AI CHIPS

Cerebras's 100MW Gimlet Deal Powers a New AI Inference Cloud

The wafer-scale chipmaker will supply roughly 100 megawatts of AI hardware to startup Gimlet Labs, which plans to run frontier models on a multi-silicon cloud promising up to 3,000 tokens per second. It's the latest sign that the AI compute war is moving from training to inference — and from Nvidia's turf to everyone else's.

Published September 28, 2026 · By Daniel Rao

File photo: a close-up of a silicon wafer. Cerebras's Wafer Scale Engine occupies an entire wafer — this generic macro is for illustration only. Photo: Sangitiana Fararano, via Wikimedia Commons (CC BY-SA 3.0).

Cerebras Systems has agreed to supply roughly 100 megawatts worth of its AI chips and hardware to Gimlet Labs, a young cloud-computing startup that plans to run frontier AI models on a new "multi-silicon" inference cloud, the companies announced Monday. The systems — based on Cerebras's CS-4 platform — will be delivered over the next one to two years, with Gimlet targeting cloud availability in 2027.

The deal, first reported by Reuters, is one of the largest third-party hardware commitments Cerebras has disclosed since going public in May, and it lands in the middle of the most aggressive expansion the AI chipmaker has ever attempted. Financial terms were not disclosed. Gimlet will operate and maintain the systems itself, targeting customers in cybersecurity, voice, finance, and other sectors building what the startup calls "AI-native" companies.

For Cerebras, the agreement adds a second major cloud-scale customer win in a year that has already included a $10 billion computing commitment from OpenAI and a hardware purchase by Amazon Web Services. For Gimlet — a startup that emerged from stealth less than a year ago — it is a bet-the-company move: 100 megawatts of wafer-scale silicon, enough compute to power roughly 80,000 American homes if it were electricity for houses, dedicated to a single thesis about how AI inference should be built.

The deal, as announced

Under the agreement, Cerebras will supply its CS-4 systems — the newest generation of its wafer-scale AI computers, introduced in August — to Gimlet over a one-to-two-year delivery window. Gimlet, in turn, will combine the Cerebras hardware with graphics processors from other vendors in what it describes as a multi-silicon system, and sell access to the combined capacity through its Gimlet Cloud.

The performance claims are ambitious. In a joint announcement, Gimlet said the collaboration is designed to deliver up to 3,000 tokens per second at production scale — a figure that, if achieved across real customer workloads, would rank among the fastest inference speeds commercially available. The companies added that their integrated solution is already serving tokens in private deployments, meaning the partnership predates Monday's announcement and has been running in some form with early customers.

"Inference is the new frontier," said Zain Asgar, Gimlet's co-founder and CEO, in the announcement. "With Cerebras wafer-scale technology integrated into our multi-silicon cloud, we can deliver the speed and efficiency breakthroughs that AI-native companies need to build the next generation of real-time applications." Sean Lie, Cerebras's co-founder and chief technology officer, called Gimlet "exactly the kind of partner we built our technology for," adding that the startup's multi-silicon approach "unlocks a new class of performance for demanding AI workloads."

Notably, Gimlet is positioned as a launch partner for the CS-4 — the first customer cohort for hardware that only debuted last month. That suggests Cerebras is using the deal not just for revenue, but as a reference deployment for its newest platform: a live, paying customer running frontier models on CS-4 systems at a scale few others can yet demonstrate.

Who is Gimlet Labs

Gimlet Labs is barely a year old in public, but it arrived with unusual pedigree. The company emerged from stealth in October 2025, founded by Zain Asgar, Michelle Nguyen, Omid Azizi, and Natalie Serrino — the team behind Pixie Labs, the observability startup acquired by New Relic in 2020. Asgar is a former Nvidia GPU architect and Google AI researcher who also serves as an adjunct computer science professor at Stanford.

The startup's funding history explains how it can contemplate a 100-megawatt hardware deal. In early September — less than four weeks ago — Gimlet closed a $300 million Series B round at a $3 billion valuation, led by Menlo Ventures, which had also led its $80 million Series A in March. Total funding stands at roughly $392 million, a war chest that makes the company one of the best-capitalized inference startups in the market.

