<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Alex Reinhart on Medium]]></title>
        <description><![CDATA[Stories by Alex Reinhart on Medium]]></description>
        <link>https://medium.com/@alexmreinhart?source=rss-949b7a03ae4f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*pV_Fg1M8HG4dXj71sPk9qQ.png</url>
            <title>Stories by Alex Reinhart on Medium</title>
            <link>https://medium.com/@alexmreinhart?source=rss-949b7a03ae4f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 17:49:07 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@alexmreinhart/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Rapid Fire, learn the latest in AI — International Conference for ML (ICML), favorite papers]]></title>
            <link>https://medium.com/@alexmreinhart/rapid-fire-learn-the-latest-in-ai-international-conference-for-ml-icml-favorite-papers-2b8ca767edc8?source=rss-949b7a03ae4f------2</link>
            <guid isPermaLink="false">https://medium.com/p/2b8ca767edc8</guid>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Alex Reinhart]]></dc:creator>
            <pubDate>Thu, 01 Aug 2024 18:50:04 GMT</pubDate>
            <atom:updated>2024-08-01T18:50:04.627Z</atom:updated>
            <content:encoded><![CDATA[<h3>Rapid Fire, learn the latest in AI — International Conference for ML (ICML), favorite papers</h3><p><em>Some colleagues and I spent the week at ICML in Vienna meeting with researchers at the cutting edge of ML. There were publishers from Google, DeepMind, Brain, Meta, Apple, MIT, Stanford, Harvard, and dozens of other top institutions from around the world.</em></p><p><em>In this essay I will highlight the most interesting posters I saw, and why they matter.</em></p><h3>ML Diagnostics in Healthcare</h3><p><em>I saw surprisingly few of these this year but this is an area I believe has significant opportunity.</em></p><h4><a href="https://openreview.net/forum?id=QXqXGDapkQ">SleepFM: Multimodal monitoring for diagnostic evaluation.</a></h4><p>This work focused on developing a model that could ingest data from <strong>Polysomnography (PSG)</strong>, the current state of the art for studying sleep. PSG measures brain activity, ECG, and respiratory signals, and is used to diagnose a variety of conditions — Sleep apnea, arrhythmia, even early signs of Parkinson’s.</p><p>I’d imagine that there are even more correlations and early diagnosis opportunities that could only be discovered though large data sets (which are currently hard to access) and ML based approaches (that can see patters not visible to the human eye).</p><p><em>Just look at the raw data that you are working with (below), this is exactly the scenario where models can do a much better job than humans.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/625/1*UwVRQRHZCHnlb8Pa9-xzmA.png" /><figcaption>Sample PSG data &amp; SleepFM processing technique</figcaption></figure><p>I spent a while at ICML talking with Magnus, one of the authors of this paper. My main takeaway from the discussion was — <strong>So many things in medicine are subject to physician interpretation.</strong></p><p>The definition for sleep apnea is ever-evolving but the mentality of a physician doesn’t necessarily update every time the FDA releases a new opinion, so you learn that complying with the FDA definition doesn’t necessarily mean matching the doctors analysis of a dataset. The SleepFM work is training their models to match the FDA specifications.</p><p>Outside of FDA specifications there are also just individual biases, risk tolerances, biases to action/patience amongst different physicians that lead to different diagnoses.</p><p>For these reasons it’s particularly hard to get adoption of ML tools. What perfect for one practitioner might seem wrong to another. The FDA is the group you need to convince, but that wont matter if nobody adopts the tool.</p><p>It seems to me that we want our medical diagnoses to follow FDA standards and be as objective as possible. Thus I hope that practitioners continue to understand the value of tools of this sort.</p><p><strong>The startup landscape. </strong>One of our investments at Pillar, <a href="https://www.pathai.com/">PathAI</a> is the perfect example of a ML diagnosis assistant, in pathology, navigating the FDA and eventually getting approval to independently categorize and diagnose tissue samples.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8mZ0iS458-lvl-Lv-naLCg.png" /><figcaption>Path AI</figcaption></figure><p>Companies like PathAI blazed the trail and I expect there will be a proliferation of startups developing ML based diagnostic tools.</p><p>Excited to see how this project evolves.</p><h3>Evaluating Model Fairness</h3><h4><a href="https://arxiv.org/abs/2402.12572">FairProof: Confidential and Certifiable Fairness for NNs</a></h4><p>Usually paper abstracts require some translation to layman’s terms but this one is excellent and I couldn’t have said it better myself-</p><blockquote>‘Machine learning models are increasingly used in societal applications, yet legal and privacy concerns demand that they very often be kept confidential. Consequently, there is a growing distrust about the fairness properties of these models in the minds of consumers, who are often at the receiving end of model predictions. To this end, we propose a system that uses Zero-Knowledge Proofs (a cryptographic primitive) to publicly verify the fairness of a model, while maintaining confidentiality.’</blockquote><p>Say you are a bank using algorithms to evaluate lone applications. How do you prove to applicants that your algorithms are fair without giving them any information on your proprietary model. Particularly if you use opaque algorithms such as ML to made decisions.</p><p>There are two things that matter here. What is fairness, and what is your system for proving fairness without sharing any information. This paper tackles the former. So- <em>you define what’s fair, and we show you how to prove that. </em>For this paper they proposed a definition of fairness to demonstrate the principle.</p><p>Fairproof, used a fairness definition of<strong> <em>fairness for the individual</em></strong><em> </em>(fairness to the lone applicant). The basis for individual fairness is that <strong>similar inputs must have similar output</strong>, aka lone applicants who are the same in every characteristic except 1 (gender for example) should’ve very similar outputs).</p><p>The process is to identify your sensitive variables {x} and quantify how much a change in {x} has on your output.</p><p>The aggregation of this across your most sensitive variables, defines your fairness. This process would also provide transparency to <em>the bank</em> on what they are most sensitive to (intentionally or not), likely income bracket, credit score, for this example.</p><p>There are also many possible definitions of fairness ( buyer, seller, collective).</p><h4><a href="https://arxiv.org/abs/2312.02592"><strong>FRAPPE: A group Fairness Framework for Post-Processing Everything</strong></a></h4><p>There was also a paper out of google tackling fairness from a different angle.</p><p>The question they were trying to answer was — can you have one model that is a combination of two tasks, the first task being any general task, lets use classification for example, and the second task being a fairness normalization of the output (based on whatever definition of fairness you want).</p><p>The value of this would be to build your fairness normalization into your model rather than running post processing to clean things up (slower, rule based).</p><p><strong>An example-</strong> using our prior definition/case study of fairness, lets say you are overly sensitive to a variable that you don’t want to be (like gender), you would like to neutralize this bias in real time, within the model. Or if you cannot modify the model, you can used their technique to remove the bias in post.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/388/1*dWW8rRz4B9pxso1xKiDyow.png" /><figcaption>Fairness mitigation strategy</figcaption></figure><p>Constructing effective ways correct <em>unfairness</em> is the other critical piece to increase adoption. A complete solution requires both.</p><h3>Fairness as a Buisness</h3><p><strong>Why this is important</strong>- A major barrier to ML adoption is trust, and accepted systems for fairness could fix that and increase acceptance.</p><p>I had a fun debate with <a href="https://www.linkedin.com/in/parker-mckee-491118a7/">Parker McKee</a> on how model fairness and interpretability relate to each other. To me it seems like fairness is the narrow subset of interpretability that we may actually be able to do well. And it is the minimum level of interpretability that would allow adoption in new categories. As far as I know, I have not seen someone demonstrate a complete interpretability solution. If you have, let me know.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*mNzpWb7n7Q_BYFxO" /><figcaption>ML Fairness</figcaption></figure><p>The other discussion we were having was if fairness could be a business on its own. My perspective is that if a company could define fairness standards that were accepted by regulatory bodies (FDIC/OCC for loans) and could help companies design models that complied with the regulations. That could be a big standalone business.</p><p>As an example of this, you can look to companies who have defined standards for and dominated the voluntary carbon markets such as <a href="https://verra.org/">Verra</a> or <a href="https://pachama.com/">Pachama</a>.</p><p>Would love to hear others perspectives, please reach out.</p><h3>LLMs Simulating Human Behavior</h3><h4><a href="https://arxiv.org/abs/2402.05863"><strong>How Well Can LLMs Negotiate? NegotiationArena Platform and Analysis</strong></a></h4><p><em>This one was funny, obviously we need to put LLMs into game arenas and see how well they replicate human behavior. It’s awesome.</em></p><p>The key finding here is that language models can exhibit human behavior in really creative ways.</p><p>Rather than taking a game theoretic approach to modeling game play, they constructed an open sandbox for negotiation ( infinite turn, all valid answers in bounds), construing games that didn’t have a right answer (negotiate a price to buy/sell a car), and using a LLMs to simulate human behavior.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/646/1*DHu2rovoq_vPybVBzQFMyQ.png" /><figcaption>Reasoning patters from NegotiationArena</figcaption></figure><p>In the Negotiation Arena the researchers observed behavior such as the model pretending to be elderly, being angry, cunning, with little to no prompting.</p><p>Right now we use game theory for a lot of human modeling, but actually this tends to overestimate the logic/intellect/patience of most participants (no offense, but how often are you thinking about a Nash equilibrium when you make decision, exactly).</p><p>My point is maybe the LLMs that are a conglomeration of general human activity online, might be able to more accurately predict human behavior.</p><p>I would be curious to see these models play more complex based games like poker or Katan where both negotiation/bluffing and rule following are required.</p><h3>ML for Control Systems</h3><h4><a href="https://arc.net/l/quote/jnqdrvsv">Opportunities Exist for Machine Learning in Magnetic Fusion Energy</a></h4><p>For background on fusion check out my article <a href="https://medium.com/@alexmreinhart/the-new-nimby-nuclear-in-my-backyard-please-3b4243f9e436">here</a>.</p><p>I was excited to see ML being applied to controls systems. Controls systems are a Mechanical Engineering concept in which you use force inputs to keep a item in a certain state. Think about balancing, for example a hoverboard uses controls to balance you/move forward, etc.</p><p>Well in magnetic confinement fusion (MCF) you use magnets, to hold a plasma in a certain shape that we deem stable (the sun is a stable plasma, stabilized by its own mass/gravity).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/897/1*EldId5h4OIvHkgExuArCMA.png" /></figure><p>Researchers have successfully demonstrated that ML models can beat state of the art controls algorithms for the controls stabilizing a plasma in a <a href="https://www.energy.gov/science/doe-explainstokamaks">tokamak</a> (shown below).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/708/1*QbK4-k9wNwO6YuCGspVN3Q.png" /><figcaption>From Paper</figcaption></figure><p>Lucas (first author) was sharing that he believes there’s line of sight to Q positive Magnetic confinement fusion within the next two years, from a scientific de-risking perspective, and likely from an engineering de-risk perspective.</p><p>Using models for live control of complex dynamical systems is a relatively new thing and it has now shown success in one of the most complex control systems — confining a plasma (piece of sun). Where to next?</p><p>Seems like it’s time to start making some tangential fusion investments.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2b8ca767edc8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Funding Capital Intensive Businesses as an early stage VC]]></title>
            <link>https://medium.com/@alexmreinhart/funding-capital-intensive-businesses-as-an-early-stage-vc-6e39e9871967?source=rss-949b7a03ae4f------2</link>
            <guid isPermaLink="false">https://medium.com/p/6e39e9871967</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[deeptech]]></category>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[finance]]></category>
            <dc:creator><![CDATA[Alex Reinhart]]></dc:creator>
            <pubDate>Mon, 08 Jul 2024 14:01:29 GMT</pubDate>
            <atom:updated>2024-07-15T17:20:35.943Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How to invest in something that requires BIG money, with little money, and still make a return.</em></p><p>Most cutting edge technology with the potential to — change the world, be big business, make lots of money, or however you define success — comes from science, not software.</p><p>But most investors these days only invest in software, because it’s easier. You don’t have to worry about bankruptcy in nearly the same way, you can launch a feature and see exponential growth in a matter of days.</p><p>That’s just not going to happen in a deep tech or other capital intensive business.</p><p>Deep tech businesses are harder to build, but in that barrier is also an inherent staying power. Non-software businesses have very different capital journeys, but they also have huge potential to drive returns in funds.</p><p>6 of the top 10 companies in the world develop and sell physical assets!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lkwIXTPDqlcGHV7LYmjx_Q.png" /><figcaption><a href="https://www.forbesindia.com/article/explainers/top-10-largest-companies-world-market-cap/86341/1">https://www.forbesindia.com/article/explainers/top-10-largest-companies-world-market-cap/86341/1</a></figcaption></figure><p>I meet so many investors who won’t touch a hardware business. That is because it takes a deeper level of patience, commitment, and diligence to place one of these bets.</p><p>In this article we will discuss how to think about investments with these characteristics.</p><h3><strong>Setting the stage — What matters for a VC?</strong></h3><p>What matters to a VC is that every investment has the potential to return the fund while owning only 5–10% at exit. With a fund size of 200M, and exit ownership of 5%, you need an exit of $4 Billion.</p><p>This means that when a VC is evaluating your deal they asking themselves can this business plausibly be worth 4 billion.</p><p>Below is an example valuation journey of Stripe which is obviously not a hardware business but most companies don’t share their valuations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*As4DbD-7nLFjGEuY.png" /></figure><p><strong>Valuation 101</strong> — As your company becomes less risky and more profitable, its valuation goes up. Valuations are set at each round and negotiated between founders and participating investors. This allows you to raise money selling off smaller portions of your business for more money. Until you IPO and then your valuation is decided upon by the public markets.</p><p><strong>What’s different in a capital intensive business? </strong>You need more money to build your product, and so you raise more pre-revenue rounds that most software businesses. This means to increase your valuation you better be de-risking in some other way.</p><h3><strong>What defines a capital intensive business?</strong></h3><p>There are two variables I evaluate these investments with:</p><ol><li><strong>Capex to first revenue — </strong>How many dollars do you need before you make a dollar (running experiments, building a plant, running a clinical trial, etc)</li><li><strong>Run rate</strong> <strong>margins </strong>— The steady state operations of the business you plan to build.</li></ol><p><strong>Capex to first revenue </strong>is an estimate of how much money must be invested to de risk the product.</p><p>This is where you win the right to exist. These raises come from venture investors who can tolerate the risk of a pre-revenue company. And you need a lot of dollars, which likely means various funding rounds.