The ZML LLMD inference server launched this week as a free product, promising to run open-source large language models across a range of chips, from Nvidia and AMD to Google’s TPU, Apple Metal, and Intel Arc, without forcing enterprises to commit to any single vendor’s hardware stack.
ZML, the Paris-based AI startup backed by Turing Award winner Yann LeCun, wants to make chip-agnostic inference a practical reality rather than a conference-circuit talking point. That’s a reasonable ambition. Whether a team of 20 can deliver it at scale is the part worth watching.
What the ZML LLMD Inference Server Actually Does
Founder Steeve Morin frames the problem clearly enough: software and architecture barriers have created vendor lock-in in AI inference, leaving enterprises stuck with whatever chip ecosystem they happened to bet on first. The ZML LLMD inference server is designed to break those silos, running a variety of open-source models on whichever silicon is available at the best price or energy cost.
‘The idea is to give people back the power to create their own system and achieve real efficiency gains that allow [AI] to be disseminated,’ Morin said.
The product is free for now. Morin’s stated reason is measured: ‘I’d rather measure and [then generate revenue] where it is most effective without hindering my growth stupidly because I have been too greedy from the get-go.’ A sensible position, though the monetisation timeline remains deliberately unspecified.
ZML/LLMD differs from the company’s earlier work: its first public project, an inference-focused ML framework released in 2024 and updated in March, is open source. LLMD is not, which is a meaningful distinction for enterprises evaluating supplier risk.
Morin also sees LLMD as a tailwind for emerging AI chipmakers, citing eight European silicon startups: Axelera, Fractile, Kalray, OLIX, Q.ANT, SiPearl, SpiNNcloud, and VSORA. His interest is less about geography than about reaching hardware ‘things that haven’t been done before anywhere in the world,’ he said. He is equally clear that ZML maintains a good relationship with Nvidia, whose existing supply dominance makes it an unavoidable partner rather than a target.
A Crowded Race, and Why Chip Agnosticism Matters
ZML enters a field that is attracting serious capital. Baseten, founded in 2019 and one of the more prominent inference competitors, is reportedly finalising a $1.5 billion funding round, though the structure is worth unpacking. According to the Wall Street Journal, as reported by Yahoo Finance, the round is split-priced: some investors are coming in at a $13 billion valuation and others at $11 billion. The headline $13 billion figure in circulation is, in other words, the top of a range.
Morin’s other named competitors include Inferact, from the creators of the open-source vLLM project, and RadixArk, the commercial entity behind SGLang. Both overlap with LLMD on inference serving, though Morin’s stated ambitions reach further: ZML is already ‘co-designing silicon’ with chip partners, he said.
ZML raised $20 million from a roster of venture firms including 20VC, >commit, AALVC, Drysdale Ventures, Kima Ventures, Kindred Capital, LocalGlobe, and Puzzle Ventures. The cap table also includes founders from Dagger, Docker, and Hugging Face, alongside LeCun, now at AMI Labs.
Morin’s credibility with that investor group traces partly to his exit track record. He served as VP of Engineering at Zenly for seven years, scaling the social location app to millions of users before Snapchat acquired it in 2017. Snap’s SEC filings, as reported by Business Insider, put the cash consideration at $213.3 million, comprising $186.8 million paid to sellers and $9.3 million of liabilities due to sellers, with additional retention and milestone bonuses potentially pushing the total higher. The snippet’s ‘nine figures’ description, while technically accurate, understates it by some margin.
The lean team structure is deliberate. Morin credits 20 people as the reason ZML can move quickly, with more releases planned. Whether that agility can survive the complexity of true multi-chip optimisation at enterprise scale is a different question, one the market will answer once adoption data starts coming in.
Morin, for his part, does not plan to leave Paris to find out: ‘I couldn’t do ZML anywhere but in Paris,’ he said. More releases are scheduled, and the moment ZML starts talking about converting free users to paying customers will be the real test of whether chip agnosticism is a product or a pitch.
