France's ZML lands $20M, launches chip-agnostic LLMD server

The interesting bit in ZML's launch this week isn't the $20 million or the celebrity cap table, though both help. It's the pitch itself: a single inference server that reportedly runs open source LLMs across Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc without forcing the customer to pick a silicon horse. TechCrunch reports that the French startup, founded by former Zenly VP of engineering Steeve Morin, has drawn Turing Award winner Yann LeCun, Solomon Hykes, and Hugging Face's Clément Delangue and Julien Chaumond onto the cap table alongside Harry Stebbings' 20VC and Xavier Niel's Kima Ventures.

The reason chip-agnostic matters right now is capacity. Nvidia allocation is the practical constraint on most serious inference workloads, and teams are quietly re-pricing everything against whatever accelerators they can actually get their hands on. If ZML's LLMD delivers on running the same open source models across AMD, Google TPU, and Apple Metal without a full port per backend, the calculus for a lab or startup under GPU rationing shifts in a real way. Morin's own framing, via TechCrunch, is that the idea is to give people back the power to create their own system and achieve real efficiency gains.

The competitive read is that ZML is walking into a crowded lane. Baseten, recently valued at $13 billion, is the incumbent inference-as-a-service teams already benchmark against; Inferact comes from the creators of vLLM; RadixArk is the commercial company behind SGLang. Chip-agnostic is a real differentiator only if per-backend performance stays close to what a chip-specific runtime delivers, and the reporting does not include head-to-head benchmarks against any of them.

The honest caveat is that LLMD is not open source but launching as a free product with the goal of learning about usage, per TechCrunch's phrasing. That is a distribution choice, not a durable business, and the piece does not name customers or publish latency or throughput numbers. What the reporting also does not give you is how the team keeps pace with vendor-specific optimizations at five backends at once, or whether endorsements from Hugging Face convert into distribution or just goodwill.

For anyone running open source model inference on a budget, or holding an AMD or TPU allocation they cannot fully exploit, LLMD is worth pulling down and testing quietly. That is the audience Morin appears to be aiming at, and the one most likely to move the needle if the pitch holds up under load.