Why this matters: the AI race is shifting from training bigger models to serving answers faster — and speed may decide who wins the agent era.
September 28, 2026 — Cerebras Systems will supply AI chips and hardware with capacity to consume roughly 100 megawatts to cloud startup Gimlet Labs, the companies announced Monday, in a push to deliver AI inference at up to 3,000 tokens per second. The collaboration pairs Cerebras's giant wafer-scale processors with Gimlet's cloud software, targeting the real-time and agentic workloads where slow responses are a dealbreaker.
Image: Original illustration of a giant glowing wafer-scale AI processor in a datacenter, symbolizing the ultrafast inference push. Credit: Red Eye Eagle News / original AI illustration.
Editor's Highlight: Training gets the headlines, but inference pays the bills: this deal is a bet that the next AI gold rush belongs to whoever can generate tokens fastest at production scale.
Table of Contents
- Key Facts
- The deal: 100 megawatts of wafer-scale silicon
- What 3,000 tokens per second means
- Why inference is the new battleground
- Who Gimlet wants to sell speed to
- Bottom Line
- FAQ
- Sources
Key Facts
| Detail | Information |
|---|---|
| Announced | September 28, 2026 (GlobeNewswire; Reuters) |
| Partners | Gimlet Labs and Cerebras Systems (NASDAQ: CBRS) |
| Speed target | Up to 3,000 tokens per second for agentic and real-time apps |
| Hardware scale | Capacity to consume roughly 100 megawatts (Reuters) |
| Systems | Cerebras CS-4, unveiled over the summer; delivery over 1-2 years |
| First datacenter | Expected online later this year, per the companies |
| Technology | Cerebras Wafer Scale Engine plus GPUs, with inference disaggregation |
| Gimlet backers | Andreessen Horowitz and Menlo Ventures; based in San Francisco |
The deal: 100 megawatts of wafer-scale silicon
The core of the announcement is hardware. Cerebras plans to supply Gimlet Labs with AI chips and hardware with capacity to consume roughly 100 megawatts of power, according to Reuters, which reported the story on Monday. That is a datacenter-scale commitment, delivered over one to two years.
The systems at the center of the deal are Cerebras's CS-4, unveiled over the summer. Reuters reports that Gimlet plans to make the CS-4 hardware available in its cloud in 2027. The companies declined to disclose the financial terms, and Gimlet will be responsible for maintaining and operating the systems once Cerebras delivers them.
In parallel, the companies say the first Cerebras-powered Gimlet Cloud datacenter is expected to come online later this year — likely using existing wafer-scale hardware while the newer CS-4 systems roll out.
Newsroom Focus: A 100-megawatt hardware commitment is not a pilot project. It signals that both sides believe ultrafast inference is a market that will exist at real scale, soon enough to justify the power bill.
What 3,000 tokens per second means
The headline number is 3,000 tokens per second — the speed the companies say their combined system is designed to deliver for demanding agentic and real-time applications. Tokens are the chunks of text an AI model generates, so this is a measure of how fast a model can think and reply.
For comparison, a typical AI chatbot today generates answers at tens to hundreds of tokens per second. At 3,000 tokens per second, a model could draft a full page of text in roughly the blink of an eye — fast enough for real-time voice assistants, live translation and agents that take many reasoning steps without making the user wait.
The technical trick, according to the companies, is inference disaggregation. Gimlet Cloud combines the Cerebras Wafer Scale Engine — a processor nearly the size of an entire silicon wafer — with GPUs into one system, and uses software to route each phase of inference to whichever silicon handles it best. Fast chips take the latency-sensitive work; high-throughput GPUs take the rest.
| Partner | What it brings |
|---|---|
| Cerebras | Wafer Scale Engine, CS-4 systems, ~100MW of hardware |
| Gimlet Labs | Multi-silicon software, Gimlet Cloud, disaggregation tech |
| Jointly | APIs, developer tooling, optimization, production operations |
"Inference speed matters. It determines how productive AI can be. Fast inference creates magical user experiences and opens new markets," said Zain Asgar, co-founder and CEO of Gimlet Labs, in the announcement.
Analysis Note: The 3,000-tokens-per-second figure is a target, not a measured result — the companies say they "plan to deliver" it. Treat it as a design goal until independent benchmarks confirm it.
Why inference is the new battleground
For most of the AI boom, the story was training: who could build the biggest model on the most GPUs. But the economics are shifting. As companies deploy AI agents that run continuously — browsing, reasoning, calling tools — the cost and speed of generating responses (inference) has become the binding constraint.
That is why hardware capable of fast inference is suddenly in demand across the industry. Reuters notes that OpenAI recognized the same need: earlier this year, Cerebras signed a deal to supply OpenAI with chips. And last year, Nvidia signed a licensing deal with Groq for that company's chips and hardware — a striking move from the GPU giant itself.
