Gimlet Labs takes Cloud brains and promises AI inference to 3,000 tokens per second

Gimlet Labs will integrate Brains' wafer-scale computing into its inference cloud, with a promise to deliver speeds of up to 3,000 tokens per second. The announcement, published by Cerebras itself, places Gimlet Cloud as a platform dedicated to high-performance AI inference, available in both cloud and on-premises environments.

What Gimlet Labs offers for AI inference in practice

Gimlet Cloud is a infrastructure focused specifically on inference — the stage in which a trained AI model responds to questions and performs tasks. By adding Cerebras hardware, the company offers dedicated capacity to customers who cannot run their applications in general-purpose GPUs without sacrificing latency.

The differential of the proposal is in the delivery model. In addition to the cloud, Gimlet offers the same infrastructure in on-premises facilities, a relevant option for organizations that need to keep data within their own perimeter. The ad can be read in full in the press release of Cerebras.

Why 3,000 tokens per second is the central number of the AI inference

Token is the basic unit that an AI model processes — approximately a fraction of a word. The token rate per second defines how much information the system delivers per unit of time, and this translates into two things: user perceived response speed and cost per request.

In AI agent applications, which chain several calls to the model to complete a task, the accumulated flow weighs more than in a simple chatbot: the higher the sustained rate, the more simultaneous sessions fit the same hardware. The 3,000 tokens per second are the main commercial argument of companies — and the point that still needs to be confirmed in real production loads.

Wafer-scale: which makes the brain chip different for AI inference

Brains became known for building wafer-scale processors instead of slicing silicon into many smaller chips. The approach maintains memory and computer cores in the same component, reducing the go and come from data that usually limit the speed of inference.

This is not the first time the manufacturer has appeared in such movements. The company had already announced a joint solution of inference with AMD, focused on ultra-low latency — see in AMD and Cerebras launch a joint solution of AI inference with ultra-low latency. The pattern repeats itself: Cerebras seeks to position its hardware as a specialized layer inside third-party clouds, instead of selling only isolated machines.

What changes for developers and companies in Brazil with accelerated AI inference

For Brazilian teams, the practical question is where this capacity will be hosted. The further the data center, the greater the network latency — and the gains of milliseconds in processing can be swallowed by the data travel time. The on-premises option solves part of this problem, but requires own investment in hardware.

There is also the regulatory dimension. Loads involving personal data under the LGPD usually favor environments controlled by the organization itself, which gives weight to the local offer. The infrastructure market for AI goes in the same direction, as shows the advance of initiatives of scale AI computing for multi-inquiry factories.

What the AI inference announcement still does not answer

The statement does not detail prices, availability deadlines, assisted regions, compatible models or service level agreements. There is also no indication of customers who already use the platform or independent benchmarks that confirm the 3,000 tokens per second in real conditions.

Until these points appear, the partnership must be read as a positioning bet: Brains expanding distribution channels and Gimlet differentiating its cloud by speed.

Frequently Asked Questions about AI inference with Brains

Gimlet Labs is replacing GPUs with brain chips?

Not necessarily. The announcement describes an addition of expertise in inference, not the complete exchange of existing infrastructure. GPUs remain the basis for training and for much of the inference loads.

What does it mean to offer the on-premises solution?

It means that the same infrastructure can be installed within the client's dependencies, rather than running only in Gimlet's data centers. It matters who needs to control where the data is.

When will the offer be available and how much will it cost?

The press release does not inform general availability dates or prices. This information should come in subsequent announcements or in direct commercial contact with companies.