ZML's complimentary AI server operates on any leading chip.

      A startup based in Paris aims to diminish Nvidia’s dominance in AI, not by developing a new chip, but through software. ZML has launched a free tool that operates open-source models rapidly across various platforms including Nvidia, AMD, Google, Apple, and Intel.

      Though Nvidia currently leads in AI hardware, its control is becoming less secure. According to TechCrunch, ZML, a Parisian startup supported by AI pioneer Yann LeCun, has introduced free software that facilitates the operation of open-source language models on a diverse range of chipsets. This range includes five different targets: Nvidia, AMD, Google’s TPUs, Intel, and Apple.

      The tool, named ZML/LLMD, serves as an inference server. Inference refers to the process of executing a trained model to respond to inputs, which currently consumes the majority of computing resources. Founder Steeve Morin stated that the objective is to dismantle the barriers that confine users to a specific vendor and to maximize the performance of each chip.

      The importance of utilizing a variety of chips lies primarily in cost. As AI expenses increase, companies and cloud services seek the flexibility to choose less expensive or energy-efficient silicon for specific tasks. “The aim is to empower people to construct their own systems,” Morin explained. Achieving this effectively could serve as a significant advantage against Nvidia's stronghold.

      This innovation might also benefit a surge of emerging chip manufacturers, many located in Europe. Morin mentioned companies such as Axelera, Fractile, Kalray, SiPearl, and VSORA, among others. Software that recognizes their chips as primary options rather than inferior ones provides customers with valid incentives to explore them.

      The competition is fierce and expensive. Morin does not dismiss Nvidia's influence and notes that ZML maintains a positive rapport with the chip leader. However, the market is crowded. The “inference gold rush” has led to competitors like Baseten, recently valued at $13 billion, along with the teams behind open-source projects like vLLM and SGLang, all pursuing the same goal of reducing AI operating costs.

      Morin believes ZML has a broader vision. “We have reached the stage where we are co-designing silicon,” he remarked. His nimble team of 20 has quickly delivered results, with more updates on the horizon.

      The significance of LLMD lies in its current availability for free, aimed at building user engagement rather than being a paid product at this time. Its unique origin sends a broader message. A tool designed to weaken Nvidia’s hold and to bolster Europe’s AI infrastructure emerged from Paris, not Silicon Valley. Morin, who secured $20 million from investors including Xavier Niel’s Kima Ventures, stated clearly, “I couldn’t do ZML anywhere but in Paris.”