA dominate the current AI market, providing general-purpose solutions that are powerful and flexible but inherently inefficient for the specific, repetitive workloads AI demands.
• AI-specific chip startups like Etched aim to address this inefficiency by building hardware optimized for AI tasks, though their designs typically don’t achieve the extreme integration Taalas offers by embedding entire models directly into silicon.
• Neuromorphic chips, which attempt to mimic the structure of biological brains to achieve low-power, high-efficiency processing, are on the cutting edge of emerging tech, but they remain in their infancy and are not practical for today’s widely used AI architectures.
• Edge AI chips from players like Hailo and Ambarella focus on delivering efficiency and real-time processing for constrained devices, but their scope is limited to edge applications and lacks the scalability required for large-scale inference and training.

Taalas sets itself apart by targeting the inefficiencies of all these approaches: delivering purpose-built hardware that’s not only tailored to specific AI architectures but also scalable and dramatically more efficient than general-purpose or edge solutions.

The competitive differentiation becomes clearer when examining specific use cases. While NVIDIA’s GPUs excel at training and running multiple different models, they sacrifice efficiency for flexibility. Etched and similar startups improve efficiency but still rely on traditional compute architectures. Taalas’s approach, though less flexible, potentially offers orders of magnitude better performance for specific, well-defined AI tasks. This trade-off — sacrificing flexibility for extreme efficiency — could prove particularly valuable as AI applications become more specialized and performance requirements more demanding.

Challenges and Opportunities

Taalas is tackling some of the most pressing inefficiencies in AI hardware, and the opportunities they’re chasing are huge — but they’re not without real challenges. By embedding models directly into silicon, they’re pushing the boundaries of what AI infrastructure can be, rethinking the foundation of how models run and scale.

Opportunities

• Lowering the Cost of AI Infrastructure: Efficiency gains from Taalas’s hardcore models eliminate the need for external memory and software, which will significantly reduce the cost of running AI models by subverting the reliance on large GPU clusters.
• Higher Efficiency: Custom chips designed for specific AI tasks will result in greater efficiency by bypassing limitations common with general-purpose hardware.
• Consumer Devices: Taalas’s technology could bring powerful AI models to consumer devices by embedding models directly into lightweight, energy-efficient chips. This opens up opportunities for AI applications in smartphones, IoT devices, and other edge environments where traditional hardware cannot operate effectively.
• Supporting Custom-Built AI Models: By enabling AI models to be hardwired into silicon, Taalas allows for the creation of highly specialized and optimized chips for unique applications. This could empower companies to deploy AI solutions tailored specifically to their needs, breaking free from the one-size-fits-all constraints of general-purpose GPUs.
• Facilitating Scalability: The compact, efficient nature of Taalas’s chips makes scaling AI models more practical and cost-effective. Instead of requiring massive data center infrastructure, AI models could be scaled through distributed, energy-efficient silicon implementations.

Challenges

• Model Flexibility: A significant trade-off of hardwired AI models is their lack of flexibility. Once a model is embedded into silicon, updating its architecture or weights becomes challenging. This rigidity could limit adaptability, particularly for tasks requiring frequent retraining or architectural evolution. While Taalas claims their chips support fine-tuning, the extent of this capability remains unclear. Additionally, the initial training of models would still require traditional GPU infrastructure, meaning Taalas’s efficiency gains apply specifically to model deployment and inference rather than the training process.
• Complexity and Cost of Custom Chip Development: Designing and fabricating custom chips is an intricate and capital-intensive process. Developing application-specific integrated circuits (ASICs) involves high upfront costs, long design cycles, and significant risks if market conditions or customer needs change. Achieving profitability at scale will require balancing these risks with the long-term benefits of efficiency and specialization.
• Rapid Evolution of AI Models: AI architectures are evolving at an unprecedented pace, with new paradigms frequently emerging (e.g., transformers, diffusion models, or state-space models). Taalas may struggle to keep up with these shifts, as each new architecture could require significant design adjustments or even entirely new chips, creating a lag between technological innovation and hardware readiness.

Summary

Taalas represents a fundamental reimagining of AI hardware architecture. While significant questions remain about the viability and scalability of their approach, their vision of embedding AI models directly into silicon could transform how we think about AI deployment and scaling. The team’s deep expertise in semiconductor design and AI hardware, combined with backing from industry legends like Pierre Lamond, suggests they have the technical and strategic capability to execute.

I’m bullish. What do you think?

Thanks for reading.

Updates to this article will be made with every subsequent funding round. Stay tuned.

I put out a deep-dive like this one every week. If you enjoyed, follow me here.

If you’re interested in contributing to Garden Research as a writer, email me at pjlabarbera11@gmail.com.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Embedding Intelligence into Silicon was originally published in Garden Research on Medium, where people are continuing the conversation by highlighting and responding to this story.

