ZML targets Nvidia's grip with free cross-chip AI software

Published: 2026-07-09T06:52:43.000Z

A Paris startup has decided that the way to loosen Nvidia's grip on artificial intelligence is not to build a rival chip but to make everyone else's chips work properly. ZML has released a free inference server called ZML/LLMD that runs open-source large language models across Nvidia, AMD, Google's TPUs, Apple Metal and Intel Arc silicon, all from a single codebase and without rewriting anything. The ambition, as founder Steeve Morin puts it, is to break the silos that keep AI workloads shackled to one vendor, and to squeeze maximum speed out of whatever hardware a company happens to own. The timing is deliberate, since inference has quietly overtaken training as the industry's real cost problem. Training a model happens once, while running it to answer prompts happens millions of times a day, and that is where the electricity bills and cloud invoices pile up. The software barriers around this stage have been the hidden mechanism of lock-in, and Morin offers a concrete illustration. An AMD Radeon card ships with 24 gigabytes of memory for under a thousand dollars, a fraction of a comparable Nvidia data center GPU, yet it has spent years treated as a second-class option because almost nothing runs on it without manual rewriting. ZML compiles the same model graph straight to AMD's stack with no separate code path. Giving the product away looks less like generosity than distribution strategy, and Morin is candid that he would rather measure usage first and monetize later than throttle growth by being greedy too early. The tool is still a technical preview rather than production-ready, and the field is crowded with better-funded rivals including Baseten, recently valued around 13 billion dollars, alongside the teams behind vLLM and SGLang. What sets ZML apart is a lean team of 20 and a claim that it has moved beyond software into co-designing silicon, work reinforced by a French sovereign-stack partnership with Scaleway and the chipmaker VSORA, and by backers ranging from Yann LeCun to Hugging Face's founders. Morin insists he could not have built the company anywhere but Paris. The Gulf has an obvious stake in this argument. Saudi Arabia and the UAE are pouring billions into sovereign compute through HUMAIN and G42, and both have found themselves navigating American export controls that shape which chips they can buy and deploy. Software that lets a data center mix Nvidia, AMD and emerging processors without rewriting its models is precisely the kind of insurance policy a region hedging between technology spheres would want. With regional operators chasing cheap, energy-efficient inference to serve Arabic-first models, ZML's chip-agnostic pitch travels rather well from Paris to Riyadh.