</p><p>The key thing for these companies to think about is — what internal milestones will increase my valuation before I go out for more funding. Some examples may be high value experiments, functional prototypes, regulatory approvals, etc.</p><p>Revenue generation is a key value inflection point for a company, and unlocks higher valuations and non-dilutive financing. But until you reach that point, you need to understand <strong>what investors will give you credit for</strong> de-risking, and what they wont.</p><p><strong>Run rate margins how many dollars of assets are required to produce a dollar of sales when operating at scale.</strong></p><p>This is where the financial models come into play.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G__JxA7s_vtzOejOD5FACQ.png" /></figure><p>Run rate margins are typically extrapolated using a combination of whatever info we have on the new technology (typically from a techno-economic analysis TEA), and variables from comparable companies.</p><p>The purpose of this modeling is to understand <strong>what your upside looks like</strong>. We will model to benchmarks like 10M, 50M, 100M in revenue, and work backwards from a target valuation — if this chemicals business were to be worth 1 billion in public markets with an EBITDA multiple of 9x, that would require a revenue of 111M, so we model to that.</p><p>Of course when you model a pre-seed investment, so so many aspects of the business are unknown, but you are evaluating that with all reasonable numbers this can be a good business. You are making sure that you don’t need to break the laws of physics for things to work out… unless you’re investing in fusion that is ;).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/275/1*7ULaKFGPhjkqA_RlIUEFxQ.jpeg" /></figure><p>The other critical part of a model is understanding where your sensitivities lie. A great example of this is a commodities business in which the price of nat gas can make you a monstrous business at one price, but with a drop of $50 you are bleeding money (aka the history of fracking).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lPuU_uSp-fx8GSD9GwnltQ.png" /></figure><p>More on this in the industry heuristics section.</p><h3><strong>Got to spend money to make money — How much de-risking do you get per dollar invested</strong></h3><p>The reason that most investors shy away from capital intensive businesses is that unlike a sass business which has the potential for exponential growth at any moment, the growth of a capex intensive business is constrained by capital input.</p><p>The pre-revenue de-risking process looks very different for different industries, so we will talk through a few of the major categories now.</p><h3>Industry Heuristics</h3><h4><strong>Biotech:</strong></h4><p>Biotech is probably the most established and accepted of all the capex intensive startup models and that is because the path is well trodden, and the exit opportunities are known and lucrative.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NFPdBY6Z-4D4NX55ZcrvNw.jpeg" /></figure><blockquote>Technical risk — High, Market risk — Low</blockquote><p>De-risking:</p><p>The path of a biotech is largely the same (though there are many flavors). Develop a drug candidate, test it in vitro, go through clinical trials phase 1, 2, &amp; 3, and finally scale up and sell.</p><p>Exits:</p><p>Functional drugs for pervasive diseases are extremely lucrative and so the burden of proof for a startup is really only to create a drug that works, and then you are likely to be acquired by a large pharmaceutical company.</p><p>For this reason early stage investors are looking to get you through one or multiple phases of de-risking, and upon success, there are many investors likely to fund you.</p><h4><strong>Hardware</strong></h4><blockquote>Technical risk — Med, Market risk — Med/High</blockquote><p>De-risking:</p><p>In a hardware business you have to design your product, then prototype, then mass produce &amp; sell. There are many forms this could take but let’s use Tesla as an example.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/198/1*O-PK0hI8qDjYJ4mnoNxIZw.png" /></figure><p>All of your technical de-risking happens pre-revenue. You make sure that you can make an electric car, with specifications that meet people’s needs, at a price point that people will pay. And your early rounds of funding surround your accomplishment of these internal goals.</p><p>Once you are able to sell your product, you can finally conclusively answer the question of will people buy this. But likely you have been trying to answer this question every day until then. (unless you’re Elon musk and you can sell the ugliest car to grace the planet, because you’re a savage)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/273/1*5A_z0TTPWKmNCsRdonCYrw.jpeg" /></figure><p>Exits:</p><p>For hardware businesses, an exit via the public markets is just as likely as an acquisition.</p><p>Hardware businesses have an added layer of risk that biotechs do not, which is the risk that the market may be smaller than you had estimated.</p><p>Yes! It’s risky, but risk generates returns.</p><p>I worked with my colleague Russ Willcox to outline some of the most powerful models for hardware:</p><ul><li>The hardware is really a way to sell software. For example, credit card readers.</li><li>The hardware is a component in something bigger, so for a small effort you can gain a big premium. Like NutraSweet in diet drinks.</li><li>There is a big scale economy, or the hardware is so complicated that in either case, once you achieve breakout, others cannot copy. Like Tesla.</li><li>The hardware delivers so much patented value, you actually can charge fat margins and protect them. Like Adaptilens we hope.</li><li>Is this the <strong>first</strong> or <strong>most qualified</strong> of its kind? The market creator. The Tesla of EV’s, the SpaceX of Satellites, the Boom of supersonic commercial planes, the Apple of smartphones …</li></ul><p>While hardware is more complex than software, the reason for its attraction is market size. The vast majority of the economy is physical. We live in houses, move in cars, create energy in power plants, eat food, rush to hospitals when ill, and wear clothes. You can access those markets to some extent with software… but only to an extent. Beyond that you have to deal in atoms, not bits.</p><h4><strong>Materials (chemicals, oil &amp; gas, mining)</strong></h4><p>Now for the really hard and yet most important of them all (if you care about climate).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*3Q_h75UJjTSz4kqOUVDxxA.jpeg" /></figure><blockquote>Technical Risk — High, Market Risk — Med</blockquote><p>De-Risking:</p><p>These are the most challenging because, if you are doing something truly novel (as all of our companies are), you have the great burden and joy of creating a <strong><em>first of a kind plant.</em></strong></p><p><em>First of a kind plants</em> deserve an article of their own, and I’ll get to that someday.</p><p>A materials co’s capital journey involves proving you can make <em>thing x</em> with meaningful process improvement in a scientific setting . Then you must create an initial <em>demo plant</em> to demonstrate creation of <em>thing x</em> at some sort of scale. THEN you must capitalize a full plant for <strong>thing x. </strong>Luckily at this point there is likely non-equity funding available to you (more on that in a moment).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lFJXVwQFTElbPrkx_wybLg.png" /></figure><p>Exits</p><p>The reality is that in materials, rarely can you exit without a scaled operation. I have seen this with many of the successful founders I know in the space.</p><p>What’s good about these businesses:</p><p>The world runs on materials and unless you are inventing a new material, you know your buyers. And you know your buyers only care about getting <em>thing x</em> at a good price, and preferably with minimal emissions (if available for a good price).</p><p>Why I still put market risk as high:</p><p>These businesses are price sensitive and materials prices can be volatile. If you are taking the clean angle you need to compete with everyone else <strong>at cost </strong>in the long run (even if some tax credits are going to prop you up for a few years, you can’t build a business on that).</p><h4><strong>The good news</strong></h4><p>There are lots of non-dilutive ways to fund a business. These include government loans, asset backed loans, grants. which are the key to scaling once you have proved that your tech works and people want it.</p><p>If you are a good business you will be able to grow. But there is an arduous journey to proving you are a good business.</p><p>Look out for an addendum where we’ll dissect some case studies of companies in each category — materials, hardware, and biotech.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6e39e9871967" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Robotics Industry Part 3, Humanoids TO THE MOON]]></title>
            <link>https://medium.com/@alexmreinhart/the-robotics-industry-part-3-humanoids-to-the-moon-e7e6c52fd686?source=rss-949b7a03ae4f------2</link>
            <guid isPermaLink="false">https://medium.com/p/e7e6c52fd686</guid>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[Alex Reinhart]]></dc:creator>
            <pubDate>Wed, 03 Jul 2024 17:08:50 GMT</pubDate>
            <atom:updated>2024-07-05T21:26:52.113Z</atom:updated>
            <content:encoded><![CDATA[<p>Humanoids are having their Crypto moment.</p><p><em>Money is flowing into humanoid robot startups. Companies are raising hundreds of millions at crazy valuations.</em></p><p><em>Imagine, a robot that can replace all human labor.</em></p><p><em>I am hearing accounts from established robotics companies whose investors are asking them — “but can you make it a humanoid?”</em></p><p><em>As a former robotics researcher, I’m here to tell you that making your hardware as complicated as possible is not going to help you automate things more effectively.</em></p><p><em>Please, draw your own conclusion from the accounts collected bellow.</em></p><h3>The Story of Boston dynamics:</h3><p>With the guiding principle that you must build something useful to make money, I have long watched the Boston dynamics dancing robot videos, and wondered, how the heck do they make money. Let’s unravel the myth.</p><p>Boston Dynamics was a spinout of a lab at MIT in 1992 working on legged robots funded by military agencies like DARPA. In 2004, Boston Dynamics created Big Dog, a quadruped robot capable of carrying heavy loads over rough terrain.</p><p>There is no public information about how many Big Dogs were sold, but the sole application was in the military for carrying supplies.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nOrWu2fBuuLnzDZq" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/690/0*45Gp3O2Lu2KWjLcF" /></figure><p>I<strong>n 2013, Google acquired Boston Dynamics for $500 million,</strong> but the company struggled to fit into Google’s plans.</p><p><strong>Boston Dynamics’ Atlas humanoid robot was created around 2015 and was never intended to be sold to the public.</strong> It was initially developed for the DARPA robotics challenge to enhance the role of robots in disaster relief and was never sold, always described as a research platform.</p><p><strong>In 2017, SoftBank acquired Boston Dynamics at a 60% discount for $165 million,</strong> and the company began focusing more on commercializing its technology.</p><p><strong>In 2020 Boston dynamics released Spot, </strong>its first commercial robot charging $75,000 for the quadruped and sold a couple hundred over the next few years. This puts a rough estimate of revenue $10 million.</p><p><strong>That same year Hyundai</strong><a href="https://spectrum.ieee.org/hyundai-buys-boston-dynamics"><strong> acquired Boston dynamics for nearly $1 billion</strong></a>.</p><p>In 2024 Boston Dynamics unveiled plans to sell a newer, fully electric generation of the Atlas humanoid robot designed for real-world applications like moving objects in factories. The cost per robot is estimated to be around $150k.</p><p>The story of Boston dynamics is a quite interesting one.</p><p><strong>They are a state-of-the-art, money-losing robotics research company.</strong></p><p>Their evolution is a metaphor for the challenges faced by robots companies generally. Robots are an epic technology, but they’ve struggled to find product market fit in industry, and much like Boston Dynamics, robotics found their home (or their use) to be inside a car company.</p><h3>The Age of the Humanoid:</h3><p>For the past few decades Boston Dynamics was THE humanoid company, now there is a proliferation of humanoids and VC investment flowing into the space, and I think everyone should take a breath.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gMNLP0XN3_zVw5ow" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cDcrtkTgfCcd8xHF4WwhWA.png" /></figure><p>It may seem surprising, given how frequently we see humanoid robots in YouTube videos, but not one company has actually sold a humanoid<strong> for any use other than entertainment or research</strong>. And not one humanoid is being used as a true resource to drive efficiency.</p><p>Obviously there are many who believe that humanoid robots are the solution to all our automation problems and maybe someday we’ll live in a world like this…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/497/0*h1eujSg-oPfjMIuX" /></figure><p>Well a few weeks ago I went to a Robotics in Industry summit and spent a lot of time talking to the leaders in commercialization, this was my experience:</p><blockquote><em>I went to an event that was mostly populated with robotics researchers, and a majority of the discussion revolved around humanoids &amp; AI foundation models for robotics.</em></blockquote><blockquote><em>Later that day I went to the industry expo for companies in robotics, and</em><strong><em> there was </em></strong><em>not a humanoid to be seen. I asked around at the 100&#39;s of booths, and each company laughed when I asked them about humanoids telling me that they had enough complexity with 6 degrees-of-freedom.</em></blockquote><blockquote><em>What this confirms in my mind is the propensity for academic &amp; media interests to gravitate to what’s possible, and industry to gravitate to what’s useful.</em></blockquote><blockquote><em>It’s easy to conflate what you see often, with the true prevalence of an idea.</em></blockquote><p>In pondering the value of all this media attention and investment with the head of engineering of <a href="https://vervemotion.com/">Verve Motion</a> a robotic exoskeleton startup, he said something that I quite liked:</p><blockquote><em>Going to the moon didn’t solve any problems or generate any revenue, but it did drive innovation, and create many valuable technologies that went on to find broad market penetration.</em></blockquote><blockquote><em>So in theory, the development of humanoid robots would have the same effect on the robotic technology, whether or not there is an immediate use case for them.</em></blockquote><p>The same day I had a discussion with various industry veterans from Guardian Ag and Realtime Robotics, on choosing the right hardware platform. I found this comment funny and insightful:</p><blockquote><em>Let’s say you are choosing whether to put wheels or legs on your robotic system. Wheels are about 1/100 of the bill of materials cost, so you would nearly always choose wheels if you’re thinking practically.</em></blockquote><blockquote><em>What if there are stairs? Well, why are there stairs? and can you not change the surroundings that this robot will operate in. Navigating a ramp is going to be much more reliable than climbing a set of stairs.</em></blockquote><blockquote><em>Reliability is the single most important thing in robotics. Simplify the problem. Don’t over engineer the solution.</em></blockquote><p>In summary, humanoid robots are an amazing intellectual pursuit for the good of humanity and I’m so glad that the best minds are focusing on it.</p><p>Given the rate of change of technology today,<strong> I wouldn’t bet against anything</strong>. I remain reserved on the proximity of monetization and I prefer to see robotics companies solving a task, not trying to automate a human.</p><p>I would love to be proven wrong.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e7e6c52fd686" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Robotics Industry Part 2, the Control Conundrum]]></title>
            <link>https://medium.com/@alexmreinhart/the-robotics-industry-part-2-the-control-conundrum-32b5265d0ad6?source=rss-949b7a03ae4f------2</link>
            <guid isPermaLink="false">https://medium.com/p/32b5265d0ad6</guid>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[robotics]]></category>
            <dc:creator><![CDATA[Alex Reinhart]]></dc:creator>
            <pubDate>Wed, 03 Jul 2024 13:02:44 GMT</pubDate>
            <atom:updated>2024-07-03T13:02:44.497Z</atom:updated>
