Gimlet Labs closed a $300 million Series B at a $3 billion valuation this week, and the pitch cuts against how most cloud providers operate. Clouds usually keep customers on one chip family, because that’s simpler to support. Gimlet wants the chip itself to stop mattering.
Andreessen Horowitz led the round. Sapphire Ventures, M12, Arm, Menlo Ventures and Factory also joined, pushing total funding to $392 million. All of it feeds Gimlet Cloud, the company’s managed inference service, which it says already carries billions in contracted revenue.
The software sits above the hardware and splits inference jobs across NVIDIA, AMD, Intel, Arm, Cerebras and d-Matrix chips, absorbing the differences in software rather than forcing cloud customers onto one vendor. Most hyperscale clouds work the other way around: you pick a chip, then build the stack on top of it. Whether Gimlet’s approach holds up under real production traffic is genuinely unclear, since these chips don’t share memory models, kernels or networking behavior. The tenfold throughput claim comes from Gimlet itself, not an outside benchmark, so treat it accordingly.
What’s likely driving the timing is where cloud spending is heading. Training costs hit once. Inference keeps running, and it multiplies fast once agents start calling models and other cloud services on their own, without a single new human user in the picture. Gimlet is essentially betting that the layer deciding which chip handles which request becomes as valuable as the chips it’s routing to.
The company says it added a top-three frontier lab and a top-three hyperscaler as cloud customers this year, tripling its customer base. It won’t say who. Gimlet Cloud can run as a managed service or inside a customer’s own environment, which helps enterprises already juggling mixed hardware fleets, though it doesn’t retire the existing cloud tooling those enterprises run for observability, security and model serving.
Then there’s the physical reality. Hundreds of megawatts of planned capacity runs straight into power, cooling and rack-density limits, and no orchestration layer gets around those on its own. Gimlet has the money now. Running this cloud at real scale is the part still ahead of it.