Private venture investment has historically been the domain of an elite few. Accredited investor requirements, high minimum investments, and regulatory restrictions have locked out over 80% of Americans from participating in a market that has consistently outperformed traditional investment vehicles.

In October, two innovative companies raised seed funding to reshape this landscape, leveraging blockchain technology to democratize opportunities traditionally reserved for institutional investors and high-net-worth individuals.

BIO, backed by Binance Labs, is creating a decentralized ecosystem to drive scientific breakthroughs. By utilizing DAOs, BIO aims to establish an open system accelerating funding, governance, and innovation for research projects, biotech companies, and cutting-edge initiatives.

Fission, which secured $1.6 million in seed funding led by Kraynos Capital Management and SALT Fund, is tokenizing startup equity into digital assets. These tokens enable investors worldwide to access venture capital opportunities, expanding private market participation while bridging traditional finance with decentralization.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Market and Opportunity

In the United States, the SEC is tasked with protecting investors, enforcing regulations requiring securities offerings to be registered prior to sale to the public. This process involves and mandates strict disclosures, including financial statements, governance policies, operations, and strategic plans for the future, providing potential investors with the necessary information to conduct proper due diligence and risk assessment activities and helping to ensure public offerings are legitimized, well-vetted, and de-risked as much as possible.

This process, while well-suited for large companies looking to raise significant amounts of public capital, is no fit for younger startups and growth-stage companies. High costs, complex requirements, and constant reporting requirements would put undue strain on limited resources. Not to mention, exposure to increased scrutiny from a broad base of investors while companies are still “figuring it out” would drastically jeopardize strategic flexibility and competitive advantage creation.

Rather, startups opt to raise money under Regulation D, a provision allowing private companies to raise an unlimited amount of capital from a smaller base of accredited investors — individuals or entities that either:

1. Earn an annual income of $200,000 (or $300,000 combined with a spouse or partner) in each of the last two years
2. Have a net worth over $1M, either individually or with a spouse
3. Are a business entity with total assets exceeding $5M, or are owned by entirely accredited investors
4. Hold a Series 7, Series 65, or Series 82 License

Accreditation requirement to invest in private ventures is, like the rest of the SEC’s actions, based on protection.

Private companies are generally riskier investments than publicly traded companies on a one-to-one comparison model, and accredited investors are presumed to either have the financial sophistication necessary to make informed investment decisions or have enough capital to withstand potential losses. By limiting investment opportunities in “high-risk” startups to these investors, the SEC aims to shield less experienced individuals or those without the funds to lose from undue risk.

However, this restriction excludes over 80% of Americans from participating in a market that has historically outperformed the S&P 500. According to data from Cambridge Associates, the US Venture Capital Index showed a 10-year annualized return of 15.2% from 2013 to 2023, and top-quartile funds could be expected to achieve annual returns ranging from 15% to 27%.

It should be noted that approximately 80% of all venture-backed startups fail to return capital to investors. Most of the annualized return is attributed to a power law distribution, where a small number of extraordinary successes — such as Uber, Airbnb, and similar unicorns — generate the majority of returns.

That said, this concentration of returns highlights a key paradox in venture investing: while individual investments carry high risk, a diversified portfolio of venture investments has historically provided superior returns compared to public markets. Yet current regulations prevent most investors from building such portfolios, effectively reserving these opportunities for those who already have significant wealth.

Many, including myself, believe that individuals should have the freedom to make their own investment decisions and assess risks they are comfortable with. Broadening public participation in venture investing could transform the landscape — empowering individuals to access high-growth opportunities, driving funding to a wider array of startups, and fostering unprecedented levels of innovation and economic growth.

The market for private funding isn’t slowing down. Technological acceleration across AI, BioTech, FinTech, Consumer Health, CleanTech, and other areas is driving a landscape ripe with high-growth opportunities and in need of scaling capital. One estimate projects the market size for venture investment will be $553.5B by 2028. Hopefully, a higher percentage of that will come from a broader investor base.

Two key technological innovations are enabling this potential democratization of private markets: asset tokenization and decentralized autonomous organizations (DAOs). Understanding these mechanisms is crucial to grasping how BIO and Fission are approaching market transformation.

Emerging Technologies Shaping the Future

Tokenization of Private Assets

Tokenization involves converting ownership of physical assets, financial holdings, or intangible property into digital tokens that exist on a blockchain. These tokens represent a stake in an underlying asset.

How Tokenization Works

• Creation of Tokens: Ownership of an asset is divided into digital tokens, each representing a small share of the asset.
• Blockchain Integration: These tokens are recorded on a blockchain, ensuring immutable records of ownership and transaction history.
• Smart Contracts: Automated contracts govern the transfer, rights, and conditions tied to the token, eliminating the need for intermediaries.

