Cerebras and Gimlet to Build AI Cloud at Speeds of up to 3,000 Tokens per Second

Cerebras Systems and Gimlet Labs have announced a collaboration to build a new cloud infrastructure for AI inference, combining Cerebras’ chip-level computing with Gimlet’s cloud software in a move targeting applications that require ultra-fast responses.

Nearly 100 megawatts of capacity

According to Reuters, Cerebras plans to supply Gimlet Labs with AI chips and systems with an operating capacity of nearly 100 megawatts, with the systems to be deployed over one to two years. The two companies did not disclose the financial value of the agreement.

Gimlet plans to use the hardware to run advanced models that require substantial computing power, with a focus on the “inference” stage — the stage at which a model generates answers and results after training.

Up to 3,000 tokens per second

In the official announcement, the companies said the planned infrastructure is targeting speeds of up to 3,000 tokens per second for AI agent applications and real-time applications, with the first Gimlet Cloud data center powered by Cerebras systems expected to become operational in 2026.

Speed becomes a competitive battleground

Fast inference is becoming increasingly important as voice applications, cybersecurity, financial analysis and AI agents proliferate. Rather than focusing solely on training larger models, companies are moving toward reducing response times and improving the efficiency of running models for end users.

Sources: Cerebras Systems, September 28, 2026; Reuters, September 28, 2026.