Hot French startup ZML releases free product to speed inference across lots of AI chips

ZML, the Paris-based startup endorsed by Turing Award winner Yann LeCun, has shipped ZML/LLMD — a proprietary inference server that runs open-source LLMs across Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc silicon. It's free today. Monetization is "later." Founder Steeve Morin's pitch: let enterprises mix chips, cut vendor lock-in, and stop overpaying for inference capacity. That's a direct assault on the "inference gold rush" economics shaping the AI infrastructure stack right now.

The cap table

Morin's track record is the load-bearing wall here. As VP of engineering at Zenly, he steered the company to a nine-figure Snapchat acquisition in 2017. That pedigree unlocked $20M from Harry Stebbings' 20VC, >commit, AALVC, Drysdale Ventures, Xavier Niel's Kima Ventures, Kindred Capital, LocalGlobe, and Puzzle Ventures — a dense syndicate for a 20-person shop with hardware co-design ambitions.

The product itself isn't open source — a deliberate pivot from ZML's earlier inference-focused ML framework (released 2024, updated March). LLMD is closed but free. Morin told TechCrunch he'd rather "measure and generate revenue where it is most effective" than "hinder growth stupidly" by charging too early. Translation: burn cash on adoption, figure out pricing later. Classic zero-to-one playbook.

The competitive set is deep. Baseten recently commanded a $13B valuation on adjacent territory. Inferact emerged from the vLLM open-source community. RadixArk backs SGLang. All partially overlap. Morin's differentiation: "co-designing silicon" and a broader ambition that spans the entire inference stack, not just serving layers.

Who actually benefits

The software pitch — peak inference performance across heterogeneous chips — reads like a hedge against Nvidia supply risk and pricing power. Morin names European chipmakers including Axelera, Fractile, Kalray, OLIX, Q.ANT, SiPearl, SpiNNcloud, and VSORA as potential beneficiaries. That's a politically convenient narrative for a Paris startup backed by French and UK venture capital. Whether any of those players ship silicon at production scale remains the operative question.

For enterprises running AI at the edge — think appliance manufacturers optimizing energy consumption through intelligent firmware — the promise of vendor-agnostic inference is tangible. Less dependency on a single chip vendor means more negotiating leverage and, potentially, lower unit economics on inference. Morin claims ZML is already "co-designing silicon," which, if true, puts the company further up the value chain than most inference middleware plays.

The reality check

Free product. No disclosed revenue. No published benchmarks outside ZML's own claims. The "inference gold rush" framing is apt — and it's crowded. ZML has a credible founder, credible investors, and a technically ambitious roadmap. What it doesn't yet have is proof that LLMD delivers measurable, reproducible speed advantages across chip families in production environments.

Morin says more releases are planned. The clock on converting free users into paying customers is already running. In a market where Baseten fetches a $13B valuation on the same thesis, ZML's $20M war chest needs to buy not just runway — but traction. Lean teams ship fast. Revenue follows, or it doesn't.