Cerebras has agreed to supply AI chips equivalent to 100 megawatts of capacity to Gimlet Labs, a cloud startup, according to a report by Quartz. The agreement commits Cerebras to delivering its wafer-scale accelerator hardware at a scale measured not in unit counts but in total power draw — a metric that has become a standard yardstick for large AI compute deals as demand for training and inference capacity outpaces conventional supply channels.

The 100-megawatt figure places the arrangement among the larger AI hardware supply commitments recently disclosed in the sector. Buyers of AI compute increasingly frame procurement in power terms because data center capacity, grid interconnection and cooling now constrain deployment more than chip fabrication alone. For a young cloud provider such as Gimlet Labs, securing a power-denominated allocation of accelerators signals an intent to compete for AI workloads that require dedicated, high-density infrastructure rather than rented slices of hyperscaler capacity.

Cerebras differentiates itself from mainstream GPU suppliers through its wafer-scale technology. Instead of manufacturing individual processors and packaging them onto boards, the company builds a single chip that occupies most of a standard wafer. The approach yields one large device with an extensive on-chip memory hierarchy and high interconnect bandwidth, characteristics the company positions as advantageous for large-model inference in particular.

For Cerebras, the Gimlet Labs agreement adds a second pillar to a commercial strategy that has so far rested heavily on direct access to its hardware through its own cloud service. Selling 100 megawatts worth of chips to an external cloud operator moves the company further into the supply model dominated by established accelerator vendors, where chipmakers ship hardware to partners who then operate it as a service.

The deal also reflects a broader shift in how AI compute reaches customers. A growing cohort of startups is building businesses around offering access to specialized accelerators — hardware that differs architecturally from the GPU fleets most cloud providers stock. These neocloud operators bet that customers will route workloads to whichever platform delivers the best throughput and latency per dollar, regardless of whether the underlying silicon comes from a market leader or a challenger.

Under an agreement of this scale, deployment will proceed in phases that align hardware deliveries with data center readiness. Power capacity of 100 megawatts typically corresponds to a multi-building or multi-hall data center footprint once cooling, power distribution and redundancy are accounted for. The parties did not publicly specify a delivery schedule, financial terms or deployment locations in the initial report.

The transaction arrives amid sustained competition for AI accelerator supply, in which cloud operators, sovereign programs and enterprise buyers have signed multi-year capacity agreements covering power allocations of comparable or larger size. Within that context, a 100-megawatt commitment directed to a single startup cloud provider stands out for its concentration: the entire allocation supports one operator's platform rather than being divided across a broad customer base.

For prospective Gimlet Labs customers, the agreement implies future access to Cerebras-class inference capacity at scale. For Cerebras, it provides a committed offtake for its silicon through a partner that will bear the operational burden of running it. Quartz reports the deal as a supply agreement covering 100 megawatts of AI chips; additional details on pricing, timelines and initial deployment sites were not disclosed.