Benefits of Tokenization

• Fractional Ownership: By dividing assets into smaller units, tokenization enables investors to own a portion of high-value assets that may have been previously inaccessible to smaller investors. Instead of requiring $1 million to invest in a private company, an individual could own a fractional token for a much smaller amount.
• Enhanced Liquidity: Tokenized assets can be traded on blockchain-based secondary markets, addressing the illiquidity traditionally associated with private investments. Investors no longer need to wait for events like acquisitions or IPOs to realize returns, as tokens can be bought and sold with ease.
• Transparency and Security: Blockchain’s inherent transparency ensures clear records of ownership and transaction history, reducing fraud and disputes.

For private markets specifically, tokenization addresses three fundamental challenges: minimum investment requirements that exclude smaller investors, illiquidity that locks up capital for years, and limited access to top-tier opportunities. By fragmenting ownership and enabling secondary trading, tokenization could fundamentally alter how private market investments function.

DAOs

Decentralized Autonomous Organizations (DAOs) are blockchain-based entities that enable decentralized governance and collective decision-making. By removing traditional hierarchies and intermediaries, DAOs create a flexible, inclusive, and community-driven framework for organizing and funding initiatives.

How DAOs Work

• Definition and Governance: A Decentralized Autonomous Organization (DAO) is a blockchain-based entity governed by smart contracts, which automate rules and decision-making processes. Members participate by holding governance tokens, granting them voting rights to make decisions collectively without centralized leadership.
• Smart Contracts: These self-executing agreements define the DAO’s operations, such as fund allocation, voting mechanics, and proposal approvals, ensuring that rules are transparent and tamper-proof.
• Decentralization: DAOs eliminate the need for intermediaries, enabling a distributed group of participants to operate the organization collectively through blockchain technology.
• Global Accessibility: Anyone with an internet connection and the required governance tokens can join and participate, fostering an inclusive, borderless structure.

Benefits of DAOs

• Decentralized Decision-Making: DAOs democratize control by giving voting rights to token-holders, ensuring that no single party dominates decisions. Collective governance reduces the risks of centralized mismanagement and aligns decisions with the community’s interests.
• Transparent Operations: All transactions and governance actions are recorded on the blockchain, offering a fully auditable and transparent history. This transparency builds trust among participants and external stakeholders.
• Global Resource Pooling: DAOs enable individuals worldwide to pool resources into a collective fund, providing significant capital to support initiatives like startup investments, research, or public goods.
• Inclusivity and Accessibility: Fractional participation allows individuals to contribute small amounts of capital or effort while still having a voice in decision-making. This structure opens investment and project participation to a broader and more diverse audience.
• Alignment of Incentives: Members often benefit directly from the success of the DAO, either through financial returns tied to token value or by advancing shared goals, such as funding impactful research.
• Innovation and Agility: DAOs can quickly adapt governance models or funding priorities based on real-time community input, fostering innovation and responsiveness to new opportunities.

How BIO and DAO Are Revolutionizing Access to Private Markets

BIO is a platform designed to make investing in decentralized projects more accessible by supporting the creation and development of DAOs (Decentralized Autonomous Organizations). Through its incubator program, BIO equips individuals and groups with the tools to launch and manage these DAOs, which can focus on a variety of ventures, including funding private projects. By removing the need for traditional intermediaries, BIO enables broader participation in private markets, allowing anyone with a stake in a DAO to be part of decision-making and potential financial outcomes.

The core of BIO’s products is BioDAOs — decentralized autonomous organizations designed specifically to fund and advance biotech and scientific research. Each DAO focuses on a particular domain or research area, enabling community members, scientists, institutional funders, and outside investors to collaborate and accelerate scientific discovery.

Whether it’s a researcher, a patient with a personal stake in the projects, or an investor with a passion for science, members can join BioDAOs by purchasing or earning governance tokens, which represent their stake in the organization. These tokens empower members to participate in funding decisions, access project updates, and directly influence the direction of research.

Projects within BioDAOs are funded either through token purchases, with accumulated funds allocated to initiatives, or via direct crowdfunding, allowing participants to support specific efforts without purchasing tokens. These projects span a range of activities, from scientific research and clinical studies to early-stage startups developing innovative solutions.

By opening access to a wide array of participants, BioDAOs benefit from a diverse set of resources, including expertise, funding, data, technology, and services that drive the growth and success of each project. Governance is facilitated through token-based voting, where influence correlates to token ownership, aligning participation with each member’s level of investment or involvement.

For scientists, this decentralization of funding addresses a critical challenge in traditional fundraising. Securing grants can take years, is highly competitive, and often excludes high-risk, early-stage projects. BioDAOs provide a faster, more inclusive alternative, empowering researchers to secure resources and pursue groundbreaking work.

