The era of Nvidia’s dominant position in AI hardware is facing new competition, as French startup ZML launches software designed to optimize AI performance across multiple chip architectures.
Backed by Turing Award winner Yann LeCun, ZML has introduced ZML/LLMD, an LLM inference server that enables open-source large language models to run efficiently across a range of hardware platforms, including Nvidia GPUs, AMD processors, Google’s TPU, Apple Metal, and Intel Arc graphics cards.
Founder Steeve Morin told TechCrunch that the software aims to eliminate existing technical silos and maximize the performance potential of diverse chip architectures, sometimes exceeding their typical limits.
As AI systems become increasingly integrated into enterprise and consumer applications, inference optimization—the processing of prompts—has become more critical than model training. However, current solutions often suffer from software and architectural barriers that create vendor lock-in situations, according to Morin.
ZML’s cross-chip compatibility solution represents both a technical achievement and potential market disruption, particularly as AI infrastructure costs continue to rise. The platform allows enterprises and cloud providers to mix different chip types, potentially using less expensive or energy-efficient alternatives.
“We’re giving users the power to build their own systems and achieve real efficiency gains that enable broader AI adoption,” Morin explained.
The software could benefit emerging AI chip manufacturers, many of which are European companies including Axelera, Fractile, Kalray, OLIX, Q.ANT, SiPearl, SpiNNcloud, and VSORA. Morin emphasized that success depends on enabling innovations that haven’t been accomplished elsewhere.
Despite the competitive landscape, Morin remains optimistic about Nvidia’s position, noting ZML’s strong working relationship with the company as it expands focus on inference workloads.
The inference optimization market has attracted significant investment, earning the nickname “inference gold rush.” ZML faces competition from companies like Baseten (valued at $13 billion), Inferact (from vLLM creators), and RadixArk (behind SGLang).
ZML’s approach extends beyond existing solutions through what Morin describes as co-designing silicon alongside software. The 20-person Paris team has already delivered multiple releases and has more planned.
The startup’s $20 million funding round, led by investors including Harry Stebbings’ 20VC, was facilitated by Morin’s previous success selling his Zenly location-sharing app to Snapchat for a nine-figure sum in 2017.
While ZML/LLMD launches as a free product to gather usage data, it is not open source. This freemium approach allows ZML to refine its offering before potential monetization. “I’d rather optimize where revenue generation makes sense without constraining growth due to premature monetization,” Morin stated.
Though commercialization timelines remain undefined, ZML has attracted notable industry figures to its cap table, including Dagger and Docker co-founder Solomon Hykes, Hugging Face co-founders Clément Delangue and Julien Chaumond, and Yann LeCun of AMI Labs. The backing reinforces growing evidence that Europe’s AI ecosystem can successfully develop cutting-edge technology domestically.
“I couldn’t build ZML anywhere other than Paris,” Morin said.
