Cerebras to Supply Gimlet Labs With AI Systems Designed for About 100 Megawatts of Capacity

Cerebras plans to supply AI chips and hardware to Gimlet Labs with capacity to consume roughly 100 megawatts. Cerebras and Gimlet Labs are linking AI accelerator supply to a much larger question: how much computing infrastructure will be required to serve AI models continuously. The companies said Cerebras will supply systems to Gimlet Labs with capacity designed around roughly 100 megawatts.

A large infrastructure commitment

Gimlet said the hardware is expected to become available in its cloud in 2027.

Gimlet intends to use that infrastructure for an AI-focused cloud business. The arrangement points to the growing distinction between training infrastructure and inference infrastructure. Training is often associated with large bursts of compute, while inference can create a persistent demand as models are used by applications and end users.

The deal’s financial terms were not disclosed and Gimlet will operate and maintain the systems once delivered.

Cerebras’ product information provides context on its wafer-scale approach and the AI infrastructure involved in the Gimlet Labs project.

The Cerebras-Gimlet relationship is about inference infrastructure rather than model training alone. Gimlet is building a cloud platform intended to make large-scale AI inference easier to deploy, while Cerebras supplies specialized systems designed for high-throughput workloads.

A capacity target of roughly 100 megawatts is significant because it describes an infrastructure footprint rather than a single server installation. Power, cooling, networking and facility design become part of the AI service itself at that scale.

A project at this scale also brings power and facility constraints into the technology discussion. Chips are only one part of the equation. A cloud operator needs electricity, cooling, networking, storage and software capable of keeping expensive accelerators busy.

Cerebras systems are built around wafer-scale processing, which takes a different approach from assembling large numbers of conventional accelerator cards. The value for an inference provider depends on how that architecture performs under real application workloads and how efficiently it can be operated.

The deal therefore provides a useful indication of where the AI infrastructure market is heading. Performance per chip still matters, but operators increasingly have to evaluate the entire system around the accelerator and the cost of delivering useful inference at scale.