Gimlet Labs Raised $300 Million to Make GPUs and Custom AI Chips Work as One System
A $300 million Series B pushed Gimlet Labs to a $3 billion valuation this month. The pitch: stop buying more GPUs and start using the chips you already have more efficiently.
In this article 6 sections
Quick answer: Gimlet Labs, a San Francisco startup building what it calls a multi-silicon inference cloud, raised $300 million in a Series B led by Andreessen Horowitz, pushing its valuation to $3 billion and its total funding to $392 million. The pitch is narrower than most AI infrastructure raises this year: instead of asking customers to buy more GPUs, Gimlet's software splits a single AI workload across different types of chips at the same time, running GPUs and purpose-built accelerators as one coordinated system.
Most of 2026's AI infrastructure money has gone toward building more of the same thing, faster — more GPUs, more data centers, more megawatts. Gimlet Labs is making a more specific bet: that the chips already deployed across the industry are being used inefficiently, and that the bigger unlock is orchestration, not raw capacity. Protocol Wire covers AI infrastructure funding.
The Round, By the Numbers
Gimlet Labs announced the $300 million Series B on September 4, led by Andreessen Horowitz with participation from Sapphire Ventures, new investors M12 and Arm, and previous backers including Menlo Ventures. The round values the company at $3 billion, up sharply from its prior mark, and brings total funding to $392 million, according to the company's own funding announcement. For a company that's still relatively early in its build-out, that's a lot of capital lined up before the product has had to prove itself at full commercial scale — which is exactly the bet Andreessen Horowitz is making on the technology underneath it, and part of a broader run of capital pouring into the chips actually powering this year's AI boom.
What Multi-Silicon Inference Actually Means
Gimlet's core claim is that no single chip type is optimal for every part of an AI inference workload, so its platform matches each piece of a job to whichever silicon handles it best — GPUs for some stages, purpose-built accelerators for others — and runs the whole thing as one heterogeneous system rather than forcing everything onto uniform hardware. The company says this delivers meaningfully more throughput and interactivity from the same power budget, a claim that matters more than it might sound like it does: power, not chip supply, is increasingly the hard ceiling on how much inference capacity any provider can actually deploy. Coordinating across silicon families instead of adding more of one kind is a different lever on that problem than most of the industry has been pulling this year — including the wave of enterprise coding agents like the ones Factory has been scaling on a tripled valuation of its own this month, all of which ultimately run on inference capacity someone has to build and pay for.
Why Andreessen Horowitz Wanted Back In
The funding will go toward expanding infrastructure capacity for Gimlet's serverless offering, with the company aiming to scale to hundreds of megawatts of managed heterogeneous infrastructure. Gimlet also says it intends to move into custom hardware, developing an inference-optimized server designed without a traditional motherboard, built to be deployed outside conventional data centers entirely — a detail that signals the company sees its future less as a software layer on top of someone else's chips and more as a full-stack infrastructure vendor in its own right.
How Does Gimlet Decide Which Chip Runs What?
It splits each inference job by the kind of work involved. In its own description of the round, Sapphire Ventures said Gimlet sends compute-bound work to GPUs and memory-bandwidth-bound work to inference chips built for fast data access. The same post claims up to 10x gains in inference throughput and interactivity within the same power envelope. Those are investor and company claims, and no independent benchmark has been published alongside them.
The hardware list is broad. Gimlet has partnered with NVIDIA, AMD, Intel, Arm, Cerebras and d-Matrix to support their chips in its architecture, according to Sapphire. Gimlet does not make its own chips today, so its value depends on how well it can schedule across hardware it does not control. Sapphire also cites a figure that frames the opportunity: existing GPU fleets often run at 10 to 50 percent utilization.
Who Is Behind Gimlet Labs?
Gimlet was co-founded by Zain Asgar, its CEO, who is on the Stanford faculty and previously served as a general manager at New Relic. Sapphire lists the other co-founders as Michelle Nguyen, Omid Azizi, Natalie Serrino and James Bartlett. The founding team previously built Pixie, an eBPF-based Kubernetes observability company that New Relic acquired.
Commercial traction is the other part of the story. Pulse 2's coverage of the round says the company emerged from stealth in October 2025 and reported in March 2026 that it had tripled its customer base. It adds that Gimlet has since signed one of the top three frontier AI labs and one of the top three hyperscalers, and that the company says it has secured billions of dollars in contracted revenue for Gimlet Cloud. Pulse 2 also notes that Factory, among the existing investors, took part in the round, and that Gimlet offers its service both through its own managed cloud and as managed deployments inside customer environments.
The Spending Backdrop
The pitch lands against very large spending forecasts. Pulse 2 reports that Gimlet cited Goldman Sachs estimates of roughly $765 billion in AI capital expenditure in 2026 and $7.6 trillion cumulatively from 2026 through 2031. Sapphire, for its part, points to a Gartner projection that AI infrastructure spending will grow from $1.5 trillion in 2026 to $2.8 trillion in 2030.
Andreessen Horowitz managing partner Raghu Raghuram, who sits on Gimlet's board, argued in the announcement, as quoted by Pulse 2, that AI demand is growing exponentially while data centers and silicon cannot keep pace. His conclusion is that the answer is better architecture, not just more infrastructure. That is the thesis Gimlet now has to prove at scale.
What to watch next is whether the contracted revenue turns into live capacity. Neither source says how much of the hundreds of megawatts Gimlet is targeting is operating today, and the custom server remains an intention rather than a shipping product.
That ambition is worth weighing against a market that a working procurement comparison of this month's best-funded infrastructure and AI startups already flags as one where capital intensity, not product-market fit, is the real diligence question. Building hundreds of megawatts of custom infrastructure is a fundamentally different scaling problem than shipping software, and it's the reason Gimlet's $300 million round reads more like the opening chapter of a much larger capital story than a typical Series B. Protocol Wire's interest in rounds like this is the capital intensity underneath them, not only the headline valuation.
