Gimlet Cloud to deploy 100MW of Cerebras wafer-scale compute for inference

One hundred megawatts of Cerebras wafer-scale compute will run on Gimlet Cloud under a partnership the two companies have announced, with the first data center expected online later this year and no address attached to the commitment yet.

What the agreement does not arrive with is an address, and Data Center Dynamics, which reported the partnership, notes that Gimlet has not provided details of its data center footprint and is hiring for a data center facilities operations lead. No site, landlord, utility or contract value appear in the coverage, which leaves a 100MW commitment reading more like a capacity announcement than a project.

Gimlet's money is new: the San Francisco company raised a $300 million Series B led by Andreessen Horowitz at a $3 billion valuation weeks before the Cerebras agreement, drawing a syndicate that included Arm, Samsung Ventures, Menlo Ventures, M12, Tiger Global Management and 645 Ventures among a longer list, and an $80 million Series A closed in March. Gimlet was founded by Zain Asgar, Michelle Nguyen, Omid Azizi, Natalie Serrino and James Bartlett; Asgar, the chief executive, also teaches computer science at Stanford.

At the time of the round, Asgar said Gimlet had secured billions of dollars in contracted revenue since March and was scaling toward hundreds of megawatts of managed capacity, with the funding earmarked for building that infrastructure, expanding the team and reaching more customers. The Cerebras agreement is one installment against that number, but the coverage does not say where the rest of the capacity would sit, whether Gimlet intends to build or lease it, or who pays for the shells and the power.

Gimlet's technical argument is that no single chip wins inference: GPUs and wafer-scale systems are good at different parts of the job, and the opportunity lies in splitting one inference workload across both, with each stage going to the silicon that handles it best, in pursuit of higher throughput and lower latency. Cerebras gets a cloud channel, and Gimlet gets a silicon mix that GPU-only competitors have to assemble deal by deal. Gimlet will also be one of the launch partners for Cerebras' next-generation CS-4 hardware, with availability for Gimlet customers in 2027, which puts part of the operator's roadmap on a generation the source describes as upcoming. The coverage describes Cerebras as a wafer-scale chip and cloud company, making it both Gimlet's supplier and, on inference workloads, a competitor.

Cerebras comes to the arrangement with a buildout already staged against demand; PWD reported in September that its Mikkeli campus in Finland had 50MW under construction and a further 115MW left to be earned by demand, a structure that kept the chipmaker from carrying an entire campus on its own balance sheet before customers existed. A 100MW partnership with a funded cloud operator extends the same approach in the other direction: the capacity gets announced, and the proving-out moves to a partner. Which party buys the buildings, the power and the wafers is not something the coverage addresses.

The unit of the deal matters as much as its size: Cerebras is selling in megawatts, the currency of the data center market rather than of the chip order book, which places the agreement alongside the campus and power deals that underwrite this build-out and raises the question those deals raise: who signs for the load, and who owns the asset at the end of it.

Gimlet's investor list puts two trading firms on the cap table, Hudson River Trading and XTX Markets. Hudson River Trading signed its own multi-billion-dollar AI cloud agreement with CoreWeave in August, a deal that came with no dollar figure, capacity number, or schedule made public, the same withholding that surrounds the Cerebras agreement. Crusoe's reported $13 billion Jane Street agreement noted a five-year term with a trading firm as the anchor, a different instrument from the hyperscaler leases that usually underwrite a campus. A trader that buys GPU capacity at that scale and holds equity in a wafer-scale cloud has exposure on both sides of the inference trade, which is either conviction about demand or a hedge on its price.

Who carries the megawatts decides whether any of this becomes infrastructure. Data center capital now prices the anchor rather than the megawatt, and a financing without a named tenant is merchant risk in disguise. Gimlet is its own anchor: a cloud operator assembling capacity for customers it has not named, on hardware it does not make, at sites it has not disclosed. That is a coherent way to sell inference and a difficult position to underwrite from the outside.

The first data center under the agreement is expected online later this year, and CS-4 capacity is promised to Gimlet's customers in 2027. Before then, the details that would turn 100MW from an announcement into an asset: a site, a utility, a landlord, a contract value.