            <content:encoded><![CDATA[<p>The evolution of robotics control systems. Is model based control the answer to our prayers?</p><h3>Setting the stage, what is a robot?</h3><p>We all know robotic arms &amp; humanoids, but what exactly defines a robot?</p><p>A robot is a piece of mechanical equipment that can perform a <em>variety</em> of tasks. When thinking about robotics in this way it separates robotics from the rest of the advanced manufacturing world. In automated factories, machines automate the manufacturing of products in a single purpose way (<em>how its’ made</em> style).</p><p>Ice cream sandwiches on <em>How it’s Made-</em></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FVlb4mfQv6-s&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DVlb4mfQv6-s&amp;image=http%3A%2F%2Fi.ytimg.com%2Fvi%2FVlb4mfQv6-s%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/9ff114df2b106e94b6e49e456843dd6f/href">https://medium.com/media/9ff114df2b106e94b6e49e456843dd6f/href</a></iframe><p><strong>Automation</strong> is a broad term used for special purpose machines or systems that are designed to automate tasks.</p><p><strong>Robots </strong>are a piece of equipment that can be programmed to perform a variety of tasks efficiently.</p><p><strong>Though robots were intended to be complex reprogrammable systems for automation, the reality of their use is that they tend to do one thing over and over.</strong></p><p>Re-programming robots is not easy, so for reliable use, we put them in factories to weld cars, solder transistors, move heavy parts, in all cases just following the same predetermined path over and over.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*KYj0amDfrcJzMqqq" /></figure><p><a href="https://www.therobotreport.com/automotive-industry-sets-record-1m-robots/">Automotive manufacturing using robots</a></p><p>Using robots in this way has allowed us to replace significant amounts of human labor. But to achieve the next level of labor replacement we must make robots more dextrous and more adaptive.</p><p>There are many tasks that we’d like to automate that don’t happen in controlled settings OR where the cost of adoption of robots is insurmountably high.</p><p>The main challenge in robotics has long been connecting the dots between robotic hardware and real value producing use cases. The crux of that is our control systems, and in this article we will discuss how robotic control systems have evolved and how model based control is attempting to unlock new applications.</p><h3>The Evolution of Robotic Control:</h3><p>Most robotics today use rule based control. This means that whether the robot is following a path, generating a path, navigating a warehouse, or any other application. It’s control system is mainly a set of hardcoded rules, actions and responses. Why? Because this is really reliable, and reliability is the #1 most important thing for commercial robots.</p><p>Roboticist&#39;s maintain a distinction between the hardware capabilities and software controls of their systems. Thus they have defined two categories of robotic intelligence — <strong>athletic intelligence</strong> and <strong>cognitive intelligence</strong>.</p><p><strong>Athletic intelligence</strong> refers to the hardware and physical systems. For complex robotics systems, athletic intelligence encompasses how to make robots that mimic human/animal motion, balance, increasing precision and dexterity, etc.</p><p>Boston Dynamics has pushed the limits of athletic intelligence with their quadruped and biped robots. At the same time companies that are working to miniaturize hardware, or make soft robotic grippers for grasping delicate things, are also advancing athletic intelligence.</p><p>This is one of the coolest demonstrations of technology I’ve ever seen-<a href="https://youtu.be/fn3KWM1kuAw"> Boston dynamics video</a></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Ffn3KWM1kuAw%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dfn3KWM1kuAw&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Ffn3KWM1kuAw%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/d5f462100ddb7be9baa1e363ffd3de37/href">https://medium.com/media/d5f462100ddb7be9baa1e363ffd3de37/href</a></iframe><p><strong>Cognitive intelligence</strong> is the connection between robots and their surroundings. This pursuit began with path planning for the trajectory that your robotic arm will take, and now encompasses complete autonomy, goal based path generation, resilience to perturbance, failure recovery.</p><p>The latest development in cognitive intelligence has been the concept of <strong>foundation models for robotic control</strong>, we will discuss this in depth in a moment.</p><p>Lack of cognitive intelligence is the barrier between most robots and practical use.</p><p>The world sees the progress with language models and thinks that cognitive intelligence is solved. But doing physical things with physical robots is different than predicting the next word or image.</p><p><strong>There are 3 components to physical systems that make them much harder.</strong></p><ol><li>There are “hard” constraints on generated actions (i.e. you cannot violate physics)</li><li>The costs of failures are higher. Whether it’s a manufacturing mistake, calibration errors, or a crashed robot, the time to perceive the mistake, and time to correct are much longer.</li><li>Finally physical systems are much more data limited, and collecting specifically applicable data takes significant time and effort. Though there are some impressive simulation to reality demonstrations that suggest that our simulations are becoming realistic enough to be useful.</li></ol><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//x.com/JasonMa2020/status/1786425771034374609&amp;image=" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/b793317b544b1b7dbe1204e0596157af/href">https://medium.com/media/b793317b544b1b7dbe1204e0596157af/href</a></iframe><h3>Comparing the pivotal papers in robotic foundation models</h3><h3>Control of robots via language</h3><p><em>Attention is all you need</em>, vs<em> RT-1 &amp; 2: Robotics transformer 1 and 2</em>.</p><p>The ‘attention is all you need’ paper first published in 2017 popularized the transformer as a mechanism for improving foundation model performance using a concept called attention. <em>Attention</em> a mathematical way to encompass how tokens affect each other, how an adjective affects a noun, how a prefix affects a suffix, and all the other ways that grammatical components come together to create meaning.</p><p>What we can learn from the success of transformers in large language models (LLMs) is that you need to do more than vectorize your current state and predict the next step.</p><p>There is a level of pre-processing, added understanding that makes your input more content-full. This data centric premise was soon applied to other domains, robotics being one of them.</p><p><strong>The key learnings here are — You can train on of task-agnostic data &amp; scale matters, a lot.</strong></p><p>In 2022 RT-1 was published introducing the concept of Robotic Transformers (RT). The first RT was trained on 800k samples of data using 2 brands of robotic arm.</p><p>This first RT foundation model was able to generalize <strong>known tasks</strong> (variants of pick and place) across many different scenarios and work across the two robotics arms.</p><p>RT-2 focused on bringing this dataset to the internet scale, taking another learning from LLMs that more data creates better models.</p><p>The Robotics Transformer 2 (RT-2) model released in late 2023, <strong>enabled robots to perform tasks they have not been explicitly trained for</strong>. This transformer also focused on pick and place, but this time it included recognition of many different objects, and many different settings for the action.</p><p>RT-2 was trained on <strong>general image data </strong>to give it a foundational understanding of the world, and then combined with a similar<strong> robotic action model</strong> to what was used in RT-1.</p><p>This increased understanding helps the model generalize better to novel objects, scenarios and tasks. RT-2 can understand natural language instructions and translate them into robotic actions.</p><p><strong>TLDR:</strong></p><p><strong>Best robotic control system performance = General vision model + General Language Model + Robotic action model</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4ZXDOGSziC7hCAYxrD9Wog.png" /></figure><h3>Control of Robots via demonstration</h3><p>Another strategy being pursued is the concept of fine tuning via demonstration data to teach a specific task.</p><p>Based on the demo videos I’ve seen I believe this is how Tesla is training their Optimus robots. — <a href="https://www.instagram.com/reel/C8QJDF0PkM1/?igsh=ejQ1dWRydnU4N3cz">https://www.instagram.com/reel/C8QJDF0PkM1/?igsh=ejQ1dWRydnU4N3cz</a></p><p>The Toyota Robotics Institute (TRI) demonstrated success teaching robots new actions starting with foundation models and teaching via repeated demonstration.</p><p>They were able to teach a pair of 6 degree-of-freedom robotic arms to wash dishes, flip a pancake, mop, beat eggs, and do dishes.</p><p>This is unquestionably impressive but the real test of its usefulness will be the robustness with which robots can follow these directions.</p><p>Today there are some startups commercializing this technology for simple tasks (placing components on a circuit board or assembling boxes). Automating these tasks is not simple, but constraining the problem may allow for a robust product to be built.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fw-CGSQAO5-Q%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dw-CGSQAO5-Q&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fw-CGSQAO5-Q%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/6130de8fb48dab409a26b66e0000f28b/href">https://medium.com/media/6130de8fb48dab409a26b66e0000f28b/href</a></iframe><p><a href="https://www.youtube.com/watch?v=w-CGSQAO5-Q"><strong>TRI demo video</strong></a></p><h3><strong>In Summation</strong></h3><p>There are two components of robotic intelligence: cognitive and athletic.</p><p>Most robots today still use rule based control.</p><p>Language model learnings are being adapted to robotic foundation models in an attempt to make them robust and useful.</p><p>Model based control is taking robots to the next level of cognitive intelligence, but the first test will be <strong>can a model based robot compete with a rule based robot on simple tasks.</strong></p><p>The second test will be can model based control unlock new business applications for robotics.</p><p>I look forward to academics and companies alike putting forth reliability metrics. What I bwant to know is — How often is the plan the model creates perfect, how often is the robot able to execute on that plan, how well is it able to recover from failure, and how are reliability numbers affected if the surroundings are perturbed.</p><h4>The robotics investments I‘m interested in making —</h4><ol><li><strong>New robotic simulator technologies. </strong>The best simulators will create the best synthetic data, and hold the keys to the best robotic models.</li><li><strong>The next pick and place.</strong> By this I mean that pick and place was a vertical robotics play — robotics startups owned that problem and were able to automate a complex thing effectively. I’m intrigued by putting robots into slightly more complex automation problems such as truck unloading and loading, bicycle and car repair, diagnostics and decision making. Show me a new automate-able problem with a big market, and I am interested.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=32b5265d0ad6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Robotics Industry Part 1, a Market-Map of Sorts]]></title>
            <link>https://medium.com/@alexmreinhart/robotics-part-1-a-market-map-of-sorts-b23573780d1f?source=rss-949b7a03ae4f------2</link>
            <guid isPermaLink="false">https://medium.com/p/b23573780d1f</guid>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Alex Reinhart]]></dc:creator>
            <pubDate>Mon, 01 Jul 2024 19:33:09 GMT</pubDate>
            <atom:updated>2024-07-03T17:09:24.193Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The startup landscape of robotics is vast, so rather than trying to collect a bunch of meaning-less logos together of which you recognize half, we will instead take a segment by segment approach to organize the robotics industry.</em></p><p><em>Specifically, we’ll look to uncover, how much funding and tangible success (in the form of exits) has accrued to each of these applications.</em></p><h3>The Factory Automation Stack</h3><p>37% of all industrial robots installed worldwide are in the automotive sector. But any robotic factory will have the same characteristics, here are the businesses that come together to create robotic factories.</p><h3><strong>HARDWARE</strong></h3><h4><strong>Robot Manufacturers:</strong></h4><p>Robot manufacturers make robots, control systems, and platform level software for control of their hardware. Most of these companies have been long established, founded in the 70s — 90s and dominating the industrial robotics space.</p><p><strong>Example companies:</strong> FANUC (market cap $27 billion), ABB (market cap $103 billion), Mitsubishi Electric (market cap $34 billion), Kawasaki Robotics (market cap $6.5 billion), DENSO (market cap $9 billion).</p><p>There are also robotics manufacturers who produce more research friendly and cobot oriented robots have a similar profile if a slightly smaller market — Universal Robots (market cap $1.5 billion).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*t2KxvgTLvTR6xJpH" /></figure><h4><strong>Robotic Components Manufacturers:</strong></h4><p>Components manufacturers make specialized components for robots, often end-effectors or sensing systems. These companies will try to dominate the market in one skill whether it’s a dexterity related skill or perception based skill (for example a unique set of sensors to enable autonomy).</p><p>These companies were founded in the 2000’s when industrial automation had experienced wide adoption, the low hanging fruit had been automated and there was opportunity for new sensors and grippers to unlock new applications.</p><p>We are early in the game to see how these robotic tools businesses will play out.</p><p><strong>Example companies: </strong>The company<strong> </strong><a href="https://www.softroboticsinc.com/">Soft robotics</a> makes novel grippers &amp; autonomy software for grasping soft objects (raised 26M Series C),<a href="https://toposens.com/"> Toposens</a> makes sensors for robot autonomy (raised 4M seed round), <a href="https://www.shadowrobot.com/">Shadow Robot</a> makes robotic hands pictured below (funding undisclosed).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zoJ0G4OSKUzycPM1" /><figcaption><a href="http://shadowrobot.com/">Source</a></figcaption></figure><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F4OoFQ3Asj74%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D4OoFQ3Asj74&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F4OoFQ3Asj74%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/46cdaf98269670f092947d5383333c47/href">https://medium.com/media/46cdaf98269670f092947d5383333c47/href</a></iframe><h3><strong>HARDWARE + SOFTWARE</strong></h3><h4><strong>Robot Integrators:</strong></h4><p>Robot integrators are the companies that automate your factory line for you. They audit the current manufacturing process, design the robotic system, and deploy that system. Integrators typically also install the software for ongoing monitoring &amp; control of that system, but don’t build these systems themselves.</p><p>These are services business.</p><p><strong>Example companies:</strong> JR Automation (acquired by Hitachi for 1.425 billion in 2019), ASG, Reis Robotics (Acquired by KUKA amount undisclosed) , Acieta (acquired by Mitsui &amp; co), Siemens</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/388/0*4AV2EOT6APNYpxu_" /></figure><p><strong><em>Observation</em></strong><em> — Many of these integrator businesses were swept up and acquired by the robotics manufacturers, this makes sense in that owning the full stack automation service helps you sell more of your robots.</em></p><h3><strong>SOFTWARE</strong></h3><h4><strong>Robotic Simulators &amp; programming tools:</strong></h4><p>Robotic simulators allow users to simulate, program, and optimize industrial robot applications offline without disrupting production.</p><p>These 3D modeling + physics engine tools allow you to import all relevant CAD files, program paths across all robotic arms and simulate the sequences you want to run with no risk of collision. This is particularly critical in programming co-bot (collaborative robot, multi robot) systems.</p><p><strong><em>Aside</em></strong><em> — Many believe that the unlock for creating AI systems to control robots is simulation, and these industrial simulators have a wealth of real data in this category.</em></p><p><strong>Example companies:</strong><a href="https://robodk.com/"> RoboDK</a> (funding unknown), Gazebo (open source), Cadence Design Systems (market cap of $87B, do many things) , Simscape multi body (created by Matlab), Isaac Sim (created by NVIDIA), FANUC ROBOGUIDE,<a href="https://intrinsic.ai/flowstate/"> Flowstate</a> by Intrinsic AI (started by Alphabet)</p><p><strong><em>Observation</em></strong><em> — Interesting that there haven’t been many startups accruing value in this space just platforms created by the big players and open source tools.