Nvidia’s grip on the AI market is loosening as new options emerge from every angle. ZML, a Paris-based startup backed by Turing Award winner Yann LeCun, has launched software allowing open-source large language models to run on Nvidia, AMD, Google TPUs, Apple Metal, and Intel Arc chips.
With ZML/LLMD, a newly released LLM inference server, the company aims to remove existing barriers and ensure different hardware runs at maximum speed. Founder Steeve Morin told TechCrunch that the goal is to make various chips available for AI use cases at their fastest possible rate, often outperforming current limits.
As artificial intelligence integrates into daily work and life, optimising inference has become more critical than model training. However, the process remains patchy due to software and architecture barriers that force vendor lock-in, Morin said.
While achieving peak performance across diverse hardware is a technical feat, it could disrupt the market amidst rising costs associated with AI. ZML hopes to offer enterprises and cloud providers the flexibility to mix chips, selecting those that are cheaper or consume less energy.
“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.
This approach could support new chipmakers, many based in Europe. Morin cited Axelera, Fractile, Kalray, OLIX, Q.ANT, SiPearl, SpiNNcloud, and VSORA. What matters most to him is that ZML can work with them on projects not attempted anywhere else globally.
Morin is not opposed to Nvidia. His stance reflects the reality of Nvidia’s existing supply chain. He noted a good relationship with the chip giant, which is preparing for the surge in inference demand.
Investment in inference has been intense, described as an “inference gold rush.” ZML faces competition from Baseten, recently valued at $13 billion; Inferact from the creators of vLLM; and RadixArk, the commercial entity behind SGLang.
Both vLLM and SGLang partially overlap with LLMD, but Morin’s ambitions for ZML cover a wider scope. “We have reached the point where we are co-designing silicon,” he said. He credited his lean team of 20 people for enabling rapid movement and planned future releases.
That small team is well funded. Morin raised $20 million from venture firms including 20VC, >commit, AALVC, Drysdale Ventures, Kima Ventures, Kindred Capital, LocalGlobe, and Puzzle Ventures. His background as VP of engineering at Zenly, which Snapchat acquired for nine figures in 2017, helped secure this backing.
Unlike ZML’s first public project, an inference-focused ML framework released in 2024 and updated in March, ZML/LLMD is not open source. It launches as a free product to gather usage data. “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,” Morin said.
It is too early to determine when ZML/LLMD might become paid or how widely it will be adopted. However, the startup’s cap table shows interest from other founders, including Solomon Hykes from Dagger and Docker, Clément Delangue and Julien Chaumond from Hugging Face, and LeCun, now at AMI Labs.
This support strengthens the argument that European AI startups can build from home. “I couldn’t do ZML anywhere but in Paris,” Morin said.
