ZML gives away the software that lets AI models run on almost any chip. Now it's using that free code as its ticket into France's push to build an AI stack that doesn't depend on Nvidia.
On June 18, at VivaTech in Paris, the Île-de-France region, cloud operator Scaleway, chipmaker VSORA and the inference startup ZML announced a partnership to build a homegrown European AI stack, from silicon to cloud. Scaleway agreed to fold VSORA's Jotunn8 inference chip into its data centers. ZML agreed to make sure the software running on top of that chip doesn't care whose silicon it's talking to. That's the entire company in one sentence, and its founder said as much himself.
"The Jotunn8 is the best inference processor on the market today," ZML CEO Steeve Morin said in the joint announcement. "When the silicon is available and the infrastructure is in place, there is only one thing to do: connect the stack and run the models."
You don't need to know what Zig, MLIR or OpenXLA are to get why that matters. Every AI company today runs its models through Nvidia's CUDA software, which only works on Nvidia chips. That's fine until you want to run on a cheaper AMD card, a Google TPU or Amazon's Trainium chip instead, at which point you're rewriting your code from scratch. ZML's pitch is that you shouldn't have to. Its inference stack, built in Zig instead of the usual Python and CUDA layers, compiles a model once and runs it on Nvidia, AMD, TPU or Trainium hardware from the same codebase. The whole thing is open source and free, sitting on GitHub with more than 3,300 stars.
Morin knows something about building things people actually use. Before ZML, he spent seven years running engineering at Zenly, the friend-mapping app Snap bought in 2017. Snap shut Zenly down in February 2023 rather than sell it, worried it would end up helping a rival. Morin started ZML that same year.
The AMD case shows what the free-software pitch is worth in practice. A Radeon RX 7900 XTX card ships with 24GB of memory for under $1,000, a fraction of what a comparable Nvidia data center GPU costs, but it's spent years as a second-class citizen for AI work because almost nothing runs on it without a manual rewrite. ZML compiles the same model graph straight to AMD's ROCm stack, no separate code path required. That's the kind of detail that makes the vendor lock-in argument concrete instead of theoretical.
ZML isn't alone in chasing inference efficiency, and its rivals are raising money on a scale it isn't. Gimlet Labs, which slices AI workloads across Nvidia, AMD, Intel, Arm, Cerebras and d-Matrix chips depending on which stage of the job is compute-bound versus memory-bound, closed an $80 million Series A led by Menlo Ventures in March, bringing its total to $92 million, and claims 3x to 10x faster inference at the same cost. According to TechCrunch, Baseten is reportedly raising $1.5 billion just months after its last mega-round, and Modal Labs was in talks in February to raise at a $2.5 billion valuation. Against that backdrop, ZML giving its core product away for nothing looks less like generosity and more like a distribution strategy: get the software everywhere first, worry about a paid tier later.
Here's the thing about the VivaTech deal, though. Strip away the product talk and what's underneath is policy. The partnership landed eleven days after the European Commission published its Technological Sovereignty Package on June 3, which bundles a Chips Act 2.0 with a new Cloud and AI Development Act aimed at cutting Europe's reliance on American and Asian chip supply chains. A French region, a French cloud provider, a French chipmaker and a French inference startup lining up together right after that announcement is not a coincidence. It's ZML positioning itself as the software layer of a sovereign European AI stack, backed by regional government money instead of venture capital.
There's a catch worth saying plainly. VSORA's Jotunn8 chip isn't in production yet. The entire arrangement, the sovereignty pitch, Morin's quote, the political timing, rests on a piece of silicon that doesn't currently exist outside a lab. Scaleway and ZML are committing to integrate hardware they can't yet buy.
Also read: SambaNova Is Closing In on a $1 Billion Round That Values It Near $10 Billion • Anthropic Leases a Hudson Square Tower Even as It Rations Its Own AI Model • China Will Debut the World's First AI Agent Phone at Its Shanghai AI Summit