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OwrglQFqR4jp5kEs" /></figure><h4><strong>Robot Operating Systems (ROS)</strong></h4><p>You can think of ROS as the lowest level software for robots, analogous to what CUDA is for NVIDIA GPUs. But unlike CUDA this is a non-proprietary industry-wide software</p><p>It allows developers to write robot software without needing to know the specifics of the underlying hardware. These libraries provide easy control of joint angles, movement, and failure recovery.</p><p>ROS also offers communication protocols for how robots should interact with other systems.</p><p>ROS is an open source project that was first a non-profit before it was acquired by Alphabet. Still the dominant platform for robot programming.<a href="https://github.com/eranet/rhombus"> Rhombus</a> is an alternative, also an open source product.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*rNGnznN1eAoICigG" /></figure><p><a href="https://www.theconstruct.ai/start-coding-on-ros/">Source</a></p><h3>The Warehouse Automation Stack</h3><p>The second wave of robotic automation took place in warehouses.</p><p>Factories are the most controlled and constrained environments, which was why they were the primary automation target. The warehouse was the next challenge, requiring the addition of navigation, and tackling of a fundamentally less constrained problem.</p><p>For this reason you see little overlap between the companies targeting the two sectors.</p><p>Warehouse robots are all about <strong>mobility</strong>.</p><h4><strong>Warehouse Fleets</strong></h4><p>Robotic fleets companies provide hardware and often full stack solutions for warehouse automation. The two companies mentioned below focus on locating the target object in the warehouse, and bringing that object to a central picking location.</p><p><strong>Startups</strong>: Locus robotics (raised an $117M series F), KIVA (Acquired by Amazon for $775 million), Symbotic is the largest of the warehouse automation companies with a market cap of $21 billion.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FGcRiiJLuO7zuz06" /></figure><h4><strong>Pick and place ‘in a box’:</strong></h4><p>The other component of object retrieval is pick and place. This is an area that robotics startups have been able to build a lot of value.</p><p>The primary form of pick and place was station based in which you have a station optimized for pick and place with items coming to you. The more recent wave has been more complex pick and place such as truck unloading.</p><p><strong>Startups</strong>: Covariant ($222 million raised to date) is an autonomous pick and place company. <a href="https://www.nimble.ai/">Nimble Ai</a> ($65 million series B) is building their own robotic warehousing solutions for CPG and consumer brands.<a href="https://apptronik.com/"> Apptronik</a> ($32.8 million) is working to create the next generation of this in the form of a humanoid. Pickle Robots is building robots for truck unloading (raised a $26 million series A in 2022)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UeiwMGvhU8jtORIr" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uwV4vajjIpBDQl_M" /></figure><h4><strong>Robotic Warehouse Management Software:</strong></h4><p>Warehouse management software is needed when you are using multiple robotic solutions, maybe that’s pick and place + item retrieval to coordinate robotic activities. This is the <em>hive-mind</em> of your automated warehouse, and its actually one of the most important components. Think of this as a digital twin of your warehouse that controls the tasks of your robots.</p><p><strong>Startups</strong>:<a href="https://www.inorbit.ai/"> InOrbit</a> digital twin software for robots shown below, tracks product as it moved through a warehouse ($3.3 million seed in 2023),<a href="https://unmanned.life/technology/"> Unmanned Life</a> does the same offering a cross hardware robot management solution ($4.8 million seed in 2022),<a href="https://www.dexory.com/"> Dexory</a> (previously known as BotsAndUs) ($13 million in 2022) captures real-time insights for warehouses using fully autonomous robotic fleets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*aSbByWoT9go6LlDw" /></figure><p><strong><em>Observation</em></strong><em> — if you look at the value accrued by companies in each of the above categories it tells an interesting story.</em></p><p><em>The first startups to build meaningful value were the startups solving the warehouse navigation problem, in one way or another they would ‘carry the boxes for you’. The next set of companies tackled pick and place.</em></p><p><em>We have yet to see the exits, but the valuations in the space suggest meaningful value. The final category, Warehouse Management, has only become necessary because of the success of the earlier two categories of warehouse robots.</em></p><p><em>Increasing automation will make businesses like these increasingly valuable.</em></p><h3>Healthcare Robotics</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*KgBaypUggRC3CM6c" /></figure><p>The healthcare robotics is not an autonomy game but rather for a game of precision, ease and dexterity. All of these surgical robots are manually controlled meaning that there is no cognitive intelligence risk.</p><p><strong><em>Observation</em></strong><em> -The Da Vinci robot is a general platform for surgery and thus has a huge market but as with any platform solution, there is opportunity in specialization.</em></p><p><strong>Startups</strong>: <strong>Intuitive Surgical </strong>($141 billion market cap) is the creator of the Da Vinci robot for robotic surgery. <strong>Moon surgical</strong> ($55.4 million raised) is creating robots for surgery, <strong>Wandercraft</strong> ($45M series C) develops an exoskeleton for inpatient rehab and outpatient daily life, <strong>CMR Surgical</strong> ($164M Series D) develops robots for endoscopic surgery</p><h3>Construction Robotics</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vwMnSyDReVppja5h" /><figcaption><a href="https://iconbuild.com/">Icons 3D printed home technology</a></figcaption></figure><p><strong>Startups</strong>: <strong>Monumental AI</strong> ($24M seed) is creating autonomous bricklaying robots, <strong>Built robotics</strong> ($112 million) built an autonomous pile driver. <strong>Icon</strong> ($451.5 million raised) is ‘3D printing’ homes.</p><h3>Lab Automation Robotics</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*aZ1t9CiTQkTxxxoj" /></figure><p><strong>Startups</strong>: <strong>Reshape Biotech</strong> ($20M Series A) is building lab automation robots, <strong>Opentrons</strong> ( $240 million raised to date), <strong>Automata</strong> ($99.8 million raised to date).</p><h3>Underwater Robotics:</h3><p>The underwater robotics industry has tackled the problem of building tools for uninhabitable environments doing everything from imaging &amp; sensing, to manipulation. The big consumers of this technology include oil and gas, ocean construction, and military.</p><p><strong>Startups</strong>: <strong>Kraken Robotics</strong> ($209 million market cap) builds a wide variety of underwater autonomous vehicles and robotic manipulators, <strong>Terra Depth</strong> ($30.1 million raised) is autonomous underwater vehicles (AUVs) for ocean data collection and research. <strong>Neptune</strong> ($17.3 million raised) robotics developing hull cleaning robots that can operate above and below water to remove fouling while protecting anti-fouling paint.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/920/0*FMzK6CgZBbZxUPvu" /><figcaption>Source Kraken Robotics</figcaption></figure><h3>Agriculture Robotics</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/653/0*YkjsTBFFl6wbqMv2" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/708/0*TyF7ZwboB754AcYJ" /></figure><p>Agriculture robots are taking on problems ranging from harvesting, to weeding, to spraying fertilizer, to collecting data on crops. Crop field autonomy is a logical step up from warehouse autonomy.</p><p><strong>Startups</strong>:<a href="https://sagarobotics.com/"><strong> Saga robotics</strong></a> makes farming robots (11M Series A), <strong>Abundant</strong> <strong>Robots</strong> has raised $12 million in funding to build robots that pick strawberries, <strong>Mineral</strong> was a google X company that was creating autonomous systems for crop data collection, but eventually pivoted to adding their sensing suite onto existing equipment (as shown in image above). Guardian Ag ( $35 million raised to date), one of our portfolio companies, is building autonomous drones that spray fertilizer on large fields.</p><h3>Home Robotics</h3><p><a href="https://maticrobots.com/blog/matic-comes-out-of-stealth-with-30m-in-funding/">Matic’s home robot</a> is quite cute. Looks like a bulldozer toy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zpIFklKkbi-LyqrX" /></figure><p>Home robotics is the long established market dominated by vacuum cleaners and more recently lawn mowers. This space has been largely stagnant, as it is quite difficult to put machines in people’s homes, and consumers are a fickle group to sell to. I’m optimistic that there are more processes that can be automated in the home.</p><p><strong>Startups</strong>: <strong>iRobot</strong> builds the autonomous vacuum cleaners that created the home robotics market ($385.35 million dollar market cap), <strong>Matic</strong> ($30 million raised to date) is another home cleaning robotics company that builds an autonomous vacuum with “the most advanced 3D mapping and spatial technology on the market”. <strong>Electric Sheep</strong> ($21.5 million dollars raised to date) is building autonomous lawnmowers.</p><h3>Robotic Foundation Models:</h3><p>Most of the startup landscape is targeting solving a tactical problem with robots. But there are some companies making a bet that building the best control technology will aggregate value.</p><h4><strong>Targeted Foundation models:</strong></h4><ul><li><strong>Covariant</strong> is making foundation models for pick at place.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/0*h9i0nUFHj8kx83_r" /></figure><ul><li><strong>Field AI</strong> ($30.2M series A) is creating Field Foundation Models (FFMs) which is a model focused on robot autonomy.</li></ul><h4>General Foundation models:</h4><ul><li>NVIDIA created a project called Gr00T a general-purpose foundation model for humanoid robots.</li><li><a href="https://www.crunchbase.com/organization/skild"><strong>Skild</strong></a>, a startup out of Carnegie Mellon that foundation models robots control, is raising $300 million from <a href="https://www.crunchbase.com/organization/lightspeed-venture-partners">Lightspeed Venture Partners</a>, <a href="https://www.crunchbase.com/organization/coatue">Coatue</a> and others at a $1.5 billion valuation.</li><li><a href="https://www.crunchbase.com/organization/collaborative-robotics"><strong>Collaborative Robotics</strong></a> raised<a href="https://news.crunchbase.com/robotics/startup-collaborative-raise-general-catalyst/"> $100 million</a> in a Series B led by <a href="https://www.crunchbase.com/organization/general-catalyst-partners">General Catalyst</a> is also making a software foundation model for robots.</li><li><strong>Field AI</strong> ($30.2M series A) is creating Field Foundation Models (FFMs) which is just a different take on a robotic foundation model. <strong>Figure</strong> (675M) is developing Robotics foundation models + humanoids for robot control.</li><li><a href="https://everydayrobots.com/"><strong>Everyday Robotics</strong></a> a Google X company and has also been a major contributor to the RT1 &amp; 2 Papers.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/503/0*Ntls78lCI4qU1i8B" /><figcaption>Everyday Robots</figcaption></figure><h3><strong>Humanoids</strong></h3><p>I know what you’re thinking.</p><p>She’s forgetting humanoids?? There is far to much to be said. Check out part 2 for more on that.</p><h3>In Summation.</h3><p>There are many plays in robotics that have grown into successful and massive businesses. Most of these make money automating narrow tasks.</p><p>Venture money is accruing to new strategies of model based control. This is a promising strategy, but the models must be robust, today almost no robotics systems use pure model based control. Demos mean little in the game of automation. I look forward to seeing customers and data.</p><p>I prefer to see targeted strategies aiming to automate tasks, that inch down the complexity road. (aka an increase in complexity akin to the factory -&gt; warehouse jump)</p><p>Automation markets are big because they replace people, and people are often the most expensive part of a business. <strong>There is plenty left to automate, from repair, to cleaning, to rescue, to space manipulation.</strong></p><p>I look forward to meeting roboticists solving those problems.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b23573780d1f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The New NIMBY, Nuclear in my backyard please]]></title>
            <link>https://medium.com/@alexmreinhart/the-new-nimby-nuclear-in-my-backyard-please-3b4243f9e436?source=rss-949b7a03ae4f------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b4243f9e436</guid>
            <category><![CDATA[nuclear]]></category>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[climate-change]]></category>
            <category><![CDATA[energy]]></category>
            <category><![CDATA[fusion]]></category>
            <dc:creator><![CDATA[Alex Reinhart]]></dc:creator>
            <pubDate>Fri, 07 Jun 2024 16:31:50 GMT</pubDate>
            <atom:updated>2024-06-07T16:31:50.055Z</atom:updated>
            <content:encoded><![CDATA[<p><em>jokes.</em></p><p><em>But the purpose of this article is to break the misconceptions about Fission energy, discuss the state of Fusion technology, and give us an informed perspective from which to evaluate the policies that currently have the US nuclear industry at a standstill.</em></p><p><em>There are many paths to decarbonization of our energy infrastructure, and nuclear is a strong candidate. You should know the pros and cons yourself.</em></p><h3>How does nuclear energy work?</h3><p>Nuclear energy is energy released from the nucleus of an atom through nuclear fission or nuclear fusion reactions. So if you think about the periodic table, you are turning an element into another element by changing the number of atoms in the nucleus.</p><h3>Fission: [currently commercialized]</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/780/0*Xw9jwTr3QvLWKE9H" /></figure><p>Fission is the splitting of heavy atomic nuclei like uranium or plutonium. In a reactor, neutrons are fired at the uranium atoms causing them to split and release energy in the form of heat. Heat produces steam which drives turbines to produce electricity.</p><p>A controlled chain reaction is maintained in the reactor core by regulating the flow of neutrons.</p><p><strong>Drawbacks:</strong></p><ol><li>Fission produces radioactive waste.</li><li>A typical 1 gigawatt fission nuclear reactor produces around 20–30 tons of highly radioactive spent nuclear fuel waste per year. This needs to be stored in secure geological repositories for thousands of years.</li><li>Fission has the same characteristics as an atomic bomb, for this reason the materials are highly regulated. Fission plants have the possibility of a nuclear meltdown if you lose control of the chain reaction. In history the famous meltdown incidents have been due to gross mismanagement (chernobyl) or natural disasters interfering with plant operations (3 mile island).</li></ol><p>There is a valid concern as to the risks of fission, but as with anything, looking at specific and emotionally triggering incidents can cloud your perspective to the realities of a technology.</p><p><a href="https://ourworldindata.org/safest-sources-of-energy#:~:text=Nuclear%20energy%2C%20for%20example%2C%20results,and%2097.6%25%20fewer%20than%20gas">This article</a> talks about the relative death rates of nuclear in comparison with all other energy sources. Just like with sharks and self diving cars, our instinct is to overreact.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tmz6iq_Q4ShTKxaR5gz8xA.png" /></figure><h3>Fusion: [ not yet possible]</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/684/0*eQSbPdTisOuig-If" /></figure><blockquote>Deuterium (heavy hydrogen isotope) + Tritium (extra heavy hydrogen isotope) -&gt; Helium + Neutron + Energy</blockquote><p>On the opposite end of the spectrum from fission is fusion. In fusion you are joining two light atomic nuclei, typically hydrogen isotopes, to form a heavier nucleus which releases immense energy.</p><p><strong>Fusion is how the sun works,</strong> combining hydrogen nuclei into helium. <em>Wouldn’t that be an elegant solution to all our energy problems?</em></p><p>Fusion reactions require high temperatures to overcome repulsion of nuclei. Fusion fuels are abundant (hydrogen from water) and the reaction produces little radioactive waste.