Members of a BioDAO, alternatively, are promised portions of all financial gains associated with projects. These rewards often include revenue sharing from successful ventures, equity stakes, or ownership of intellectual property rights. For example, one of BIO’s first DAOs, VitaDAO, funds projects through two primary models: equity ownership and IP-NFTs (intellectual property non-fungible tokens). Equity-funded projects allow DAO participants to own a stake in biotech startups, sharing in the financial success if the company grows or exits. IP-NFT-funded projects, alternatively, allow DAO participants to own the intellectual property rights of projects, earning revenue from licensing deals or royalties generated by the research. These models ensure that contributors not only support innovation but benefit from the breakthroughs they help bring to life.

Beyond the lab, BioDAOs democratize access to scientific investment opportunities, allowing the general public to contribute to and benefit from successful projects. By engaging individuals in the process, these organizations foster a sense of ownership, promote transparency, and create a more inclusive investment landscape — one where the collective potential of a global community drives progress and discovery.

VitaDao, BIO’s first BioDAO, is a prime example of what could be. Since its launch, the 10,000-member DAO has deployed $4.2M in funding across 24 projects, with $6M in liquid funds left to invest. VitaDAO supports groundbreaking research aimed at helping people live longer, healthier lives. Their projects include developing treatments to recycle and rejuvenate aging cells, exploring natural anti-aging compounds inspired by long-lived animals, and using artificial intelligence to discover drugs that could extend human lifespan. Other major efforts focus on eliminating harmful aging cells with advanced immune technology and improving brain health by targeting the root causes of Alzheimer’s disease. These initiatives are driving progress in fighting age-related diseases and enhancing quality of life as we grow older.

While BIO’s model presents compelling opportunities, it also faces unique challenges in the biotech space. Scientific research often requires long development timelines and significant capital, which could test the patience and commitment of DAO members. The complexity of biotech investments creates additional hurdles in decentralized governance — how do you ensure voting members have sufficient expertise to evaluate complex scientific proposals? Moreover, the inherent uncertainty in early-stage biotech research, combined with the need for sustained funding through multiple development phases, could strain the DAO model. BIO will need to carefully balance democratization with effective decision-making to navigate these challenges successfully.

Yet these challenges are precisely what make BIO’s potential impact so significant. By creating a new funding paradigm that combines scientific expertise with broader community involvement, BIO could unlock resources for critical research areas that traditional funding models overlook. The ability to pool global capital and expertise through DAOs might actually be uniquely suited to the long-term, high-impact nature of biotech research. Early evidence from VitaDAO suggests that when given the opportunity, communities will rally behind meaningful scientific endeavors, potentially creating a new model

Team, Traction, and Progress

BIO was founded by Paul Kohlhaas. Previously, he founded Molecule, the platform credited with first introducing intellectual property NFTs to the market.

To date, BIO has facilitated the creation of 10 BioDAOs:

• VitaDao: Early stage longevity research
• AthenaDao: Women’s health research
• PsyDAO: Tokenized psychedelic science
• ValleyDAO: Synthetic biology for climate and cleantech
• HairDAO: Hair loss
• CryoDAO: Cryopreservation
• CerebrumDao: Brain Health and Neurodegeneration
• Curetopia: Rare diseases
• Long Covid Labs: Cure for long Covid
• Quantum Biology DAO: Building a quantum microscope to advance bioresearch

BIO’s role extends beyond just enabling these DAOs — they’re effectively creating a new model for research funding that combines the efficiency of traditional venture capital with the inclusivity of crowdfunding. Each successful BioDAO serves as both proof of concept and network strengthener, attracting more participants and expanding the potential for future DAOs.

Fission is a blockchain-based platform that simplifies and expands public access to private equity investments through tokenization. In addition to tokenization, Fission provides tools for trading, lending, and maintaining liquidity, which helps make investments easier to buy, sell, or use as collateral. By combining these features, Fission democratizes access to private markets, allowing more people to invest in opportunities that were previously out of reach.

Fission’s flagship token, TECH, represents a stake in a curated selection of private companies originally bought, held, and managed by an offshore holding company. The Fission team performs all diligence and underwriting of investments, focusing on high-growth companies backed by major VC firms.

There are two ways to acquire TECH tokens: mint or trade

1. Minting:

• When users mint TECH, they deposit USD into the TECH holding company. These funds are used to purchase private company shares on the secondary market, typically within four weeks.
• To ensure responsible fund deployment, Fission limits minting to 20% of the total value of assets already held by the holding company. For example, if the holding company manages $10M in private shares, users can mint up to $2M in additional tokens.
• While investments are being processed, deposited funds are held in high-yield stablecoins, currently generating a 5% return, providing interim earnings for users.
• Minting prices are based on the Net Asset Value (NAV) of the underlying portfolio, calculated daily by the Fission team using institutional valuation estimates for the private company shares.

