A French startup called ZML has released software that lets artificial intelligence models run efficiently on multiple different computer chips, breaking what has long been a major bottleneck in the AI industry. The new tool, called ZML/LLMD, is designed to eliminate the vendor lock-in that forces companies to stick with one chip manufacturer, even when other options might be cheaper or more energy-efficient.

Why Does It Matter That AI Can Run on Different Chips?

For years, Nvidia has dominated the market for AI chips because its hardware and software work seamlessly together. But as artificial intelligence becomes more central to business operations, the cost of relying on a single supplier has become a real problem. Companies want flexibility; they want to mix and match chips based on what makes sense for their budget and power consumption.

ZML/LLMD solves this by allowing large language models, or LLMs (the AI systems that power chatbots and text generation), to run at peak performance across chips from multiple manufacturers. This includes Nvidia's GPUs, AMD's processors, Google's TPUs (specialized AI chips), Apple's Metal framework, and Intel's Arc graphics cards.

"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," said Steeve Morin, founder of ZML.

Steeve Morin, Founder at ZML

Morin founded ZML after serving as VP of engineering at Zenly, a location app that Snapchat acquired for nine figures in 2017. His track record helped him raise $20 million from venture firms including 20VC, Kima Ventures, and Kindred Capital.

How Can Enterprises Use This New Flexibility?

• Cost Optimization: Companies can now choose cheaper chips for certain workloads instead of being locked into premium Nvidia hardware for everything, potentially reducing infrastructure spending significantly.
• Energy Efficiency: Different chips consume different amounts of power; enterprises can select hardware that minimizes electricity costs while maintaining performance for their specific AI applications.
• Vendor Independence: Organizations are no longer forced to rely on a single supplier's roadmap, pricing, or availability, giving them negotiating power and supply chain resilience.
• Support for European Chipmakers: The software enables emerging AI chip companies like Axelera, Fractile, Kalray, OLIX, Q.ANT, SiPearl, SpiNNcloud, and VSORA to compete in the market without building their own software ecosystems from scratch.

The timing is significant because inference, the process of running a trained AI model to generate responses, has become more important than training the models in the first place. This shift has sparked what industry observers call an "inference gold rush," with multiple startups competing to optimize how AI systems process prompts.

ZML faces competition from other inference optimization companies like Baseten, valued at $13 billion; Inferact, created by the team behind the open source project vLLM; and RadixArk, the commercial entity behind SGLang. However, Morin's ambitions extend beyond just optimizing existing chips; he says ZML is now "co-designing silicon" with hardware makers, meaning the company works directly with chipmakers to ensure their hardware performs optimally with ZML's software.

Unlike ZML's first public project, an inference-focused machine learning framework released in 2024, ZML/LLMD is not open source. Instead, it launched as a free product so the company can learn how it's being used before deciding on a pricing model. Morin explained that he prefers to measure adoption and understand where the software creates the most value before monetizing it.

The startup's investor list reveals how seriously the AI community is taking this problem. Backers include Yann LeCun, a Turing Award winner now with AMI Labs; Solomon Hykes, founder of Docker and Dagger; and Clément Delangue and Julien Chaumond from Hugging Face, a major open source AI platform. This level of founder and researcher participation suggests that breaking vendor lock-in is seen as a critical infrastructure challenge.

Morin also made a point about where this innovation is happening. "I couldn't do ZML anywhere but in Paris," he said, highlighting how Europe's AI startup ecosystem is maturing enough to tackle infrastructure-level problems that were previously dominated by Silicon Valley. The company operates with a lean team of just 20 people, which Morin credits with enabling rapid development and multiple planned releases in the coming months.