</p><p><strong>Drawbacks:</strong></p><ol><li>We can’t do it yet</li></ol><h3>Types of Fusion</h3><ol><li><strong>Magnetic Confinement Fusion:</strong> This is the leading approach used by Commonwealth Fusion Systems (2B raised). It requires using a magnetic field to suspend a plasma (a piece of sun), and using that to achieve fusion.</li><li><strong>Inertial Confinement Fusion:</strong> This strategy requires compression of a pellet of D &amp; T using lasers. What’s the inertial component? This compression is achieved by firing lasers or particle beams onto the outer surface of the pellet, which ablates the outer layers and drives an inward-moving shock wave, aka pressure.</li><li><strong>Muonic fusion:</strong> Muonic fusion is a process that allows nuclear fusion to occur at much lower temperatures. It involves the use of muons, which are unstable subatomic particles similar to electrons but with a mass around 207 times greater.</li></ol><p><em>How are muons created?</em></p><p><em>Muons can be artificially produced in large quantities using particle accelerators</em></p><ol><li><em>First, protons are accelerated</em></li><li><em>These high-energy protons are then collided with a solid target like nickel, producing pions through the strong nuclear interaction.</em></li><li><em>The pions decay into muons and muon neutrinos, similar to the cosmic ray process.</em></li></ol><p><em>Physics has such fun names for things!</em></p><h3>State of Fusion Technology</h3><p><strong>Key Definitions:</strong></p><p><strong>Scientific Breakeven</strong> — Where the fusion reaction produces more energy than was supplied to initiate it. Q factor — Ratio used to describe ratio of energy out to energy in. Given that we want more energy out that in a q factor of &gt;1 is key.</p><p><strong>Engineering Breakeven — </strong>When the <strong>energy captured</strong> from the fusion reaction is more that the energy that was supplied to initiate it. And note that typical heat energy capture systems have efficiencies of around 35–40%.</p><p><strong>Commercial Breakeven — </strong>When the <strong>energy into the full system &lt; energy captured.</strong> This means that the facility is net energy positive so that it be a profitable energy generator.</p><h3><strong>Status of ICF fusion</strong></h3><p>TLDR <a href="https://lasers.llnl.gov/">Lawrence Livermore National Lab</a> achieved fusion ignition with Q factor (energy multiple) of ~1.5</p><ul><li>On December 5, 2022, an experiment at NIF produced 3.15 megajoules (MJ) of fusion energy output from an input of 2.05 MJ of energy delivered by NIF’s lasers to the target. This was the first time a fusion experiment achieved an energy output greater than the laser energy input, with a Q factor exceeding 1.</li><li>Achieving ignition, where the fusion reaction becomes self-sustaining after being initiated, has been a major goal for inertial confinement fusion research for decades. It demonstrates the fundamental scientific basis for controlled fusion energy production.</li><li>While still far from a practical energy source, as the Q factor needs to reach around 10 or higher, this achievement is being hailed as a “Wright Brothers moment” for fusion energy by overcoming a key scientific hurdle. Nice!</li></ul><h3><strong>Status of MCF</strong></h3><ul><li>Working on creating self sustaining plasmas where the fusion reactions provide a significant portion of the heating, getting closer to net energy output.</li><li>High score is Q of 0.33 for 5 seconds.</li></ul><h3><strong>Status of Muonic fusion</strong></h3><ul><li>The success of muonic fusion is all about how many fusion reactions can 1 muon catalyze. Today the relationship stands at 1:150, which leaves the Q factor at 0.14.</li><li>The work today is to make muons more effective — both by requiring less energy to create them and by putting them in environments where they can catalyze more reactions.</li></ul><h3>What percent of people are working on Fusion vs Fission?</h3><p><strong>Fusion (tiny)</strong> There are about 50 private fusion companies around the world who have raised around $6.2 billion in investment, 25 hq’d in the US.</p><p><strong>Fission (huge)</strong> The nuclear fission market is 305 billion dollars and provides 10% of the world’s energy.</p><h3>Nuclear in the Market today ( All Fission )</h3><p>There are 54 commercially operating fission power plants in the us. A study in 2022 found that 16% of fission power plants were unprofitable with many more plants being marginal or breakeven.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/609/0*cfK-5KvSCH3wkiUu" /></figure><p>The challenge for US Nuclear power is that our natural gas and oil are cheap. Historically US opinion on nuclear hasn’t been positive even today the data shows support around 55%.</p><p>Meanwhile in France fission is now the primary source of electricity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pi3y0ZjVwLSqSSGp" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mpDKe0aVTzu9mydK" /></figure><p><strong>History and Rationale:</strong></p><ul><li>France began rapidly expanding its nuclear fleet in the mid-1970s after the 1973 oil crisis to reduce reliance on imported fossil fuels. Between 1975–1990, 52 new reactors were connected to the grid.</li><li>This allowed France to transition from being an electricity importer to a major exporter</li><li>The nuclear buildup was one of the most rapid de-carbonizations of an industrialized power grid in history, reducing fossil fuel use from 70% to 10% by 1990.</li></ul><p>But we are not here to talk about de-carbonization, we are here to talk about economics.</p><h3><strong>Why does this work in France and not the US?</strong></h3><ol><li>Electricity is State owned by Electricite de France (EDF).</li><li>As a state-owned monopoly, EDF can set regulated electricity prices in France that ensure profitability for its nuclear fleet . This pricing power is not available to private U.S. nuclear operators who have to compete in free electricity markets.</li><li>The US aggressively regulates fission technology, artificially ballooning costs.</li><li>Because nuclear power has grown massively in France EDF benefits from economies of scale.</li></ol><h3>Nuclear in the Market Tomorrow: The Case for Fission</h3><p>Companies have already captured massive value from fission, and there have been country scale demonstrations of success. The impediment to the success of fission is in sentiment and in regulation. The reality is that it has become a bad business to run in the US, and that is why you see investments in the space drying up.</p><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//x.com/Alexmreinhart/status/1795127956982915312&amp;image=" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/d057942220c90dc4df07c8de951be481/href">https://medium.com/media/d057942220c90dc4df07c8de951be481/href</a></iframe><p>Regulating an industry into unviability is the same as regulating against it. How will we feel when there is no more nuclear left in the US. Meanwhile China doubled their nuclear capacity in the past 10 years and plans to do so again in the next 10.</p><p><strong>If policies shift, the tides of fusion might turn, and we might achieve stable renewable, abundant power, using this technology. If policies don’t, fission is going to continue to flounder, and plants will shut down.</strong></p><h3>Nuclear in the Market Tomorrow: The Case for Fission</h3><p>It is true that it doesn’t have the same risks as Fission, but it will have similar challenges of competing with commoditized technology. The question is will there be less regulatory burden, removing the costs associated with that. The answer is yes.</p><p>The Nuclear Regulator Commission (NRC) <a href="https://www.fusionindustryassociation.org/nrc-decision-separates-fusion-energy-regulation-from-nuclear-fission/#:~:text=NRC%20Decision%20Separates%20Fusion%20Energy%20Regulation%20from%20Nuclear%20Fission,-Ensuring%20Regulatory%20Certainty&amp;text=On%20Friday%2C%20April%2014%2C%20the,regulatory%20regime%20as%20particle%20accelerators.">announced that they would separately regulate Fission and Fusion</a> in 2022, which was a huge stride in terms of regulatory support for the technology.</p><p>Companies are chipping away at Fusion technology, and the government has said they <a href="https://apnews.com/article/granholm-nuclear-fusion-energy-emissions-416f497b842ce94292ec6fb009a4087a">hope to have the a Fusion plant within the next 10 years</a>.</p><p>I find this timeline ridiculous given how unpredictable research is, and the fact that the average fission plant takes 10 years to build. But hey, lets shoot for the stars!</p><h3>Conclusion</h3><p>Both Fusion and Fission are viable solutions to provide us renewable, clean, reliable, base-load energy on the grid. Fission is available today, but the current state of regulation in the US makes it a no-go.</p><p>Fusion energy will be viable at an unknowable point in the future. There are competing strategies to make it viable which adds layers of risk. As a venture capitalist who only has so many bullets, there are too many approaches to Fusion to bet on any one today. Not to mention you are too late to bet on any of them at a reasonable price.</p><p>I am intrigued by bets on the success of the industry as a whole. An example would be a bet on the materials that are used by 90% of fusion strategies (deuterium and tritium).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b4243f9e436" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Hydrogen Economy]]></title>
            <link>https://medium.com/@alexmreinhart/the-hydrogen-economy-3de5c249a044?source=rss-949b7a03ae4f------2</link>
            <guid isPermaLink="false">https://medium.com/p/3de5c249a044</guid>
            <category><![CDATA[hydrogen]]></category>
            <category><![CDATA[clean-energy]]></category>
            <category><![CDATA[electrolysis]]></category>
            <category><![CDATA[policy]]></category>
            <dc:creator><![CDATA[Alex Reinhart]]></dc:creator>
            <pubDate>Fri, 23 Feb 2024 20:53:41 GMT</pubDate>
            <atom:updated>2024-02-28T15:00:18.801Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Hydrogen is one of the most abundant elements in the universe and has long been touted as the solution for the clean energy transition because of its high energy density, low weight, and abundance.</em></p><p><em>The problem is, we’ve been saying the same thing since the 1970s.</em></p><p><em>The reality is that energy transition is a really challenging problem, but from government intervention to new science, there’s reason to believe that the time for hydrogen could be now.</em></p><h3>Understanding the Hydrogen Markets</h3><p>The market size for hydrogen was 170 billion in 2023, even with the extremely limited Hydrogen energy use cases that exist today. As shown in the chart below the demand for Hydrogen is robust with the two largest use cases being petroleum refinery and ammonia production. Ammonia is the key ingredient in fertilizers and is the bedrock of feeding the ever growing population. The transportation/power generation segments of the hydrogen economy barely exist today, but are projected to grow as we identify new applications.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1002/0*BVtM-gaByRVFYZrI" /></figure><p>Most people don’t realize how large the existing hydrogen industry is already. Everyone who is aware of climate change realizes how much CO2 airline travel contributes; but very few are aware that <a href="https://h2sciencecoalition.com/">H2 production is equal in CO2 emission</a>s to the entire airline industry.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8kxt_CzE3YzmtGuP_K_HGQ.png" /></figure><h4><strong>The Hydrogen Color Code</strong></h4><p>The tendency is to assume that hydrogen is synonymous with clean because it’s naturally occurring, but that is far from the case. Hydrogen gas CAN form naturally, deep in the earth, and it CAN be harvested using similar methods to petroleum. However most of the hydrogen we use today is instead produced by removing hydrogen atoms from other compounds, and typically by using energy that comes from gas or coal in the process.</p><h4><strong>There are 3 categories:</strong></h4><p><strong>Thermochemical</strong>: Burn x to release hydrogen ( x = natural gas, coal, biomass). The most popular method today is Natural gas reforming, also called steam methane reforming (SMR) . But there are also other reactive metals that can be used to produce hydrogen, such as aluminum.</p><p><strong>Electrolytic</strong>: Uses electricity to separate water into oxygen and hydrogen</p><p><strong>Biological</strong>: Uses bacteria, algae or micro-organisms to produce hydrogen through biological reactions.</p><p>Today a vast majority of hydrogen is produced by SMR, burning natural gas. Which makes hydrogen a dirty energy source. The challenge is that even for electrolysis produced Hydrogen, if the energy used in the electrolysis is dirty then your hydrogen isn’t clean either. This can make understanding the emissions within the hydrogen market quite challenging.</p><p>There are some novel lesser used techniques to create hydrogen from reactive metals (aluminum) and water.</p><p>To distinguish the possible origins of hydrogen there is a broadly used color code. Green hydrogen is clean through and through. Blue hydrogen is dirty but with carbon offsets purchased to equate the emission. Grey hydrogen is dirty. For the record I bet someone paid someone to make sure it was gray instead of red, to me grey suggests a medium level of concern.</p><p>Green hydrogen is the economy that the US government is trying to create, however we have not yet achieved cost competitive-ness. In<a href="https://about.bnef.com/blog/green-hydrogen-to-undercut-gray-sibling-by-end-of-decade/"> 2023</a> the cost of green hydrogen remains about 3x gray.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/874/0*9lNJ_u4WgsAWi8J4" /></figure><h3><strong>Government Activity</strong></h3><p>While I would not count on the government to be the leader of our national energy transition, there have been significant funds dedicated to helping prop up the green-hydrogen economy as it matures. Hydrogen is considered to be one of the keys to US energy independence. Hydrogen can be created from water, anywhere, making it one of very few energy sources not dependent on rare earth metals or fossil fuels, and the global supply chains connected to them. Even batteries are highly internationally dependent.</p><p>Prior to Biden’s presidency government pressures came mainly from state by state legislation that enforced taxation, and demanded climate action plans. Starting in 2022 pressure is now coming from the national government through subsidies and incentives. The two key bills have been the infrastructure and jobs act (550 billion to climate) and the inflation reduction act (369 billion to climate) passed in 2023. The chart below shows the allocation of funds from each bill.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/757/0*W9SINvizzWhM08HF" /><figcaption>Jobs Act VS Inflation Reduction Act</figcaption></figure><p>These bills and subsidies help tip the scale in favor of clean energy sources temporarily, ideally helping them achieve cost competitive-ness through economies of scale. The challenge for any alternative energy source we develop is that it has to compete with one of the most optimized infrastructures in the word — oil. Nobody is going to adopt anything without some version of this equation being true:</p><blockquote><strong>Cost Energy(clean) &lt;= Cost Energy(dirty)</strong></blockquote><p>But with these government incentives the equation now looks like this:</p><blockquote><strong>Cost Energy(clean) — Incentives &lt;= Cost Energy(dirty) + Emission Taxes</strong></blockquote><p>These stimulus packages are broad across all of clean infrastructure, but there have been some tailwinds supporting the hydrogen economy more specifically. The Department of Energy released a ‘funding opportunity announcement’ in December of 2023, for up to<strong> $59 million to accelerate the research, development, demonstration, and deployment of affordable clean-hydrogen technologies</strong>. This is significant, but it is just the first deployment of funds of the<strong> $9.5 billion that the US has earmarked for clean hydrogen</strong>.</p><p>The US government also published the<a href="https://www.hydrogen.energy.gov/docs/hydrogenprogramlibraries/pdfs/us-national-clean-hydrogen-strategy-roadmap.pdf"> US Clean Hydrogen Strategy and Roadmap</a>, in June of 2023 which sets the stage for the use cases and infrastructure that they are prioritizing.</p><p><em>[Intended clean hydrogen] markets include the industrial sector (e.g., chemicals, steel and refining), heavy-duty transportation, and long-duration energy storage to enable a clean grid. Additional longer-term opportunities include the potential for exporting clean hydrogen or hydrogen carriers and enabling energy security for our allies.</em></p><p><em>- Clean Hydrogen Roadmap</em></p><p>To achieve success in any of these markets two things are required: <strong>green hydrogen</strong> must be made cost competitive with gray hydrogen, and<strong> green hydrogen energy </strong>must be cost competitive with and as performant as the alternative in any given use case (gas/diesel/batteries/etc).