2. Trading:

• TECH tokens can also be bought and sold at any time on a decentralized exchange (DEX), ensuring liquidity for holders. This feature allows for broader participation and easier access to the token, even when minting opportunities are temporarily unavailable.
• The DEX facilitates permissionless trading, automated price discovery through smart contracts, and real-time low-fee execution.

To manage risk and provide stability, the Fission investment team employs strict diversification strategies:

• No more than 20% of the total fund is allocated to a single asset.
• No more than 40% of the fund is invested in any specific sub-sector.

These policies reduce concentration risk, offering token holders exposure to a balanced and diversified portfolio.

As compared to traditional venture capital, Fission offers investors several compelling benefits. Unlike conventional private equity investments, which are typically illiquid until an exit event such as an acquisition or IPO, TECH can be traded at any time on a decentralized exchange, providing flexibility for investors to realize returns or reallocate capital without long waiting periods. In addition, TECH tokens may be used as collateral for loans, allowing holders to access liquidity without selling equity-backed assets, enhancing capital efficiency and flexibility.

The Fission team takes a careful, calculated approach to diligence, focusing on high-growth companies backed by leading VC firms to minimize the risks often associated with private equity investments. Currently, the TECH token represents holdings in Stripe, SpaceX, Discord, Circle, Figma, OpenAI, Neuralink, X, Carta, Plaid, and Anduril — all companies projected to be successful in the long term.

Transparency is central to their model, with the Net Asset Value (NAV) of the underlying portfolio updated daily, giving investors a clear, real-time view of their token’s worth. In the meantime, any funds waiting to be deployed are held in high-yield stablecoins, earning a 5% return and providing additional value.

Fission sets itself apart by bridging traditional venture capital with blockchain innovation, offering tokenized private equity that’s accessible, liquid, and transparent. Through daily valuation updates and seamless trading on a decentralized exchange, Fission lowers barriers to entry while providing investors the flexibility to trade or leverage their holdings. This unique model combines the growth potential of venture capital with the liquidity and adaptability of decentralized finance, redefining private equity investing.

Competitive Analysis: How do these solutions compare to current options?

The primary existing avenue for unaccredited investors to participate in private markets is equity crowdfunding. In 2015, the SEC signed Reg CF into effect, a provision allowing small businesses and private entities to raise limited funds ($1M to start, later expanded to $5M in 2021) from the public.

While a step forward, Reg CF still has significant limitations. These include:

• Limitations on capital raised: Companies can raise a maximum of $5 million annually under Reg CF, which may not meet the funding needs of growth-stage startups requiring larger amounts of capital, creating a scenario in which equity crowdfunding may not be attractive to all companies.
• Illiquidity: Shares purchased through equity crowdfunding are typically not tradable on secondary markets, leaving investors unable to exit until a significant company event occurs (e.g., acquisition or IPO).
• Deal Quality: Platforms often feature early-stage companies with limited proven traction, increasing the risk of failure and reducing the likelihood of high returns for investors. Companies that might attract institutional VC funding are less likely to pursue equity crowdfunding due to the comparatively lower amounts raised.
• Regulatory Restrictions for Investors: Non-accredited investors face strict investment caps based on their annual income or net worth, limiting their ability to make meaningful allocations even if they identify promising opportunities.

Several companies operate within this space, all of which I admire. WeFunder, SeedInvest, and Republic have all helped to expand access to Venture Capital for everyday investors, but each suffers from the above challenges. I don’t believe this is a zero-sum game. For true democratization, we need equity crowdfunding platforms, but these can’t and should not be the only options.

Future Outlooks: Opportunities, Challenges, and Final Thoughts

BIO and Fission exemplify how blockchain technology is reshaping private market investments, offering solutions that enhance liquidity, broaden participation, and accelerate innovation.

However, the promise of these platforms comes with significant challenges. Regulatory frameworks remain a critical hurdle, especially for tokenized and decentralized finance (DeFi) solutions. While evolving policies could expand eligibility, platforms like Fission cannot currently be used by U.S. residents, creating a significant barrier to market penetration. Furthermore, as tokenized assets transcend borders, compliance with multiple jurisdictions adds complexity, requiring robust systems to maintain trust and scalability.

Despite these challenges, the potential impact of BIO and Fission is immense. Their models represent a significant shift toward democratizing private markets, enabling individuals to bypass traditional barriers and access high-growth opportunities. By reducing funding bottlenecks and introducing liquidity into previously static markets, BIO could drive biotech research forward, while Fission broadens access to private equity.

The long-term vision for platforms like BIO and Fission is to establish global ecosystems where investors, creators, and innovators interact seamlessly. By adopting DAOs and DeFi models, these platforms could redefine investment landscapes, making wealth-building opportunities more equitable and accessible.

While regulatory and operational hurdles remain significant, BIO and Fission represent two distinct but complementary visions for democratizing private markets. BIO’s focus on research funding through DAOs could create entirely new funding mechanisms for scientific developments, while Fission’s approach to tokenizing established private companies could bridge the gap between traditional and decentralized finance. Success for either company would validate blockchain’s potential to reshape private market access.

