Cerebras has agreed to supply 100 megawatts of its AI hardware to cloud startup Gimlet Labs, one of the largest power-denominated compute deals yet announced by the wafer-scale chipmaker.

The figure is the story. Not wafers, not units, not dollars — megawatts. That's the currency hyperscalers and cloud upstarts now trade in, because rack power, not chip count, sets how much AI compute a data center can actually absorb. A 100 MW commitment places Gimlet Labs, a startup, in the same negotiating language used by operators spending billions on GPU clusters.

Who is supplying what?

Cerebras builds its systems around the WSE family — wafer-scale engines that pack an entire wafer's worth of logic into a single package, rather than dicing silicon into conventional reticle-sized chips. The company has marketed these systems as inference machines, with claims of high token throughput for large language models served from a single rack footprint.

Gimlet Labs is the buyer — a cloud startup positioning itself to rent AI compute to developers and enterprises. For a company without the balance sheet of a hyperscaler, signing a 100 MW hardware agreement signals intent to compete for inference workloads rather than cede that market to the giants. The deal's financial terms and delivery schedule have not been disclosed, so treat the 100 MW figure as a stated commitment rather than shipped, deployed capacity.

Why measure AI compute in megawatts?

Because power is now the binding constraint. Across the industry, data center operators report that grid capacity, cooling and interconnection queues — not chip availability alone — determine how fast AI infrastructure scales. Deals framed in megawatts tell investors and customers what a provider can actually energize, not just what it has ordered.

For Cerebras, the framing does double duty. Its wafer-scale approach concentrates enormous compute per unit, and selling by the megawatt lets the company argue that fewer systems deliver more usable inference per watt and per square meter of data center floor.

What does this mean for the competitive picture?

Cerebras remains a challenger in a market where Nvidia captures the majority of AI accelerator spending. Its pitch has rested on inference economics: serve large models faster from compact systems, and sidestep the cluster-scale networking that GPU fleets demand. A cloud partner willing to commit 100 MW gives Cerebras a flagship deployment vehicle — and a reference customer that other operators can watch before writing their own orders.

For Gimlet Labs, the bet is that demand for fast, cost-efficient inference keeps outpacing supply, and that customers will rent from a nimble specialist rather than wait in queue at the hyperscalers.

What's confirmed, what isn't

Confirmed: Cerebras will supply 100 MW of AI hardware to Gimlet Labs under the announced agreement. Unconfirmed: revenue value, hardware mix across specific WSE generations, deployment timeline, and which data centers will host the systems. None of those details appeared in the announcement as reported, so any projection of deal revenue would be an analyst estimate at best.

The industry will be watching deployment milestones. If Gimlet Labs energizes that capacity on schedule, the deal becomes proof that wafer-scale inference can anchor a commercial cloud; if it slips, the 100 MW headline joins the long list of AI announcements that aged faster than the silicon they promised.