</p><p><em>Sidenote —</em></p><p><em>As with anything in government, it’s totally possible that these bills could be repealed. However because of the way these bills are structured that’s not super likely. These bills are creating a lot of economic opportunity — mostly economic rewards rather than punishments. Basically its good for business and good for the environment and thus it’s unlikely that republicans could repeal it ( both because they don’t have enough seats and because they likely wouldn’t have internal alignment) Here’s a great post about it — </em><a href="https://legal-planet.org/2023/04/17/why-the-inflation-reduction-act-cant-be-repealed/"><em>https://legal-planet.org/2023/04/17/why-the-inflation-reduction-act-cant-be-repealed/</em></a></p><h4><strong>Hydrogen Globally</strong></h4><p>Similar to the US’s green hydrogen plan, the EU, China, Australia, Japan and India have all created green hydrogen plans and dedicated substantial funding. China is the furthest along with a much more aggressive timeframe targeting to bring 50,000 hydrogen fuel-cell vehicles on the road<a href="https://www.iea.org/policies/16977-hydrogen-industry-development-plan-2021-2035"> by 2025</a> and to build a number of hydrogen refueling stations.</p><h3><strong>The Hydrogen Opportunity</strong></h3><p>The consumable marketplace for hydrogen today is over $100B each year with pretty stable growth, but the novel market is the use of hydrogen as an energy storage and fuel mechanism.</p><p>The opportunity for value creation in the hydrogen economy is twofold — create cheap green hydrogen and compete in the established commodities market, OR develop the use cases of hydrogen as a fuel source and play in the new and unproven products/energy market.</p><p>In this article we will take a look at both segments and who the players are.</p><h4><strong>Green Hydrogen as a Consumable</strong></h4><p>The name of the game in this category is creating inexpensive green hydrogen.</p><p>In 2023 the first green hydrogen unicorn raised its series C,<a href="https://eh2.com/"> Electric Hydrogen, a Massachusetts</a> based company that developed electrolysis technology that allows them to produce more hydrogen at a lower price. The company’s business model is to sell cheaper electrolyzers to existing hydrogen producers so that they can scale up their green hydrogen production, rather than creating and running green hydrogen plants themselves. It’s similar to how solar panel manufacturers sell panels to clean energy developers. The below image shows Electric Hydrogens prototype electrolysis solution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SmsiB-b9JqX1nfPQ" /><figcaption>Electric Hydrogen’s electrolysis system</figcaption></figure><p>Another green hydrogen startup targeting the same market with a similar business model is<a href="https://advanced-ionics.com/about/"> Advanced Ionics</a>. Who most recently raised a<a href="https://techcrunch.com/2023/08/24/advanced-ionics-hydrogen-series-a/"> series A of 12.5 million</a>. Advanced Ionics’ novel approach is to tap into existing processes and waste heat sources at industrial facilities and create green hydrogen plants in partnership with big industry.</p><p><strong>Electrolysis is not the only strategy for green hydrogen</strong> production at scale,<a href="https://solugen.com/"> Solugen</a> is a synthetic biology startup that develops enzymes and uses natural biological processes to create chemicals. They play not only in the hydrogen market but in all chemical manufacturing. Since their founding in 2016 they’ve raised<a href="https://www.crunchbase.com/organization/solugen"> $642.2M</a>.</p><p>Innovations in the production efficiency space are likely to come from the lab. I believe that there are also some interesting opportunities as the hydrogen market evolves. These startups are all addressing the existing hydrogen market — the consumable market, as new use cases for hydrogen develop there will be different types of production needs which leaves space for new entrants.</p><h4><strong>Transportation/Fuel</strong></h4><p>The concept of using hydrogen as a fuel source for transportation is hotly debated, but never-the-less it’s moving forward.</p><p>Hydrogen fuel can come in two forms, gaseous and liquid. Both are used as fuel sources but have very different requirements. Gaseous hydrogen must be stored at extremely high pressures, while liquid hydrogen requires extremely low temperatures, specifically cryogenic conditions.</p><p>Both forms of hydrogen fuel have <strong>gravimetric energy density</strong> higher than gasoline (density for a given weight), the challenge is rather the <strong>volumetric energy density</strong> (density for a given volume).</p><p><strong>Hydrogen is nearly weightless, but it takes up more space.</strong></p><p>Liquid hydrogen is created by cooling gaseous hydrogen to extremely low temperatures. This extreme cooling condenses the hydrogen gas into a liquid form, significantly increasing its volumetric energy density. Liquid hydrogen requires particularly careful handling due to its cryogenic nature. Both gaseous and liquid hydrogen are highly flammable substances. The explosive nature of hydrogen makes it a candidate for both fuel cell and internal combustion engine processes.</p><p>Below is a chart comparing the energy density across the most common energy sources. For reference, gasoline has a density of 730 kg/m3 and a gravimetric density of 46 MJ/kg. <strong>This means that gaseous Hydrogen takes up 8x more space than gasoline for the same amount of energy, and liquid Hydrogen takes up 4x more space.</strong></p><p>In comparing these fuels it becomes a trade off between, space, weight, and range.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VnyoksZHObjI4HnS" /></figure><h4><strong>How does hydrogen fuel work?</strong></h4><p>Hydrogen is a combustible, which means that it can be burned as a fuel source to create energy. However, it does not detonate the way that petroleum fuels do, and hydrogen fires end in a matter of seconds, compared to many minutes for heavier-than-air petroleum fuels.</p><p>Here’s a video of a comparison between a hydrogen vehicle explosion versus a gas explosion:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FOA8dNFiVaF0%3Fstart%3D18%26feature%3Doembed%26start%3D18&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DOA8dNFiVaF0&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FOA8dNFiVaF0%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/fd7ffda498b87ea15ad5cc806d9ea996/href">https://medium.com/media/fd7ffda498b87ea15ad5cc806d9ea996/href</a></iframe><p>Liquid hydrogen has been used as a rocket fuel since the 1950s. It is an excellent fuel source for airborne applications because of their sensitivity to weight. Along with its lightweight nature, hydrogen is also quite explosive (technical term being <em>high specific impulse</em>) which just means per kg of fuel it provides more force and burns for longer. Finally the cherry on top is that it burns clean, producing only water vapor as the byproduct.</p><h4><strong>Why didn’t planes adopt liquid hydrogen as a fuel source when rockets did?</strong></h4><p>The reason that planes didn’t adopt hydrogen when rockets did is due to costs. To maintain liquid hydrogen you have to keep it in cryogenic conditions which requires a vacuum insulated storage system. Ironically even though rockets have historically been single use, they have not been budget constrained, and so liquid hydrogen made sense. The other component in favor of hydrogen is that rockets hold few passengers and are extremely large. Liquid hydrogen systems take up significant space on rockets, for commercial airlines on the other hand — space, is seats, seats are money.</p><h3><strong>Hydrogen Powered Vehicles</strong></h3><p>Toyota just came out with the world’s first<a href="https://www.youtube.com/watch?v=DGL5g91KwLA"> Liquid Hydrogen combustion engine</a> that raced against traditional internal combustion engines (ICEs) on August 11 2023. Below is a model of the Hydrogen ICE used in the race. The reason for the development of this vehicle is to prove that there could be a viable clean fuel source for race cars that has the same explosive-ness and sound that people love so much today.</p><p>Realistically ICEs have much higher loss and so they are not likely to be the way we see hydrogen fuel processed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*KlbgVgPGPZJeUO3h" /><figcaption>Toyotas liquid hydrogen race car engine</figcaption></figure><p>For ground transportation use cases hydrogen is not typically combusted, but rather processed by a fuel cell. Fuel cells rely on the recombination of hydrogen and oxygen to generate an electrical current, with the product being water, electricity and heat. There are many examples of vehicles driven by gaseous hydrogen. Everything from commercially available Fuel Cell EV’s to research projects, MIT came out with a<a href="https://news.mit.edu/2024/future-motorcycles-could-be-hydrogen-0110"> hydrogen motorcycle</a> that used this technology.</p><p>The total number of hydrogen-powered passenger vehicles on the road amounts to approximately<a href="https://gmauthority.com/blog/2023/02/56000-hydrogen-powered-vehicles-on-the-road-so-far/"> 56,000 units in 2023</a>. Out of these, 30,000 hydrogen fuel-cell vehicles (FCEVs) were sold in 2021 and 2022, about 54 percent, which suggests accelerating demand for these vehicles.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mO6BC0-UXQvslQjJ" /></figure><p>In terms of the availability of Hydrogen, it’s not very available. This is one of the biggest challenges facing its adoption. <strong>For drivers there are about</strong><a href="https://www.eia.gov/energyexplained/hydrogen/use-of-hydrogen.php"><strong> 56 hydrogen-vehicle-fueling</strong></a><strong> stations in the US, all of which are in California. So good luck.</strong></p><p>There are significant infrastructural challenges with overcoming the distribution issues, gaseous hydrogen requires such high pressures none of the existing pipelines are fit to the task.</p><p>That hasn’t stopped startups from bringing hydrogen vehicles into reality.</p><p>The most well known startup is<a href="https://nikolamotor.com/tre-fcev/"> Nikola Motors</a> which created a hydrogen electric truck with a range of up to 500 miles, and a refuel time of 20 minutes. Nikola has raised $3.3B to date. However Nikola Motors was the subject of a <a href="https://hindenburgresearch.com/nikola/">Hindenburg report</a> in 2020 that exposed a number of instances of fraud and the CEO was ousted.</p><p><em>Their tag-line </em><strong><em>‘All science no fiction’</em></strong><em> maybe should’ve been ‘</em><strong><em>All science some fiction’.</em></strong></p><p>They are now a public company and launched their hydro-electric semi mid last year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UTaI1TOxzg3gxam7" /><figcaption>Nikola motors hydro-electric trucks</figcaption></figure><p>Another startup <a href="https://www.tevva.com/en">Tevva</a> is tackling the same problem with a hydrogen/electric trucking vehicle that has raised $173.6M to date. These larger vehicles are the perfect candidates for a hydrogen upgrade because not only are they ‘energy guzzlers’ they are also not particularly space constrained in comparison to a personal vehicle.</p><p>Toyota is the big player in the space and have hydro-electric semi trucks driving routes throughout California. This begs a question of the opportunity to expand refueling for these fleets that are clearly doing well.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BplPRjNi_2GexZ9oBFyiGw.png" /></figure><h3><strong>Hydrogen Aviation</strong></h3><p>Startup<a href="https://www.beyond-aero.com/prototypes"> Beyond Aero</a> developed the first hydro-electric aircraft which uses a combination of battery power and gaseous hydrogen allowing for a 7x range.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ISHC76OIyvoj1wqf" /><figcaption>Beyond aero hydro-electric plane</figcaption></figure><p>Startup ZeroAvia built a hydro-electric plane targeting 500 mile short-haul trips, at half of today’s cost. Raised $257M to date. <strong>Startup</strong><a href="https://hydrogen.aero/newsroom/"><strong> Universal Hydrogen</strong></a><strong> is the first to create a hydro-electric passenger airplane that uses liquid hydrogen. Their first prototype plane has set numerous records.</strong> The company has developed special tanks to hold liquid hydrogen (UH calls them modules), which can then be trucked to the airport.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/503/1*kaTADgvtavT-bO4tzK8mTw.png" /><figcaption>Universal Hydrogen Commercial Airliner</figcaption></figure><p>The modules are designed to slot straight into the aircraft, where they can be plugged into the propulsion system. No need for pipes, hoses and pumps.</p><p>The modules are insulated and can keep the hydrogen in its liquid form for four days. Two modules would hold 360kg of hydrogen and would be able to fly an aircraft 500 miles, plus an extra 45 minutes of flight time in reserve. An impressive concept but lacking the infrastructure to run at scale.</p><p>All in all Universal Hydrogen is the most likely to bring hydro-electric planes to commercial flight in the next 5 years.</p><p>In September of 2023 the use of pure liquid hydrogen aircraft (no battery) was demonstrated in the<a href="https://www.h2fly.de/2023/09/07/h2fly-and-partners-complete-worlds-first-piloted-flight-of-liquid-hydrogen-powered-electric-aircraft/"> world’s first piloted flight of a liquid hydrogen-powered electric aircraft</a>, marking a critical step towards emissions-free, medium- and long-haul commercial flights.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VRZoVpS3LnWIVsen" /><figcaption>first fully hydrogen plane</figcaption></figure><p>The Drone market has also begun to join the hydrogen train starting in late 2023.<a href="https://www.nrel.gov/news/program/2023/honeywell-aerospace-and-nrel-partner-to-scale-novel-hydrogen-fuel-storage-solution-for-drones.html"> Honeywell Aerospace</a> and<a href="https://www.flyingmag.com/hevendrones-unveils-2-hydrogen-powered-drones/"> Heaven Drones</a>, 2 top drone manufacturers, unveiled Hydrogen powered drones for long-range flights for both defense and non-defense applications. Newer companies in the space such as Anduril and Mach Industries are working to bring Hydrogen into the high speed drone and projectile (hydrogen propelled missiles) space.</p><p><strong>Sentiment is undeniably high.</strong></p><p>It is worth noting that the alternative (clean fuel) to these battery or hydrogen solutions is synthetic fuels, which plug into current infrastructure much easier. <strong>This still correlates well with the hydrogen economy because hydrogen is a component in all synthetic fuels.</strong></p><p><em>Personal Take —</em></p><p><em>We have all heard the stories of the Boeing 737 issues and the Alaska Airlines doors with missing bolts. No one is thrilled with the aviation companies of today. They’ve locked themselves into a stalemate of little innovation. Not dissimilar to the car companies of years past. Admittedly aviation is a tougher nut to crack, but I have no doubt that the founders lurking in labs and basements (maybe spaceX offices) can overcome these hurdles.</em></p><p><em>I think there’s opportunity for a new airline manufacturer for commercial flight that looks very different than the ones we have today.</em></p><h3><strong>Energy Storage</strong></h3><p>Hydrogen is a good medium in which to store energy because gaseous hydrogen is stable and requires no energy to maintain if stored in the right container. Batteries on the other hand lose energy over time, approximately 3% per month. This means that your average Li-Ion battery will lose all of its charge if left untouched for 3 years. The challenge is that one must consume energy to create hydrogen meaning that you lose some energy in the ‘charging’ process.</p><p>Most companies in the the hydrogen energy storage space are also hydrogen producers. This makes sense as to store energy as hydrogen you need to create hydrogen. Its unlikely that we will see a standalone version, but it is very possible that this could be a real revenue generating component of a green hydrogen business.</p><h3><strong>Natural Gas Alternative</strong></h3><p>Natural gas today makes up about 33% of US energy consumption. Many power plants operate on natural gas and there is significant pipeline infrastructure to transport this gas around the country. Hydrogen has comparable properties and has been considered as an alternative. Hydrogen however, would require completely different infrastructure — different engines, new pressure rated pipelines, and once again you would need 3x the volume of hydrogen to output the same amount of energy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*PK3Cocy67NR2-Xyf" /></figure><p>The natural gas opportunity is certainly possible, but its a challenge that would need to be attacked one small region at a time. Not to mention it would be a painful and challenging business to build.</p><h3><strong>Overview &amp; Opinions</strong></h3><p>The hydrogen economy is in flux, and that leads to opportunity. That opportunity falls in two categories — creation of green hydrogen as a consumable, and development of products with hydrogen as an energy source. The money the government has dedicated to the development of hydrogen hubs within the US opens the doors for green hydrogen processes that weren’t previously economically feasible.</p><p>I think green energy creation is an opportunity if you can show order of magnitude advances over state of the art electrolyzer technologies. Otherwise it’s just too hard to compete in that space. It’s not new and it’s also competing in the energy market which is commoditized.</p><p>The largest opportunity stems from the creation of a new market, hydrogen fuel and hydrogen transportation. Should these endeavors be successful there will also be many infrastructure changes required to support the new economies.</p><p><strong>The questions I’m thinking about are —</strong></p><p><em>What electro-hydrogen vehicles/systems can exist?</em></p><p><em>How will we equip vehicles with hydrogen?</em></p><p><em>How might we equip airports with hydrogen?</em></p><p><em>Where will we need hydrogen and how can that inform the types of refuling systesm we may need?</em></p><p><em>What kinds of long term energy storage facilities will be built with hydrogen at the core?</em></p><p>Got ideas? Let me know.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3de5c249a044" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What you should know about GPUs — The Paradigm Shift of Compute]]></title>