The implications extend beyond just investment returns. Democratizing private market access could fundamentally alter how innovation is funded, who benefits from technological breakthroughs, and how wealth is created and distributed. While challenges around regulation and implementation remain, the potential for transformation is immense.

We at Garden Research are particularly excited about these developments, as they represent a crucial step toward a more inclusive and efficient private investment ecosystem. The success of these platforms could catalyze a broader shift in how we think about private market access and who gets to participate in the next generation of breakthrough companies.

I put out a deep-dive like this one every week. If you enjoyed, follow me here.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Private Markets Are Being Decentralized. Here’s What That Means. was originally published in Garden Research on Medium, where people are continuing the conversation by highlighting and responding to this story.

In October 2024, Paterna BioSciences announced the successful closure of a $6M oversubscribed seed funding round led by SpringTide Ventures, with participation from the Utah Innovation Fund and e184. The funds will be used to advance Paterna’s mission to address male infertility, ultimately increasing the success rates of IVF procedures.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Key Gap Served

For the nearly 15% of infertile couples in America, a common, often last resort effort to achieve pregnancy is in vitro fertilization (IVF). Approximately 3% of all children born in the United States are conceived using this process, which involves the fertilization of a mature egg with a sperm cell outside of the body.

With infertility rates on the rise, some projections estimate that IVF usage, the most common method of conception for couples struggling to conceive, may increase 5–10% every year.

Albeit being the most commonly used and most successful method of conception outside of natural intercourse, failure rates are unnervingly high. For women under 35, success rates are a meager 50%, and couples must undergo an average of two to three cycles prior to achieving a successful pregnancy. This repeated effort can result in a potential spend of over $50,000 for hopeful parents.

Behind these statistics lie countless personal stories of emotional and financial strain. For couples facing male infertility, the journey often involves multiple failed treatment attempts, each carrying not just a financial burden but also a significant emotional and psychological impact. The lack of effective treatment options can strain relationships, affect mental health, and force difficult decisions about alternative paths to parenthood. This human cost, combined with the high financial investment required for multiple IVF cycles, underscores the critical need for more effective solutions in male infertility treatment.

Though it was originally designed to specifically address female infertility issues, of all unsuccessful IVF cycles, around half of the failures can attributed to the usage of invalid sperm. Male infertility contributes to this significantly, with rates hovering around 8–10%, but failure is only exacerbated by inefficiencies within current IVF practices.

Even the most infertile of men do have viable sperm. Unfortunately, common IVF practices do not properly optimize the rate at which viable gametes are used, relying on the subjective and often inaccurate manual selection of sperm under a microscope by an embryologist. This human owned visual assessment is inconsistent, prone to error, and assumedly significantly contributes to the total failure rate, though pegging a concrete percentage of blame is difficult.

While male infertility treatment could theoretically be implemented to increase the odds of viable sperm selection, there are currently no FDA-approved treatments in the market, and current options are invasive, costly, and largely ineffective.

Founding Story

From the mouth of CEO and Co-Founder, Dr. Alex Patuszak:

“There have been very few and very limited efficacy treatment options for men and their fertility problems. Typically, when I see a man in a clinic with fertility problems, there are only two things I can offer him: medicine, which has literally a flip-of-a-coin chance, fifty-fifty, of helping him solve his problem, or surgery. That is it. In the context of where we are technologically, we should be doing much better. So we asked the question, ‘Can we do better’?”

The company, led by a coalition of PhD founders with prior experience across operations of early-stage companies and supported by an array of researchers, scientists, and advisors across urology, bioinformatics, stem cell research, and startup operations, is on a mission to fill the gaps in male infertility treatments and increase the success rates of IVF treatments.

Product and Technology

Paterna’s approach centers around the development of in vitro spermatogenesis (IVS), an advanced biomedical technique that replicates the natural process of sperm production outside of the human body.

Within the testes, spermatogenesis occurs in the seminiferous tubules and consists of three main stages:

• Proliferation: Spermatogonial stem cells (SSCs), the starting point for sperm development, divide to increase their numbers and create the early-stage cells needed for the next steps.
• Meiosis: These early cells, called spermatocytes, go through two rounds of cell division. This cuts their chromosome count in half, transforming them into immature sperm cells called spermatids.
• Spermiogenesis: The spermatids transform, changing their shape and structure to become fully mature sperm. They develop a tail for swimming and condense their DNA into a compact head, making them ready for fertilization.

IVS attempts to replicate the above stages in a controlled environment. Harvested SSCs from the testes are placed in a controlled environment, where temperature, oxygen levels, and pH are highly regulated to mimic the testicular niche. In this environment, using a combination of signaling molecules, meiosis-inducing factors, and growth factors, the SSCs are manually guided through the above stages to become mature, viable sperm which may be used in IVF treatment.

