ZML raises €17.5M to run LLMs on any chip — and break Nvidia's inference grip

ZML, a Paris-based AI infrastructure company, has raised €17.5 million (approximately $20 million) in a round backed by 20VC, Kima Ventures, Kindred Ventures, Puzzle Ventures, and AAL Management. No lead investor is named in the disclosed terms. ZML's founder is Steeve Morin.

What LLMD does

ZML's core product is LLMD, an open-source inference server for large language models. Its design claim is specific: LLMD can run the same LLM workload across Nvidia CUDA, AMD ROCm, Google TPU, Apple Metal, and Intel Arc hardware without requiring the enterprise to rewrite the inference stack for each target platform.

That is a more pointed claim than it might appear. Most production LLM deployments run on Nvidia infrastructure not because enterprises evaluated alternatives and chose Nvidia, but because the surrounding toolchain — cuDNN, TensorRT, NCCL, and the frameworks that depend on them — is built CUDA-native. Multi-chip inference has been technically feasible for years; the barrier is the switching cost that CUDA compatibility creates at every layer of the stack.

LLMD's pitch is an abstraction layer that sits above the hardware-specific runtime, presenting a single interface while routing to whatever accelerator is available or cheapest. For an enterprise running a mix of on-premise Nvidia, cloud AMD, and edge Apple Silicon, the value proposition is a unified deployment model rather than three separate inference configurations.

The lock-in market

Enterprise AI infrastructure is currently a two-step procurement decision: first buy the model, then buy the chips and runtime it runs on. Nvidia's CUDA lock-in effectively merges those steps — choosing a model that performs well in benchmarks usually means choosing Nvidia hardware to run it. That merger is commercially valuable to Nvidia and inconvenient for enterprises that want chip procurement to respond to price and availability signals rather than compatibility constraints.

ZML is not the only attempt to break that lock: vLLM, llama.cpp, and Ollama are all widely adopted open-source inference frameworks with multi-hardware support. What distinguishes LLMD, per the thesis, is architecture designed from the start for production-grade multi-chip scheduling rather than retrofitted from single-GPU origins. Whether that architectural distinction is real and durable is the product question the round does not answer.

The 18-month question

The €17.5 million funds LLMD's development and distribution. The open-source model means the commercial test is distribution, not licensing: how many enterprises adopt LLMD as their default inference layer, and whether any convert that adoption into a paid support or hosted layer.

The risk runs in both directions. Nvidia continues to deepen its software moat — recent CUDA updates have made migration harder, not easier. And the open-source inference landscape is crowded enough that a technically sound product still needs developer community and enterprise adoption rates to become the default. ZML has a credible investor set with cross-Atlantic reach. The test is whether that translates into production deployments at named enterprise customers within the next 18 months.