ZML's complimentary AI server operates on all major chipsets.

      A startup based in Paris aims to challenge Nvidia’s dominance in AI, not with new hardware but through software. ZML has introduced a free tool that efficiently runs open-source models across various hardware platforms, including Nvidia, AMD, Google, Apple, and Intel.

      While Nvidia continues to lead in AI hardware, competition is increasing. ZML, supported by AI pioneer Yann LeCun, has developed free software capable of running open-source language models on a variety of processors, according to a report from TechCrunch. The targeted hardware includes Nvidia, AMD, Google’s TPUs, Intel, and Apple.

      The tool, named ZML/LLMD, functions as an inference server. Inference involves executing a trained model to generate responses, which now constitutes the majority of AI’s computing demands. Founder Steeve Morin aims to dismantle the barriers that tie users to a single vendor, ensuring that each chip can operate at maximum efficiency.

      The importance of using diverse chips lies in cost considerations. As AI expenditures rise, businesses and cloud providers seek the flexibility to choose more affordable or energy-efficient chips for specific tasks. “The idea is to empower people to create their own systems,” Morin explained. Successfully accomplishing this could serve as a disruptive force against Nvidia’s dominance.

      This approach may also provide opportunities for emerging chip manufacturers, particularly those in Europe. Morin mentioned companies like Axelera, Fractile, Kalray, SiPearl, and VSORA, claiming that software recognizing their chips as primary options rather than secondary ones incentivizes consumers to explore these alternatives.

      The competitive landscape remains intense and financially demanding. Morin acknowledges Nvidia's significance and asserts that ZML maintains a positive relationship with the industry leader. However, the environment is bustling, characterized by the “inference gold rush” which has seen competitors such as Baseten, recently valued at $13 billion, and teams working on open-source projects like vLLM and SGLang, all vying for the same goal: reducing AI operational costs.

      Morin believes ZML has broader ambitions. “We have reached the point where we are co-designing silicon,” he stated. His small but efficient team of 20 has released products swiftly, with more scheduled to follow.

      Currently, LLMD is available for free in order to increase its user base and is not yet a paid offering. The tool’s unique origin sends a significant message: a development designed to mitigate Nvidia’s stronghold and support Europe’s AI infrastructure has emerged from Paris rather than Silicon Valley. Morin, who secured $20 million in funding from investors like Xavier Niel’s Kima Ventures, expressed it clearly: “I couldn’t do ZML anywhere but in Paris.”