Much of Paterna’s work to date has been spent laying a foundation. Studying spermatogenesis to understand what it looks like at a molecular level.

This careful dissection of the “programming” that turns an SSC into a viable gamete has given the team a deep and unique understanding of the specific parameters necessary to achieve successful gamete production, and has allowed Paterna to reach the milestone of gamete generation that completed all cellular processes of DNA — intermixing, dividing, and maturing — so that they can be used for fertilization.

They’re currently in talks with the FDA to set up clinical development processes for trials, which will be split across two studies. The first, a phase zero study set to begin in 2025, will involve testing the gametes to determine if they can fertilize human eggs in vitro. If successful, they’ll progress into a true clinical trial in which the embryos are transferred into a live female subject, identical to the IVF process.

The Market

Technically, Paterna occupies both the IVF and male infertility treatment markets. Their technology directly addresses both use cases, as successful in vitro gamete creation can be used outside of IVF for direct male infertility treatment. With that said, IVF will likely be the most common and first end-use, whereas general male infertility will be expanded into later, as discussed in the following passage.

The global in vitro fertilization (IVF) market was valued at approximately USD 25.3 billion in 2023 and is projected to grow at a compound annual growth rate (CAGR) of 5.54% from 2024 to 2030. This growth is driven by factors such as rising infertility rates, advancements in reproductive technologies, and increased societal acceptance of assisted reproductive techniques. Additionally, the trend of delayed parenthood and the availability of fertility treatments have contributed to the market’s expansion.

The male infertility treatment market was estimated at USD 3.85 billion in 2024 and is expected to reach USD 4.58 billion by 2029, growing at a CAGR of 3.54% during the forecast period. Key growth drivers include increasing awareness of male infertility issues, advancements in diagnostic techniques, and the development of effective treatment options. However, societal stigma and limited access to specialized care in certain regions may pose challenges to market growth.

Business Model

Though Paterna Biosciences has not explicitly disclosed its business model, it can be estimated that revenue generation will be driven through multiple channels, including:

• Licensing of IVS Technology: Paterna could license its in vitro spermatogenesis (IVS) technology to IVF clinics and fertility centers, allowing these facilities to integrate the process into existing operations. This would provide recurring revenue streams as clinics adopt the technology to improve success rates and attract more patients.
• Direct Sales of Pre-Qualified Gametes: By offering high-quality, pre-qualified gametes for use in IVF, Paterna could provide clinics with a tangible product that ensures improved fertilization outcomes. This service could address the significant inefficiencies in current sperm selection methods, potentially becoming a gold standard in the industry.
• Collaboration with the FDA for Male Infertility Treatments: Paterna’s long-term goals likely include securing FDA approval for direct treatments addressing male infertility. This could open entirely new therapeutic markets, allowing Paterna to position itself as a pioneer in solving one of the most persistent challenges in reproductive medicine.

Our Thesis

At least to begin, Paterna will likely focus on one of the first two mentioned paths — licensing their technology or becoming a provider of viable gametes. If I were to guess, I’d bet on licensing, which would enable Paterna to scale quickly without the necessity for physical clinical infrastructure of its own. That said, there is a reality in which both eventually exist simultaneously.

As they scale, they’ll likely have the resources necessary to operate a proprietary network of clinics, becoming an independent supplier of fertile gametes and potentially expanding into operating full-service IVF clinics. This is even more likely when considering the ultimate goal is to achieve FDA approval for general male infertility treatment.

Beyond licensing and direct gamete production, Paterna’s revenue potential extends across several key channels. The specialized nature of IVS technology opens opportunities for clinical training and certification programs, creating additional revenue through the education of fertility specialists. Insurance partnerships represent another significant avenue, especially as the technology demonstrates improved success rates over traditional methods. With coverage partnerships in place, Paterna could establish more predictable revenue streams while expanding market access. Their pioneering work in male infertility and spermatogenesis also positions them to monetize their intellectual property across the broader reproductive medicine space, where their research and technology could find valuable applications beyond direct IVF use.

Competitive Landscape

Paterna operates in a specific, underdeveloped, and unique segment of the fertility market that has remained untouched. By directly addressing both the foundational challenge of sperm production in infertile men while creating a solution to the prevalence of human error involved in viable gamete selection by embryologists, Paterna finds themselves in a league of their own: despite operating in a growing, high-importance market, they face limited direct competition. While the broader IVF and male infertility sectors are saturated with ancillary solutions, few companies are tackling the core issue of infertility at a cellular level.

For example, early-stage startups in the male infertility space primarily focus on complimentary services. Companies like Legacy and Dadi provide solutions spanning at-home sperm testing and cryopreservation but fail to address underlying causes of infertility and do nothing to prevent the human error associated with IVF. Other startups focus on supplements or behavioral health tools aimed at improving sperm quality, but these approaches often lack robust clinical validation and offer limited efficacy.

