
Gimlet Labs Raises $300M at $3B Valuation to Build the First Multi-Silicon Inference Cloud
Andreessen Horowitz leads the Series B for a company that decomposes AI models across GPUs, CPUs, near-memory compute and dataflow chips — with Arm and Microsoft's M12 betting inference is about to escape the GPU.
Gimlet Labs has raised $300 million in a Series B led by Andreessen Horowitz, valuing the applied AI research company at $3 billion barely a year after it emerged from stealth. Sapphire Ventures, M12, Arm, Menlo Ventures and Factory participated, bringing total funding to $392 million.
Inference beyond the GPU
Gimlet's core claim is that agentic AI has broken the one-chip-fits-all model of inference. Its platform decomposes models across a heterogeneous fleet — GPUs, CPUs, near-memory compute and dataflow architectures — routing each phase of a workload to the silicon that executes it most efficiently. The company calls it the industry's first multi-silicon inference cloud, purpose-built for the fast, interactive, many-small-calls pattern of agentic workloads.
That thesis has structural allies on the cap table. Arm's participation signals interest in CPUs taking a larger share of inference; M12's suggests Microsoft is hedging the same GPU-concentration risk that pushed Broadcom's custom-silicon revenue to $16.7 billion last quarter.
Where the money goes
The new capital funds a major expansion of the platform's serverless edition, with plans to add several hundred megawatts of computing power. More ambitiously, Gimlet intends to enter the custom hardware market with an inference-optimized server that dispenses with a motherboard entirely and can operate outside traditional data centers — pushing agentic inference toward factory floors, retail sites and telecom edges.
The context
The round completes a striking week for AI infrastructure: Crusoe's $3 billion, Fluidstack's $1.5 billion, and now Gimlet's $300 million — three different bets (owned data centers, orchestrated capacity, heterogeneous silicon) on the same underlying conviction that inference demand is about to dwarf training. With Nvidia consolidating the open ecosystem via its $13 billion Hugging Face deal and its Groq-derived inference chips, the counter-programming is clear: everyone else is funding the world where inference escapes the GPU monoculture. Gimlet is now the best-funded pure-play version of that future.
Newsletter
Get Lanceum in your inbox
Weekly insights on AI and technology in Asia.
More in Startups

Travis Kalanick's Atoms Eyes Robotaxis — With Uber's $100 Million and His Old AV Team

Fluidstack Closes $1.5 Billion Led by Jane Street — Revenue Up From $1.8M to a Projected $660M, While Owning Zero Chips