Gimlet's technical pitch is what it calls multi-silicon inference. Rather than running an entire AI model on one type of chip, its software disaggregates inference into stages and assigns each stage to the hardware best suited for it: GPUs handle the "prefill" phase, where the model processes an incoming prompt, while memory-bandwidth-rich accelerators — like Cerebras's wafer-scale chips — handle the "decode" phase, where the model generates tokens one by one. The company claims the approach delivers up to ten times the throughput and interactivity within the same power envelope, and says it works with silicon from Nvidia, AMD, Intel, Arm, Cerebras, and d-Matrix.

That architecture is a direct response to the economics of modern AI. As models get larger and usage explodes, the cost of serving them — inference — is overtaking the cost of training them as the industry's biggest compute bill. Whoever can serve tokens fastest and cheapest owns the margin layer of the AI economy. Gimlet's bet is that no single chip does every phase of inference well, and that the winners will be software companies clever enough to orchestrate many kinds of silicon at once.

Cerebras's breakout year

For Cerebras, the Gimlet deal caps a twelve-month run that has transformed the company from a well-funded chip startup into a public-market heavyweight. The milestones have come in rapid succession:

In January, the company signed a deal with OpenAI to deliver 750 megawatts of computing capacity through 2028 for $10 billion — a contract so significant that it reportedly includes a temporary prohibition on Cerebras selling to Anthropic. In March, Amazon Web Services agreed to purchase CS-3 systems for deployment on Amazon Bedrock. In May, Cerebras went public on Nasdaq in the largest U.S. technology IPO since 2019, raising $5.55 billion at a $95 billion valuation; the stock surged 68 percent on its first day of trading.

The technology behind the run is the Wafer Scale Engine, the largest AI chip ever built. Where a conventional processor occupies a small rectangle cut from a silicon wafer, Cerebras's WSE-3 spans the entire 215-millimeter-square wafer — 4 trillion transistors and 900,000 AI-optimized cores on a single piece of silicon. The design eliminates the interconnect bottlenecks of GPU clusters by keeping data on one enormous chip, at the cost of formidable engineering challenges: 25 kilowatts of power draw per system, prices up to $3 million per node, and manufacturing so demanding that only TSMC can currently fabricate the chips.

The payoff is speed. Cerebras has long claimed its inference service runs ten to twenty times faster than systems built on Nvidia's H100 GPUs, and its recent benchmarks give the claim texture: 1,600 tokens per second on DeepSeek's R1 70B reasoning model, a speed record powering Mistral's Le Chat service, 1,200 tokens per second for Perplexity's Sonar model, and — in independent testing — more than 2,500 tokens per second per user on Meta's 400-billion-parameter Llama 4 Maverick, against 1,000 for Nvidia's Blackwell. In 2025, the company reported $510 million in revenue and $87 million in net income — rare profitability for a company still in land-grab mode.

Why this deal matters beyond the two companies

The Cerebras–Gimlet agreement is best read as a data point in three larger shifts reshaping the AI hardware market.

First, the center of gravity is moving from training to inference. The last three years of the AI boom were dominated by whoever could amass the most GPUs to train the biggest models. But trained models are now abundant — open-weight models from Alibaba's Qwen, DeepSeek, and Meta's Llama families run competitively on third-party clouds — and the scarce resource is fast, cheap serving capacity. Every startup promising real-time AI agents, voice interfaces, or instant reasoning is an inference customer. Gimlet's entire company is a bet that serving tokens becomes a bigger business than training models.

Second, the "Nvidia alternative" trade is maturing from press releases into purchase orders. Cerebras spent years as a fascinating technology with a thin customer list; in 2026 it has converted that technology into an OpenAI mega-deal, an AWS deployment, a blockbuster IPO, and now a 100-megawatt startup commitment. The company's four major customers now span a Gulf AI university, G42, OpenAI, and AWS — and Gimlet adds a fifth logo from the startup world. Nvidia remains dominant, but the field of credible challengers — Cerebras, AMD, and a clutch of well-funded startups — has never been deeper.