Simultaneously, large pharmaceutical companies are exploring male infertility, but their focus is on drug development and hormonal therapies. Though these solutions are designed to address infertility at its core, stimulating sperm health and production, many are still in the clinical trial phase and face long timelines, high costs, and moderate efficacy. Historically, pharmaceutical advancements in male infertility have been slow, with regulatory and development challenges hindering progress. Paterna, by contrast, circumvents the limitations of pharmaceutical approaches with its disruptive IVS technology, which has a significantly more limited potential for negative side effects on participants as compared to direct drug intervention.

This differentiation comes with good reason. The technological complexity of replicating spermatogenesis in vitro requires advanced expertise and infrastructure, creating a significant hurdle for potential entrants.

Paterna’s strategic lead in clinical trials and ongoing collaboration with the FDA enhances its competitive positioning, as it is poised to establish industry standards. Moreover, partnerships with leading IVF clinics and academic institutions will likely create network effects, further entrenching its leadership position.

Challenges and Opportunities

Opportunities

• Expanding Male Infertility Treatment Options: Paterna fills a critical gap in male infertility treatment and has the potential to be the first FDA-approved procedure to serve the space. By addressing this unmet need, the company could redefine infertility care and create a new standard for IVF.
• Broader Use Cases for IVS: Beyond infertility, successful IVS technology may be applied to fertility preservation. For prepubertal boys undergoing chemotherapy, radiation treatments, or suffering from diseases that leave them unable to produce mature sperm, IVS could be used to help them conceive in their adult years. Additionally, it could be leveraged in research settings to study genetic diseases, test new contraceptive methods, or explore reproductive health innovations, opening up new revenue streams.
• First Mover Advantage: Paterna is the first company to explore commercial applications for IVS treatment. This is a significant moat and will allow them to set industry standards, build partnerships with fertility clinics, and establish themselves as a market leader before anyone else has developed a competitive process.
• Global Market Expansion: With fertility rates declining in many countries, the demand for innovative reproductive technologies is rising. Paterna’s solution has global applicability, offering opportunities to capture significant market share in regions with high infertility rates and growing acceptance of assisted reproductive technologies.

Challenges

• Regulatory Hurdles: Paterna faces a complex regulatory landscape in obtaining FDA approval for its novel in vitro spermatogenesis (IVS) technology. The lack of precedent for lab-grown gametes in fertility treatment adds additional layers of scrutiny, requiring extensive clinical trials and compliance with stringent safety and ethical standards. Delays in this process could slow their go-to-market timeline.
• Technological Scaling: While the lab process for IVS has shown promise, scaling it for widespread clinical use presents significant challenges. Ensuring consistency, efficiency, and cost-effectiveness in producing high-quality, lab-grown sperm at a commercial level remains a critical hurdle that must be addressed to meet market demand.
• Market Adoption Resistance: Convincing fertility clinics and patients to adopt a disruptive technology like IVS may encounter resistance. Traditional IVF processes are deeply entrenched, and educating stakeholders about the benefits of Paterna’s approach will require substantial marketing and data-backed validation of success rates.
• Competition from Established Players: While Paterna has limited direct competition in IVS, established pharmaceutical companies with deep resources could pivot to this space if IVS demonstrates commercial viability. If they do not move quickly to form key partnerships and create a leading presence, there will come a time when larger companies may compete through their established network.

Summary

Paterna’s entry into the fertility market represents a compelling intersection of technological innovation and critical medical need. Their approach to in vitro spermatogenesis not only addresses foundational challenges in male infertility but also targets significant inefficiencies in current IVF practices. While scaling biological processes and navigating regulatory frameworks present clear challenges, Paterna’s strategic position as a first mover in commercial IVS technology, combined with their focus on both immediate IVF applications and broader fertility preservation use cases, creates multiple paths to market success. The company’s potential to fundamentally alter fertility treatment while establishing valuable IP in a rapidly growing market makes them particularly interesting to watch. I’m bullish on Paterna’s ability to transform an industry that has seen limited innovation in male fertility treatment, especially given their clear path to commercialization and multiple revenue opportunities across both clinical and research applications.

Updates to this article will be made with every subsequent funding round. Stay tuned.

I put out a deep-dive like this one every week. If you enjoyed, follow me here.

If you’re interested in contributing to Garden Research as a writer, email me at pjlabarbera11@gmail.com.

For weekly updates on the entirety of the U.S. Seed Ecosystem, including who raised, who’s exciting, and who’s hiring, subscribe to my Substack below

→ Garden Research ←

Paterna Biosciences: Deep Dive was originally published in Garden Research on Medium, where people are continuing the conversation by highlighting and responding to this story.