Cerebras, which went public earlier this year, has built its identity on wafer-scale computing: instead of cutting a silicon wafer into hundreds of small chips, it turns nearly the entire wafer into one enormous processor. The company says this keeps data on silicon instead of shuttling it across racks, cutting latency for inference workloads.
"Combining the fastest tokens from Cerebras with the highest throughput GPUs delivers the best datacenter economics for everyone," said Sean Lie, co-founder and CTO of Cerebras. He added that Gimlet is a launch partner for CS-4, giving customers a direct path to Cerebras's latest technology.
| Date | Event |
|---|---|
| Summer 2026 | Cerebras unveils CS-4 wafer-scale systems |
| Sep 28, 2026 | Gimlet Labs–Cerebras collaboration announced |
| Later 2026 | First Cerebras-powered Gimlet Cloud datacenter online (per companies) |
| 2027 | CS-4 available in Gimlet's cloud (per Reuters) |
Who Gimlet wants to sell speed to
Speedy inference matters most where waiting is expensive. In an interview with Reuters, Asgar named cybersecurity, voice applications and financial analysis as areas where faster responses make a real difference.
For this hardware, Gimlet plans to target companies and startups building their entire products around AI — the agentic coding tools, real-time voice products and automated analysis services that live or die on responsiveness. Gimlet plans to use the Cerebras hardware to run frontier models, the cutting-edge AI systems that require enormous computing power to serve quickly.
The collaboration builds on joint customer work already underway: the companies say they have been engaged with shared customers since last year, and that an integrated solution is already serving tokens in private deployments.
Practical Takeaway: If you build on AI APIs, watch inference-speed offerings like this one: faster tokens can mean shorter agent run times and lower per-task costs, even before anyone cuts prices.
Bottom Line
The Cerebras–Gimlet Labs deal turns the industry's inference anxiety into hardware: roughly 100 megawatts of wafer-scale silicon aimed at 3,000 tokens per second, with a first datacenter due this year and CS-4 systems arriving through 2027.
What to watch: whether the 3,000-tokens-per-second target holds up in independent benchmarks, how the financial terms (still undisclosed) affect pricing for developers, and whether the OpenAI and Nvidia-Groq moves signal a full industry pivot from training hardware to inference hardware.
FAQ
What is AI inference?
Inference is the process of a trained AI model generating responses — for example, when a chatbot answers your question. It is distinct from training, which is the expensive one-time process of teaching the model. Inference happens every time someone uses the model, so speed and cost per response matter enormously at scale.
What are Gimlet Labs and Cerebras?
Cerebras Systems (NASDAQ: CBRS) builds wafer-scale AI computers — processors nearly the size of an entire silicon wafer, designed for ultra-fast AI inference. Gimlet Labs is a San Francisco startup, backed by Andreessen Horowitz and Menlo Ventures, building an inference cloud that combines different kinds of chips with software that routes each workload to the best hardware.
How fast is 3,000 tokens per second?
Very fast by current standards: typical chatbots generate tens to hundreds of tokens per second. At 3,000 tokens per second, a model could produce a full page of text almost instantly. The figure is the companies' stated target, not yet an independently verified benchmark.
When will developers get access?
The companies say the first Cerebras-powered Gimlet Cloud datacenter comes online later this year, and an integrated solution is already serving tokens in private deployments. Reuters reports the newer CS-4 systems will be available in Gimlet's cloud in 2027.
How does this fit the wider AI chip race?
It fits a clear trend: the industry is shifting investment from training hardware to inference hardware. Cerebras already has a chip-supply deal with OpenAI (signed earlier this year), and Nvidia licensed Groq's inference hardware last year — both signs that fast token generation is now a strategic priority across the industry.
Sources
- GlobeNewswire, "Gimlet Labs Adds Cerebras to Deliver Ultrafast AI Inference," September 28, 2026 — joint announcement, 3,000 tokens/sec target, datacenter timeline, executive quotes: https://www.globenewswire.com/news-release/2026/09/28/3369931/0/en/gimlet-labs-adds-cerebras-to-deliver-ultrafast-ai-inference-through-gimlet-cloud.html
- Reuters (via 710 AM WDSM), "Cerebras to supply AI systems to cloud computing startup Gimlet Labs," September 28, 2026 — 100-megawatt scale, 1-2 year delivery, CS-4 in cloud in 2027, OpenAI/Groq context: https://wdsm710.com/2026/09/28/cerebras-to-supply-ai-systems-to-cloud-computing-startup-gimlet-labs/
- Valley101, "How a Chinese AI lab proved Cerebras' wild chip idea" (YouTube) — original-reporting documentary on Cerebras, wafer-scale architecture, pivot to inference: https://www.youtube.com/watch?v=9mjEH4eCrL8
- Red Eye Eagle News, original AI illustration — hero image for this article.
Join the conversation
Comments
Post a Comment