Gimlet Labs raises $300M at a $3B valuation to run AI inference across every chip
The Andreessen Horowitz-led Series B backs a multi-silicon inference cloud that splits agentic AI workloads across GPUs, CPUs and purpose-built accelerators.
Gimlet Labs has raised $300 million in a Series B led by Andreessen Horowitz, valuing the applied AI research and product company at $3 billion and bringing its total funding to $392 million.
Sapphire Ventures participated, as did new investors M12 - Microsoft's venture fund - and Arm. Existing backers Menlo Ventures and Factory also joined.
What the company is building
Gimlet is building what it calls a multi-silicon inference cloud for agentic AI. The premise is that a single AI workload does not want a single kind of processor.
The company's software disaggregates AI models and assigns different phases of an inference workload to whichever hardware suits that phase best, with the aim of improving latency, throughput, utilisation and power efficiency at the same time. In practice that means running across GPUs, CPUs and purpose-built AI accelerators rather than standardising on one architecture.
Gimlet says it works with NVIDIA, AMD, Intel, Arm, Cerebras and d-Matrix. Having Arm on the cap table alongside that list is not incidental.
The scale claim
The company says it has secured billions of dollars in contracted revenue for Gimlet Cloud and is scaling its managed heterogeneous infrastructure to hundreds of megawatts.
Contracted revenue is not recognised revenue, and megawatts under management is a forward-looking figure rather than a delivered one. Both are worth tracking rather than taking at face value. But the shape of the claim tells you what the company is being funded to do: not to research inference efficiency, but to operate a fleet.
Why it matters
Inference has quietly become the expensive part of running AI in production. Agentic workloads make it worse, because a single user request can fan out into many model calls rather than one.
The bet Gimlet's investors are making is that the answer to that cost curve is not a faster chip from any one vendor, but an orchestration layer that treats all of them as interchangeable capacity. If that is right, the value accrues to whoever owns the scheduler rather than the silicon.
Sources
Comments
Loading comments…