            <link>https://medium.com/@alexmreinhart/what-you-should-know-about-gpus-the-paradigm-shift-of-compute-1343950d2560?source=rss-949b7a03ae4f------2</link>
            <guid isPermaLink="false">https://medium.com/p/1343950d2560</guid>
            <category><![CDATA[photonics]]></category>
            <category><![CDATA[quantum-computing]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[gpu]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Alex Reinhart]]></dc:creator>
            <pubDate>Tue, 26 Dec 2023 22:16:54 GMT</pubDate>
            <atom:updated>2024-01-29T01:25:38.551Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>What you should know about GPUs — Part 3, The Paradigm Shift of Compute</strong></h3><p><em>Part 1 — </em><a href="https://medium.com/@alexmreinhart/what-you-should-know-about-gpus-part-1-history-industry-9f27ab06b50a"><em>History &amp; Industry</em></a></p><p><em>Part 2 — </em><a href="https://medium.com/@alexmreinhart/what-you-should-know-about-gpus-part-2-the-rise-of-ai-d426b1272446"><em>The Rise of AI</em></a></p><p><em>Part 3 — </em><a href="https://medium.com/@alexmreinhart/what-you-should-know-about-gpus-the-paradigm-shift-of-compute-1343950d2560"><em>The Paradigm Shift of Compute</em></a></p><h3><strong>Innovation Pushing to Change the Paradigm of Compute</strong></h3><h3><strong>Quantum Computing</strong></h3><p>To understand quantum computing you have to think back to chemistry classes when you were studying the locations of elections in atoms and looking at diagrams like this —</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/569/1*_LmiOMlL2bhPH0fet1kNsQ.png" /></figure><p><strong>Superposition</strong> — To understand quantum computing we will analogize a transistor (which is either a 1 or 0 in traditional computing) to the yellow orbital in the top left. The location of an electron (referred to as a qubit) at the time of measurement is either a 1 or 0, based on what side of the x axis it is found. The concept of superposition in electrons means that they are probabilistically in both states at once until measured (Schrödinger’s cat style). Quantum algorithms can exploit this superposition to explore many possibilities simultaneously, leading to parallelism. What that means, idk, but I’m just going to leave that there …</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y1XPa5YGBKo0S4vZEH5V2Q.png" /><figcaption><a href="https://scienceexchange.caltech.edu/topics/quantum-science-explained/entanglement">Caltech entanglement 101</a></figcaption></figure><p><strong>Entanglement</strong> — Quantum computers can leverage entanglement, a phenomenon where the state of one qubit becomes linked with the state of another, even if they are physically separated. Changes to one qubit will instantaneously affect the other, no matter the distance between them [aka 0 latency information transfer].</p><p>One of the the challenges is quantum gates (the quantum version of logic gates aka AND, OR, NOT) are error prone because the control of electrons is not a perfected science.</p><p><strong>State of Tech </strong>— <a href="https://ionq.com/quantum-systems/compare">You can use quantum computers</a> if you would like to spend approximately $<a href="https://www.google.com/url?sa=t&amp;rct=j&amp;q=&amp;esrc=s&amp;source=web&amp;cd=&amp;cad=rja&amp;uact=8&amp;ved=2ahUKEwisocGSldaCAxU7mGoFHXxMCFMQFnoECA8QAw&amp;url=https://quantumzeitgeist.com/how-much-do-quantum-computers-cost/#:~:text=Several%2520companies%2520offer%2520cloud%252Dbased,hour%2520of%2520quantum%2520processing%2520time.&amp;usg=AOvVaw0s7l6qIrdRbNlX1wi-k_Qy&amp;opi=89978449">1–5k/hour</a>. Long term, it is unlikely that quantum will scale to the accessibility and generalizability of a GPU in the next 20 years given the cost to manufacture. Quantum computers will compete on the supercomputer market, not on the ‘everyday inference’ or ‘everyday training’ markets. There is a question to be asked if they may be used for training of the largest of models and <a href="https://www.datanami.com/2023/07/24/quantum-computing-and-ai-a-leap-forward-or-a-distant-dream/">this is a possibility</a><strong>.</strong> To unlock this technology there will need to be a CUDA equivalent for quantum, not out of reach but definitely a pain.</p><p>Few startups will have the funds to compete in the quantum computer space. On top of the cost there are also challenging incumbent dynamics. Microsoft owns 49% of OpenAI and their own quantum computer. Neither of these has stopped the <a href="https://www.paymentscardsandmobile.com/quantum-computing-practical-implications-in-payments-and-banking/">money flowing into the field</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WFBUEb_mT9HDcGpRaDikJw.png" /></figure><p><strong>Startups </strong>— A recent startup <a href="https://www.extropic.ai/accelerate">Extropic</a> raised 17M to create an AI processor using what appears to be photonic technology.</p><h3><strong>Photonic Computing</strong></h3><p><strong>Concept</strong> — Photonics is a technology that replaces electrons with photons in chips. Researchers are attempting to develop photonic versions of all sorts of components, photonic memory, photonic data transfer, photonic compute (as we are discussing here).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n4fOF03znwzP-lnamO6vAA.png" /><figcaption><a href="https://techhq.com/2023/05/what-is-optical-computing-explained/">Photonics explained</a></figcaption></figure><p><strong>How does photonics work?</strong> Traditional compute uses transistors (0s &amp; 1s). Photonics uses wave propagation to do calculations. The wave property means that photonics can be analogue or digital. Light waves have additive and subtractive properties when interacting with other light waves, 0s and 1s are represented using light or absence of light. <strong>Measuring the outputted light can allow you do calculations at the speed of light</strong> without the latency induced by transistor switching. Check out <a href="https://techhq.com/2023/05/what-is-optical-computing-explained/">this video</a> for an in-depth explanation.</p><p>When using electrons, the speed of data transfer is limited by how fast electrons move. Since the speed of light is more than 100 times faster, moving data using photons can significantly accelerate computing and data transfer.</p><p>We all know the saying- “<em>you’re only as strong as your weakest link”</em>, here it’s more like “<em>you’re only as fast as your slowest component”</em>. Photon-izing the GPU, runs the risk of being bottlenecked by other components that can’t move at the speed of light. The key here is to design processes that can run phonically with out needing access to other systems, e.g memory. To experience the true lift of this technology, we need photonic systems not just photonic components.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/340/1*QjZqJflK7F-R1Jhf1sLUjg.png" /><figcaption><a href="https://www.britannica.com/science/fiber-optics">Fiber Optics 101</a></figcaption></figure><p><strong>Startups</strong> — <a href="https://tracxn.com/d/companies/lightmatter/__UzjACTIII619AffSL1FYPgASKG7oNo0LdOcRvQnFRl8">Lightmatter</a> was founded in 2021 originally targeting photonic processors. The company eventually pivoted to work on interconnect, aka connections intra chip. If I were to hypothesize this shift, I’d imagine that photonic interconnect is a low hanging fruit problem that likely can be sold to many chips on the market, data throughput is a universal issue. It is still to be seen what comes of light matter’s photonic compute technology, but it seems to be on the back burner.</p><p><a href="https://www.analogphotonics.com/products/">Analog photonics</a> was working to make photonic sensors starting with lidar, combining both the memory advantages of analogue (avoiding intermediary binary conversions and capturing more information in each calculation), and the speed of photonics.</p><p><strong>State of Tech</strong> — Photonic sensors and fiber optics have proven themselves to be viable and useful technologies. Photonic compute &amp; memory companies need to prove they can create reliable functional products, then they need to create an economy of scale to make the costs viable. If they can execute on then above they have the potential to blow silicon GPUs out of the water with 100,000x better processing power. Certainly worth paying attention to.</p><h3><strong>Opinions</strong></h3><h4><strong>Innovations in traditional silicon are not going to move the needle</strong></h4><p>A paradigm shift is coming in compute and I think its going to be analogue, photonics, quantum or some combination of the above. The 3 big AI chip startups are poised to compete with NVIDIA over the next 5–10 years, but we are reaching the true limits of current silicon technology and the next big step is going to be fundamentally different, analogue, photonic, or quantum. I lean towards the first two.</p><h4><strong>I hope we see some exits soon</strong></h4><p>Significant amounts of deep tech venture money is locked up in the big 3 AI chip startups. These investments were made in around 2016 or earlier, this means some of this money is 7 years, deep. The real game changers in this field are going to come from the more speculative research driven spaces. For the sake of the advancement of this space, I hope we see some good outcomes and money being re-allocated into the next generation.</p><p>If you are interested in or working on AI chips, Photonics, quantum or the like, please reach out.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1343950d2560" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What you should know about GPUs Part 2 — The Rise of AI]]></title>
            <link>https://medium.com/@alexmreinhart/what-you-should-know-about-gpus-part-2-the-rise-of-ai-d426b1272446?source=rss-949b7a03ae4f------2</link>
            <guid isPermaLink="false">https://medium.com/p/d426b1272446</guid>
            <category><![CDATA[semiconductors]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[gpu]]></category>
            <category><![CDATA[innovation]]></category>
            <dc:creator><![CDATA[Alex Reinhart]]></dc:creator>
            <pubDate>Tue, 26 Dec 2023 21:56:55 GMT</pubDate>
            <atom:updated>2024-01-28T22:52:51.052Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>What you should know about GPUs Part 2 — The Rise of AI</strong></h3><p><em>Part 1 — </em><a href="https://medium.com/@alexmreinhart/what-you-should-know-about-gpus-part-1-history-industry-9f27ab06b50a"><em>History &amp; Industry</em></a></p><p><em>Part 2 — </em><a href="https://medium.com/@alexmreinhart/what-you-should-know-about-gpus-part-2-the-rise-of-ai-d426b1272446"><em>The Rise of AI</em></a></p><p><em>Part 3 — </em><a href="https://medium.com/@alexmreinhart/what-you-should-know-about-gpus-the-paradigm-shift-of-compute-1343950d2560"><em>The Paradigm Shift of Compute</em></a></p><h3><strong>The Rise of AI and GPUs</strong></h3><p>In the past year, AI has been in the spotlight like never before. It’s easy for us to forget that this wow moment came with decades of hard fought development. These models have been around since the early 2000s.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XfuBRF_-PBs7mghe1qXmRQ.png" /></figure><p><a href="https://ourworldindata.org/brief-history-of-ai">Source</a></p><p>Understanding that the growth in AI has been hard fought, can also help us understand that the hardware for AI has been evolving for some time as well. GPUs continue to be the hardware backbone of AI systems, but this may be changing with the advent of specialized AI chips.</p><h3><strong>Specialized AI Chips</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/636/1*GhB500LBRxbGW-rQTx6Ejw.png" /></figure><p><a href="https://www.youtube.com/watch?v=A7SpNWzAh8U">AMD vs NVIDIA</a></p><p>In the last 5 years the ‘AI chip’ design frontier has grown crowded with Meta, Alphabet, and Microsoft joining the fabless chip designer club. Needless to say this has been very good for TSMC.</p><h4><strong>Are ‘AI Chips’ fundamentally different from GPUs?</strong></h4><p>In short, the answer is no — at least not as of today.</p><p>AI chips are really a sub category of GPUs in which the ratios of components (processing cores &amp; memory) are tuned to the application. ML tasks require more memory and specific types of cores — tensor cores that are optimized for <a href="https://www.math3ma.com/blog/the-tensor-product-demystified">tensor based multiplication</a> (matrix multiplication). Certain hardware components are also modified for the application e.g. ML tasks require less precision so memory and cores are tuned to use 16-bit or 8-bit data rather than the typical 32-bit.</p><p>The NVIDIA H100 is a GPU, it’s just a GPU optimized for AI. It’s a GPU so optimized for AI that it <a href="https://videocardz.com/newz/nvidia-40k-hopper-h100-gpu-has-been-tested-in-games-slower-than-integrated-graphics#:~:text=The%20H100%20is%20not%20a,graphics%20cards%20to%20play%20games.">performs poorly for gaming</a> compared to off the shelf GPUs.</p><h3><strong>Trends</strong></h3><h4><strong>Training and Inference</strong></h4><p>Training an ML model looks like a very different process than inference.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mrP2mS2HV1rSmySo2hveGg.png" /></figure><p><a href="https://blogs.nvidia.com/blog/difference-deep-learning-training-inference-ai/">Traning and Infrence comparison</a></p><p>In training you both calculate forward and propagate information backward to update edge weights. This requires significantly more calculations and more memory. Inference is a forward only process for which there are techniques that help minimize the calculations and memory needed, like sparceification (removing unused weights).</p><p>Large Language Models (LLMs) require huge amounts of compute for training. An AI chip (specialized GPU) is the best solution. AI chips today are tuned for training not for inference.</p><p>Inference is largely run on CPUs rather than GPUs. GPU optimization for inference may make sense for inference at the edge (on device) and/or at very high speeds (for example self driving). The current state of the art for inference is hardware optimized for the type of calculation needed, an ASIC (Application Specific Integrated Circuit). To employ ASICs you need to design a chip to your use case. For example the<strong> TPU (Tensor Processing Unit) created by Google is a </strong><a href="https://www.linkedin.com/pulse/gpus-vs-tpus-comprehensive-comparison-neural-network-workloads-joel/"><strong>ASIC for ML</strong></a>.</p><h4><strong>Digital vs Analogue</strong></h4><p>There has been a long debate on weather compute should proceed in the analogue or digital domain.</p><p>Analog signals are continuous signals, most often used to represent physical measurements. They use a continuous range of values to represent information. To understand this phenomenon think of a temperature sensor, it provides a constant stream of measurement relaying 32 degrees. Digital instead represents data in binary, so 32 would read 100000. Un-intuitively, Analogue was the first compute technology, but it was found to be unreliable for computation due to its sensitivity to noise, for example if you are reading a 50.29, noise can easily distort your results versus there’s virtually no risk in misreading a 1/0.</p><p><a href="https://www.bairesdev.com/blog/why-analog-may-be-the-future-of-computing/">Many advocate for analogue computing</a> to return, it is used frequently in sensors that require continuous streams of data, but hasn’t returned to compute. There are a few startups who think that is the answer to memory bandwidth issues in AI computing because there are less conversions and intermediate numbers be saved to memory.