The AI compute crisis just got a highly practical workaround. Paris-based startup ZML is handing developers a completely free, open-source software stack designed to dramatically accelerate AI model inference across a huge variety of hardware accelerators. And they built it using Zig, MLIR, and Bazel.
Running reasoning models at scale has become a crippling operational bottleneck for almost everyone in the industry. But ZML, founded by former Zenly VP of Engineering Steeve Morin, is compiling AI models right down to the metal. This completely decouples complex AI workloads from proprietary ecosystems. You can now take a major model and run it exactly where you want.
Compiling Llama 3 for Any Silicon
So how does this actually work? The framework currently supports heavy-hitters like Llama 3.x and Qwen. Instead of being trapped on one manufacturer’s hardware, developers can execute these models on AMD RoCM, Google TPUs, AWS Trainium, and Intel OneAPI using just a single codebase. It is a brilliant technical achievement, speeding up AI inference without locking organizations into a single silicon vendor.
Europe is aggressively pushing for technological sovereignty right now. France alone is pumping billions into its Électronique 2030 program to rebuild native semiconductor infrastructure. We are already seeing the practical fallout. European cloud provider Scaleway recently launched a sovereign AI inference ecosystem based on VSORA’s Jotunn8 processor. They are utilizing ZML’s software layer to make it run efficiently.
How ZML’s Stack Triggers the Zero Buy-In Era
This release mirrors the early open-source operating system wars. ZML is essentially trying to become the Linux for AI chips by eliminating the vendor lock-in that historically allowed hardware manufacturers to dictate software innovation.
It accelerates an industry shift known as the “Zero Buy-In Strategy.” Organizations can now dynamically migrate workloads from Nvidia to AMD or AWS Trainium based entirely on real-time cost and availability. They do not have to rewrite their software architecture to pull it off. That freedom gives smaller teams incredible negotiating power when managing cloud computing costs.