The Pulse
Gimlet and Cerebras Target 3,000 Tokens per Second
Gimlet Labs and Cerebras plan to combine wafer-scale systems and GPUs in Gimlet Cloud, targeting AI inference speeds of up to 3,000 tokens per second. The companies say their first joint data center is expected later in 2026, while Reuters
AI.info Team · Written with AI ·
A 3,000-token target, not a published benchmark
Gimlet Labs and Cerebras say their planned inference cloud could reach 3,000 tokens per second. The number is a target for demanding agentic and real-time applications, not a reported result from a disclosed benchmark: the companies have not specified the model, test conditions or number of simultaneous users behind it.
The September 28 announcement pairs Cerebras wafer-scale systems with GPUs and Gimlet’s software, which assigns different parts of model inference to different hardware. The companies say their first Cerebras-powered data center is expected to come online later in 2026. That timetable and speed remain plans, rather than delivered customer performance.
“Inference speed matters. It determines how productive AI can be.”
Zain Asgar, co-founder and CEO of Gimlet Labs
Gimlet assigns different inference phases to different chips
The proposed system is not a Cerebras-only cloud. Gimlet describes an architecture that combines Cerebras Wafer Scale Engine systems with GPUs, routing stages such as prefill, token generation and attention to the hardware it considers best suited to each one.
That division reflects a basic tension in inference infrastructure: a data center can maximize aggregate throughput while an individual user still waits for tokens. Gimlet’s stated aim is to improve responsiveness without giving up system efficiency. The company says its orchestration software will present developers with a standard inference API, even as work runs across different silicon.
Gimlet says the collaboration builds on customer work underway since last year and an integrated system already serving inference traffic in private deployments. It also says it will be a launch partner for Cerebras CS-4 systems, with access for Gimlet Cloud customers expected in 2027.
Reuters reports a 100-megawatt supply plan
Reuters reported on September 28 that the companies plan roughly 100 megawatts of Cerebras-powered capacity. Cerebras expects to supply CS-4 systems over one to two years, Reuters reported, while Gimlet plans to make the hardware available through its cloud in 2027.
The companies did not disclose financial terms. Reuters reported that Gimlet will maintain and operate the systems after delivery, putting the day-to-day running of the deployment with the cloud provider rather than the chip supplier.
Andrew Feldman, Cerebras’ chief executive, described the deployment as “further validation from Cerebras... of how easy it is to deploy, even in heterogeneous environments,” in comments to Reuters. The arrangement tests whether wafer-scale systems can be installed alongside other vendors’ hardware and managed as part of a single cloud service.
Availability will decide what the headline speed means
For developers, the practical question is not only whether a system can generate tokens quickly, but what speed it sustains for a given model and workload as more customers use it. The announcement gives a top-line target and a projected data-center schedule, but no pricing, capacity rollout by customer, or benchmark details to judge that performance against.
The next concrete milestones are the first data center’s expected arrival later in 2026 and the planned CS-4 access in 2027. Until Gimlet publishes operating results and service terms, 3,000 tokens per second describes the companies’ ambition—not a generally available service level.