Third, multi-silicon is becoming the industry's working assumption. Gimlet is not alone in arguing that the future is heterogeneous: the biggest cloud providers all now mix GPUs, custom accelerators, and CPUs in their fleets, and chip startups from d-Matrix to Etched are designing specifically for the decode phase of inference. If Gimlet can demonstrate that orchestrating Cerebras wafers alongside GPUs delivers 3,000 tokens per second at production scale, it will validate a model that favors software orchestration over any single chip — a genuinely new axis of competition.

Cerebras–Gimlet deal FAQ

What is Gimlet Labs?
Gimlet Labs emerged from stealth in October 2025, founded by Zain Asgar, Michelle Nguyen, Omid Azizi and Natalie Serrino — the team behind Pixie Labs, acquired by New Relic in 2020. Asgar is a former Nvidia GPU architect and Google AI researcher, and a Stanford adjunct computer science professor. In early September the startup closed a $300 million Series B at a $3 billion valuation, led by Menlo Ventures, which had also led its $80 million Series A in March — bringing total funding to roughly $392 million.

What is multi-silicon inference?
Gimlet disaggregates inference into stages and assigns each stage to the hardware best suited for it: GPUs handle the "prefill" phase, where the model processes an incoming prompt, while memory-bandwidth-rich accelerators — like Cerebras's wafer-scale chips — handle the "decode" phase, where the model generates tokens one by one. The company claims the approach delivers up to ten times the throughput and interactivity within the same power envelope, and says it works with silicon from Nvidia, AMD, Intel, Arm, Cerebras and d-Matrix.

What is Cerebras's Wafer Scale Engine?
The largest AI chip ever built. The WSE-3 spans an entire 215-millimeter-square wafer, with 4 trillion transistors and 900,000 AI-optimized cores on a single piece of silicon. The design eliminates the interconnect bottlenecks of GPU clusters by keeping data on one enormous chip — at the cost of formidable engineering: 25 kilowatts of power draw per system, prices up to $3 million per node, and manufacturing so demanding that only TSMC can currently fabricate the chips.

When will Gimlet Cloud's Cerebras capacity be available?
CS-4 systems will arrive at Gimlet over the next one to two years, and the Gimlet Cloud's Cerebras-powered offering is targeted for 2027. In the meantime, the private deployments already serving tokens are the proving ground — early customers in cybersecurity, voice and finance stress-testing whether the multi-silicon architecture holds up under real production load.

What happens next

The timeline is measured in years, not weeks. CS-4 systems will arrive at Gimlet over the next one to two years, and the Gimlet Cloud's Cerebras-powered offering is targeted for 2027. In the meantime, the private deployments already serving tokens will be the proving ground — early customers in cybersecurity, voice, and finance stress-testing whether the multi-silicon architecture holds up under real production load.

The risks are the usual ones for ambitious infrastructure bets. Cerebras must manufacture and deliver wafer-scale systems — among the most complex computers ever built — on schedule and at quality. Gimlet must convert its $3 billion valuation and $392 million in funding into a cloud business that can fill 100 megawatts of capacity with paying customers, in a market where hyperscalers and well-funded neoclouds are all chasing the same inference workloads. And the 3,000-tokens-per-second figure, impressive as it sounds, will need to survive contact with the messy reality of diverse customer models and traffic patterns.

But the direction of travel is clear. A year ago, Gimlet was a stealth startup with a clever architecture and a famous founding team. Today it is a $3 billion company with 100 megawatts of the world's largest AI chips on order. In an industry where compute is destiny, that is about as strong a statement of intent as a startup can make.

Related TechPulse USA coverage: VSMC Opens $7.8B Singapore Fab, Already Eyes a Second — on the parallel boom in chip manufacturing capacity feeding the AI buildout.

Sources: Reuters (Sept. 28, 2026); Gimlet Labs / Cerebras joint announcement via GlobeNewswire and The Manila Times (Sept. 28, 2026); WOWTALE on Gimlet's Series B (Sept. 5, 2026); Cerebras Systems company background via Wikipedia (updated Sept. 2026, citing SEC filings and company announcements).

Note on this story: Financial terms of the Cerebras–Gimlet agreement were not disclosed, and neither company has published the number of systems covered. The "up to 3,000 tokens per second" figure and the "10x throughput" claim are company statements about expected and architectural performance, not independently verified benchmarks. Gimlet's cloud availability target of 2027 is a company projection.