ZML, a Paris-based AI infrastructure company, has built a production-grade inference stack that decouples AI workloads from proprietary hardware, allowing any model to run on multiple accelerators (NVIDIA, AMD, TPU, Trainium and more) from a single codebase while delivering peak hardware performance. The company compiles models directly to the hardware using Zig and MLIR, rejecting hidden state, magic abstractions, and Python-heavy runtimes. ZML recently released ZML/LLMD, a powerful LLM inference server, and has gained public support from Turing Award winner Yann LeCun. The stack is purpose-built for production environments where performance and hardware flexibility matter most. The overall verdict is 4.5/5 stars, with the company described as one of the most ambitious and technically rigorous approaches to AI inference in 2026.