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/540/1*2I0Dp7jDTBAUL3QYSQJStQ.png" /></figure><h3><strong>GPU Startup Landscape</strong></h3><p>There are a few startups that have raised significant amounts of money to challenge NVIDIA in creating the best AI chip. <strong>SambaNova Systems</strong> has raised <a href="https://www.crunchbase.com/organization/sambanova-systems">1.1B dollars </a>since their founding in 2017, with the goal of providing a full stack AI service for LLMs for enterprise. Their novel chip SN40L has memory access innovations that aim to alleviate the memory bottleneck. Their technical advantage is nothing to write home about, but they are providing an easy to use compute service in a time of scarcity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/312/1*jfOWD8aYanDPp9mt9FKt2w.png" /><figcaption><a href="https://www.ai-startups.org/top/hardware/">SambaNova Team</a></figcaption></figure><p><strong>Cerebras</strong> took a different approach in their AI hardware. They are making chips as big as current manufacturing processes allow- wafer sized. In comparison to the NVIDIA 5120 core chip, Cerebras’s has 850,000 cores, to support that many cores Cerebras has also made advancement in I/O and memory access. They have raised <a href="https://www.crunchbase.com/organization/cerebras-systems">715M since their founding in 2016</a>, and today have a live training and inference cloud service as well as hardware available for purchase.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-RIJi1HHOA6eILKwpghkCA.png" /><figcaption><a href="https://www.cerebras.net/product-chip/">Cerebras Chip Size comp</a></figcaption></figure><p><strong>Graphcore</strong> a UK based AI chip manufacturer was founded in 2016 and has racked up <a href="https://www.crunchbase.com/organization/graphcore">682M in funding</a>. Graphcore’s technical edge is a new hardware they call the IPU (inference processing unit). They partnered with TSMC to bring a new 3D manufacturing technology into their IPUs. They tout a <a href="https://www.graphcore.ai/posts/the-wow-factor-graphcore-systems-get-huge-power-and-efficiency-boost">5x faster training speed</a> and 40% faster inference and offer both cloud services and hardware for purchase.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*bOlKTeocYhhDiZjGZBWzDA.png" /><figcaption><a href="https://www.graphcore.ai/products/c600">Graphcore IPU</a></figcaption></figure><h4><strong>The CUDA Moat</strong></h4><p>There are a few commonalities amongst all of these well funded NVIDIA competitors, they all continue to leverage standard silicon wafers, and <strong>they all have a massive accessibility challenge to overcome</strong> with NVIDIA’s proprietary CUDA interface.</p><p><em>‘Graphcore, Cerebras, Tenstorrent, all of these ML accelerator companies, they built chips, the chips were good, the software was terrible’</em> — George Hotz founder of Tiny Corp.</p><p>There was really only one viable strategy, create your own CUDA (your own hardware &lt;&gt; ML library interface). To mimic the competitive moat obtained by NVIDIA, each of these interfaces continues to be developed in house and remain proprietary.</p><p>If we zoom out from the AI hardware race there is one startup taking on the interface challenge. The company <a href="https://www.modular.com/">Modular</a> founded in 2022 has raised 130M to create a universal interface between AI Hardware and software that they call the <a href="https://www.modular.com/max">Max Engine</a>. The Max Engine aims to bridge the powerful moat that NVIDIA has created with CUDA. Modular is also creating a language called mojo, which is a superset of python, making it extremely easy to adopt on existing code. Mojo replaces the need to write in C++ or CUDA and makes it easier to write the code to interface with specialized hardware.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*glYJHjBDc8o2nNuqKVOXyw.png" /><figcaption><a href="https://www.modular.com/max">Modular’s Max Engine</a></figcaption></figure><h3><strong>Events, Acquisitions, etc.</strong></h3><h4><strong>Not all Rainbows and Butterflies</strong></h4><p>Not many AI hardware startups have gone to zero, but 2022 was a challenging time for many of them as NVIDIA came to dominate AI hardware and most venture funding had dried up.</p><p><a href="https://rain.ai/">Rain AI</a> a company trying to create chips that mimic neurons in the brain (I would push back on this statement, <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8777191/#:~:text=Herein,%20we%20turn%20attention%20to,are%20quanta%20of%20electromagnetic%20waves">neuron activity is photonic</a>, but alas) had sourced funding from notable AI figures including Sam Altman, but faced regulatory action after accepting money from a Saudi investor. Goes to show that the AI race is becoming a subject of national security. Or more aptly, an issue of US superiority. Rain AI was saved by a <a href="https://www.datacenterdynamics.com/en/news/openai-promised-to-buy-51m-chips-from-sam-altman-backed-neuromorphic-chip-company-rain-ai/">deal with OpenAI</a> to buy $51M in AI chips.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Kw81W6KLyWtiDxFiZ8M_rg.png" /></figure><p>Rain, pls use some of that 51M to re-do your <a href="https://rain.ai/">website</a>, it looks like a scam</p><p>In a similar vain an AI inference chip company <a href="https://mythic.ai/">Mythic</a>, <a href="https://techcrunch.com/2023/03/09/ai-chip-startup-mythic-rises-from-the-ashes-with-13m-new-ceo/">ran out of funds</a> after raising $70M. After some floundering, they were able to raise another $13M with a new CEO, Dave Flick stepping up. Their aim is to create <strong>an analogue chip</strong>, and target energy efficiency in edge computing. I covered the edge inference space less thoroughly in this article but it’s crowded — <a href="https://techcrunch.com/2022/05/25/ai-chip-startup-sima-ai-bags-another-30m-ahead-of-growth/">SiMa.ai</a>, <a href="https://techcrunch.com/2022/10/25/ai-chip-startup-axelera-lands-27m-in-capital-to-commercialize-its-tech/">Axelera</a>, <a href="https://techcrunch.com/2020/10/20/flex-logix-chip-availability/">Flex Logix</a>, <a href="https://techcrunch.com/2022/12/06/neureality-ai-accelerator-chips-startup-raises-35m/">NeuReality</a>, <a href="https://techcrunch.com/2022/12/14/encharge-ai-emerges-from-stealth-with-21-7m-to-develop-ai-accelerator-hardware/">EnCharge</a>, <a href="https://techcrunch.com/2021/10/12/hailo-raises-136m-ai-chips/">Hailo</a> and <a href="https://techcrunch.com/2021/12/20/ai-chip-kneron-25-million-funding/">Kneron</a>. The fact that they are targeting analog means they’re bringing something new to the table, excited to see how this plays out.</p><p>Even though there haven’t been many disastrous failures so far in the AI hardware space, it has become a particularly challenging space to raise money. You may have noticed that <strong>between SambaNova, Cerebras, and Graphcore, $2.5B of venture money is locked up</strong>, and at least another billion spread over other AI hardware startups that didn’t make that list. There is only a small collection of VCs with the aptitude and appetite for investing in high risk, long horizon hardware products, and most have made their bets for now.</p><h4><strong>Top Acquisitions:</strong></h4><p>Habana Labs was an Israeli company making record breaking AI hardware in 2019. Habana had released the<a href="https://www.forbes.com/sites/moorinsights/2019/12/16/intel-acquires-habana-labs-for-2b/?sh=2b811fe919f9"> Goya chip for inference and the Gaudi chip for training</a>. They were <strong>swiftly acquired by Intel for 2B</strong>. There were particularly strong synergies between Habana Labs and Intel because they were both creating separate technologies for inference and training, which has proven to be a leading philosophy.</p><p>In March 2019 NVIDIA was betting on data center expansion and anticipating models to continue increasing in size, which would begin to require <em>sharding</em> — the breaking apart of models to run on different systems. To maintain speed in cross systems processes they needed high speed networking. <strong>This inspired NVIDIA’s </strong><a href="https://nvidianews.nvidia.com/news/nvidia-to-acquire-mellanox-for-6-9-billion"><strong>aquisition of Mellanox</strong></a><strong> a high performance networking company in March of 2019 for $6.9 Billion</strong>. NVIDIA couldn’t have made a better bet, if you look back at this revenue chart, you can see that data centers have been on a steep upward trajectory ever since.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/1*7I3YVGYr9NuY1dArM-tVCw.png" /><figcaption><a href="https://www.nextplatform.com/2022/08/08/datacenter-props-up-nvidia-as-gaming-sales-collapse/">Nvidia Quarterly Revenue by Category</a></figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d426b1272446" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What you should know about GPUs Part 1 — History & Industry]]></title>
            <link>https://medium.com/@alexmreinhart/what-you-should-know-about-gpus-part-1-history-industry-9f27ab06b50a?source=rss-949b7a03ae4f------2</link>
            <guid isPermaLink="false">https://medium.com/p/9f27ab06b50a</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[gpu]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[semiconductors]]></category>
            <dc:creator><![CDATA[Alex Reinhart]]></dc:creator>
            <pubDate>Tue, 26 Dec 2023 21:52:04 GMT</pubDate>
            <atom:updated>2024-01-02T00:21:28.018Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>What you should know about GPUs Part 1 — History &amp; Industry</strong></h3><p><em>As someone who likes to understand things deeply, this is my research into understanding AI chips, GPUs and the surrounding innovation. It was particularly hard to dig up knowledge because of the abundance of non-substantive content on the subject.</em></p><p><em>Enjoy the information that I found deep in reddit rabbit-holes and TSMC blog posts.</em></p><p><em>Part 1 — </em><a href="https://medium.com/@alexmreinhart/what-you-should-know-about-gpus-part-1-history-industry-9f27ab06b50a"><em>History &amp; Industry</em></a></p><p><em>Part 2 — </em><a href="https://medium.com/@alexmreinhart/what-you-should-know-about-gpus-part-2-the-rise-of-ai-d426b1272446"><em>The Rise of AI</em></a></p><p><em>Part 3 — </em><a href="https://medium.com/@alexmreinhart/what-you-should-know-about-gpus-the-paradigm-shift-of-compute-1343950d2560"><em>The Paradigm Shift of Compute</em></a></p><h3><strong>Origin of the GPU</strong></h3><p>The first Graphics Processing units (GPUs) were created to render visuals in arcade games. As these techniques evolved the GPU became the heart of graphics compute in everything from complex multiplayer gaming systems to the display of your Mac/PC.</p><p>In the early 2000s researchers at Stanford discovered the relationship between graphics processing and other compute intensive processing tasks (AI) and partnered with NVIDIA to <a href="https://www.slashgear.com/nvidia-gpu-neural-network-makes-googles-cat-spotter-look-dumb-18286783">classify cats</a>, running the first ML model (an image classifier) on a GPU. The generalization of the GPU stems from the fact that it excels at SIMD (Single Instruction Multiple Data) operations.</p><h4><strong>Why is the GPU so good at math?</strong></h4><p>A CPU is the brain of a computer, and a GPU is the calculator. CPUs control everything, save to memory and decide when work is sent to the GPU for evaluation. A<strong> core</strong> is the component that does math, the calculator’s processor. A core is full of <strong>transistors</strong> which represent numbers in binary. Cores run <a href="https://www.google.com/search?client=safari&amp;rls=en&amp;q=boolean+operators&amp;ie=UTF-8&amp;oe=UTF-8">boolean operations</a> AND, OR, NOT, XOR (exclusive OR) representing any multiplication, division, addition, or subtraction. Read a cool description of how this works in a calculator <a href="https://www.livescience.com/14087-calculators-calculate.html#:~:text=Transistors%20are%20microscopic%20electrical%20switches,series%20of%201s%20and%200s.">here</a>. The number of cores determines the number of simultaneous processes that can be run.</p><p>The massive parallelization through cores is why GPUs excel. There are 5120 cores on a V100 (Recent NVIDA GPU) while the average CPU has 48 cores. That’s about 100x processing capacity. Though core count is a useful and frequent comparison it is not quite apples to apples. GPU cores have a small eligible instruction set while CPU cores are more versatile. See here for an <a href="https://superuser.com/questions/1516217/gpu-cores-vs-cpu-cores">in-depth explanation.</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/407/1*zNBs5d49rVjOdJPC6_IZZA.png" /></figure><p>The use of GPUs for AI was brought to the mass market via NVIDIA’s release of CUDA (Compute Unified Device Architecture) in 2006. CUDA is the toolkit that connects traditional programing languages (python, java, C++) to the machine code required to leverage a GPUs compute. With this innovation, NIVDIA sparked a revolution of GPU adoption for data centers over the following 10 years.</p><p><a href="https://www.nextplatform.com/2022/08/08/datacenter-props-up-nvidia-as-gaming-sales-collapse/">Nvidia Quarterly Revenue by Category</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*WC1YWFZKD_VJKgwVGHHW9Q.png" /></figure><p><strong>CUDA</strong> turned the <strong>GPU</strong> (Graphics Processing Unit) into the <strong>GP-GPU</strong> (General Purpose Graphics Processing Unit). Want more? See theses articles for more on the <a href="https://medium.com/neuralmagic/a-brief-history-of-gpus-27122d8fd45">GPU</a> &amp; it’s use for <a href="https://medium.com/neuralmagic/why-am-i-using-gpus-for-deep-learning-inference-7d0b10fb7624">AI &amp; ML</a>.</p><h3><strong>General Purpose-GPUs (</strong>Pre-AI)</h3><p>To understand the innovations in GPU technology it is useful to understand dynamics in the semiconductor industry.</p><h4><strong>The Semiconductor Industry</strong></h4><p>Most of us know Intel, Samsung, TSMC and most recently NVIDIA, but it is critical to understand how these how these players interact, to understand the categories of innovation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vXioG_WVaTb_-QbOpBUX_g.png" /></figure><p>The IDMs of today are the monstrous organizations that ruled semiconductor production during the tech proliferation of the 90s. Nowadays cutting edge chips come most often from fabless design companies &amp; OSATs in partnership.</p><p>Fun fact — AMD used to manufacture their own chips but they spun off global foundries in 2009. Former AMD CEO Jerry Sanders once remarked “real men have fabs.” Not anymore Jerry.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*gVSGT0QZXqnnuJyniKEKEg.png" /></figure><p>In GPU innovation the cutting edge of chip design is coming from NVIDIA and AMD. The advancement of fitting more compute into smaller spaces (<a href="https://www.tsmc.com/english/news-events/blog-article-20190814">transistor density</a>) while maintaining manufacturability comes from OSAT innovation, namely TSMC.</p><h4><strong>Design Innovation</strong></h4><p>The innovation in design has come from engineering components and transistors, arranging wires closer and closer together while avoiding crosstalk (where adjacent wires interfere with each other’s signals), and managing heat.</p><p>Design companies also spend time innovating the interface through which software engineers use GP-GPU compute. Examples range from the CUDA toolkit to complete apps designed to leverage the GPU.</p><p><a href="https://www.nvidia.com/en-us/high-performance-computing/simulation-and-modeling/">NVIDIA Apps</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*60dZxA_yLjVx0sA53Q7zBQ.png" /></figure><h4><strong>Chip Hardware Innovation</strong></h4><p>Chip hardware innovation is largely owned by the OSATs who work to push the boundary of what was possible in terms of size and manufacturability. The underling philosophy in semiconductor design has been Moores law — that the number of transistors in an integrated circuit (IC) doubles every two years.</p><p><strong>Interesting excerpt from a </strong><a href="https://www.tsmc.com/english/news-events/blog-article-20190814"><strong>TSMC article</strong></a><strong> (2019) -</strong></p><p><em>Some people believe that Moore’s Law is dead because they believe it is no longer possible to continue to shrink the transistor any further. Just to give you an idea of the scale of the modern transistor, the typical gate is about 20 nanometers long. A water molecule is 0.275 nanometer in diameter! You can now start counting the number of atoms in a transistor. At this scale, many factors limit the fabrication of the transistor.</em></p><p><em>The primary challenge is the control of materials at the atomic level. How do you place individual atoms to create a transistor? How do you do this for billions of transistors found on a modern chip? How do you build these chips that have billions of transistors in a cost effective manner?</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9f27ab06b50a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>