Gimlet Labs: what the $3B multi-silicon inference cloud does

What is Gimlet Labs?

Gimlet Labs is an applied AI research and product company. It publicly launched in October 2025 with, by its own account, eight-figure revenue out of the gate, which is unusual for a company that young.

Its homepage sums up the worldview in one line: "Today's computing systems will undergo a massive transformation to efficiently and scalably serve AI workloads." Everything else follows from that. Instead of accepting that a model runs on whatever GPU you rent, Gimlet builds the layer that spreads the model across many chip types at once.

There are two products behind the pitch:

• Gimlet Cloud is a serverless inference platform aimed at AI agents. You import an existing agentic pipeline, chain multiple models with non-model steps like search, and the platform handles scheduling and orchestration.
• kforge is the lower-level piece. It auto-generates optimized kernels straight from PyTorch across CUDA, ROCm, and Metal backends, using a multi-agent system that explores kernel designs and checks them for correctness, so engineers get speedups without hand-writing kernels.

The problem: agents broke single-chip inference

Here's the part that makes the whole thing click, and it's why I think this is more than a funding-headline story.

A large language model does not do one uniform kind of work. When it answers, it first reads your prompt (the prefill phase), which is compute-bound, it wants raw math throughput. Then it generates tokens one at a time (the decode phase), which is memory-bandwidth-bound, it spends its time shuttling data in and out of memory. And a modern AI agent doesn't stop there: it calls tools, hits APIs, and fetches documents, which is network-bound.

Run all of that on one homogeneous GPU cluster and you're always leaving something idle. During decode, your expensive compute sits half-used. During tool calls, the whole accelerator waits on the network. Gimlet's argument, straight from its Series B post, is that homogeneous hardware alone can't meet the speed and efficiency needs of agentic workloads. The more steps an agent chains, the more that waste compounds.

This is not an abstract concern for me. At eesel I build AI agents that resolve support tickets, and a single resolution is a chain: read the ticket, search the knowledge base, reason over what it found, sometimes call an order-lookup API, then write a reply. Every one of those steps has a different bottleneck. When people ask why agentic products can feel slow, this is a big part of the answer.

How the multi-silicon cloud works

Gimlet's fix is disaggregation. It splits the phases apart and runs each on the silicon that's best for it: prefill on one device set tuned for compute, decode on another tuned for memory bandwidth, and so on. It layers speculative decoding and attention-versus-feed-forward splitting on top, and rebalances dynamically as hardware frees up.

To make that possible, Gimlet partners across the chip industry rather than picking a side. It works with NVIDIA chips alongside AMD, Intel, Arm, Cerebras, and d-Matrix. The software is the product; the silicon is interchangeable. That's the opposite of the usual "buy our accelerator" story, and it's why Arm invested in the round rather than treating Gimlet as a threat.

The payoff Gimlet advertises is 5-10x speedups within the same power envelope, or the same throughput at lower latency. For agentic AI specifically, where one request can need dozens of sequential model calls, shaving latency off every call is where the compounding works in your favor instead of against you.

The funding story, and who's betting

The money tells you who believes this. The Series B was led by Andreessen Horowitz, with a long list of participants including Menlo Ventures (which led the Series A), Arm, M12, Samsung Ventures, Tiger Global, and XTX Markets. Gimlet says it has added billions in contracted revenue since March and has a datacenter pipeline measured in gigawatts.

The trajectory is steep: an $80 million Series A in March 2026 at roughly a $400 million valuation, then a jump to $3 billion six months later. In a market where every lab is chasing cheaper LLM inference, the thesis that efficiency, not just bigger models, is the next battleground is clearly resonating with investors.

The honest caveats

I want to be fair here, because a $3B valuation and a 10x claim invite exactly the kind of hype I try to avoid.

First, the performance numbers are Gimlet's own. When the company published a piece on AI-generated Metal kernels, a PyTorch engineer on Hacker News raised a pointed, reasonable objection:

"I work on PyTorch and there are many things that make me suspicious about these results. My TL;DR is unless we get a zip file of all the kernels with how they're benchmarked results like this are almost impossible to verify."

formalsystem, Hacker News

That's not a takedown, and the Gimlet team engaged with it in the thread. But it's the right frame: benchmark speedups are a claim to verify on your own workload, not a settled result. Second, Gimlet hasn't named its customers publicly, describing them only as "a major model maker" and "an extremely large cloud computing company." And third, this is infrastructure for labs and data centers. If you're building a product, you'll likely never touch Gimlet directly, you'll feel it only as faster, cheaper inference underneath the tools you actually use.

What this means if you build with AI

So where does this leave you if you're not running a frontier lab?

Gimlet is a good reminder of the layer cake. There's the silicon at the bottom, the inference layer that Gimlet is fighting to own in the middle, the models, and then the actual employees you hire on top. Gimlet is infrastructure. It makes the engine run better; it is not the thing that does your job.

That distinction is exactly how I think about eesel. eesel is an AI teammate platform: you hire ready-to-work teammates for specific jobs, and today the roster is an AI helpdesk teammate and an AI blog writer. The helpdesk teammate joins your existing support queue, learns from your past tickets and help center, and resolves conversations inside Zendesk, Freshdesk, or Gorgias. The blog writer does the same for content. Whatever inference cloud sits underneath, the teammate is what carries the job to done.

And because Gimlet's whole world is agentic and programmable, it's worth saying eesel meets that reader where they are too. The eesel CLI drives the same helpdesk teammate from a terminal: a person can run it by hand, scripts can automate it, and coding agents like Claude Code or Cursor can operate it through the CLI and its MCP server, so simulating a rollout against historical tickets or triggering a run becomes one command instead of a dashboard click. If you'd rather see it resolve real tickets than read another spec sheet, you can try eesel free and point it at your own help center in a few minutes.

If you're earlier in the journey and just comparing options, the roundups on the best AI agents and AI ticket classification are a better place to start than a $3B inference cloud.

The bottom line

Gimlet Labs is an interesting company solving a real, un-glamorous problem: the hardware under AI is a bad fit for how agents actually run, and someone had to build the software to fix it. The $3B valuation is a bet that inference efficiency is the next big fight, and given how fast agentic AI is growing, it's not a crazy one.

Just hold the speed claims loosely until they're independently tested, and remember which layer you're actually buying. Gimlet makes the engine faster. The teammate on top is still what gets the work done.