Gimlet Labs raised $300M at a $3B valuation, up from $400M six months ago

Arm designs the architecture that most non-Nvidia AI silicon is built around, and has an obvious interest in software that makes non-Nvidia silicon easier to use. Microsoft runs Azure, buys enormous quantities of Nvidia hardware, and has its own accelerator programme it would like customers to be able to reach without rewriting anything.

Both are investing in the removal of a switching cost. The reason Nvidia's position has held through three years of well-funded competition is not only that its chips are fast; it is CUDA, and the fact that a workload built for it does not simply move. Every hyperscaler is building alternative accelerators, and every one of those programmes runs into the same problem, which is that customers will not port.

Gimlet is selling the port. That is worth a great deal to everyone in the industry except one company.

The timing is not subtle

The round closed in the same fortnight that Nvidia agreed to buy Hugging Face for $12.9 billion — a deal whose logic, as reported, includes keeping the open-weight ecosystem anchored to Nvidia hardware.

So the market is now pricing both sides of the same question at once: Nvidia paying $12.9 billion to make its position stickier, and a16z, Arm and Microsoft paying $300 million for the tool that makes it less so.

Nvidia, for its part, committed to keeping the Hugging Face hub open to competing models and competing chips. Whether that survives is the term everyone will be watching, and its credibility is worth something to Gimlet either way: if Nvidia honours it, cross-chip deployment gets easier; if it does not, the demand for an independent routing layer rises.

What has to be true for $3 billion

Cross-chip routing is a hard engineering problem that has defeated well-resourced attempts before, because performance portability is not the same as functional portability. Getting a model to run on alternative silicon is achievable; getting it to run at a cost and latency that justify the move is the actual product, and it has to be re-earned with every new chip generation.

The valuation also assumes the alternatives become genuinely competitive. Routing software is worth little if there is nowhere good to route to. That depends on Arm-based accelerators, Microsoft's Maia line, AMD's roadmap and the hyperscalers' internal programmes closing the gap — none of which Gimlet controls.

A 7.5-fold markup in six months on a company at this stage is a bet on the strategic need rather than on demonstrated revenue. Gimlet has not disclosed revenue, customers, or what workloads are running through its software in production.