ZML released LLMD, a free LLM inference server that runs on Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc. The Paris-based startup, endorsed by Yann LeCun, aims to break vendor lock-in and reduce AI inference costs.
Key facts
- ZML raised $20 million from 20VC, Kima Ventures, LocalGlobe.
- ZML's team has 20 people; Morin was Zenly VP of engineering.
- Supports Nvidia, AMD, Google TPU, Apple Metal, Intel Arc.
- Baseten recently valued at $13 billion; vLLM and SGLang compete.
- Google booked Intel to package 3 million TPUs by 2028.
ZML, a Paris-based AI startup backed by Turing Award winner Yann LeCun, today released ZML/LLMD, a free inference server for large language models that supports Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc hardware According to TechCrunch. The company claims the software can run models at their maximum available speed, and sometimes faster, across a heterogeneous hardware stack.
The inference gold rush gets a multi-chip weapon
Inference has become the dominant cost in AI deployment, outpacing training in enterprise spend. ZML's pitch is straightforward: give enterprises and cloud providers the ability to mix cheaper or more energy-efficient chips without sacrificing performance. ZML founder Steeve Morin told TechCrunch the goal is to "give people back the power to create their own system and achieve real efficiency gains."
The move comes as Nvidia's market dominance faces growing challenges. While Nvidia's H100 and Blackwell GPUs remain the default choice, AMD's MI300X, Google's TPU v5p, and Intel's Arc series are increasingly viable alternatives. ZML's LLMD server abstracts the software layer, allowing a single deployment to span multiple chip architectures.
Competition and context
ZML enters a crowded field. Baseten recently hit a $13 billion valuation, Inferact emerged from the vLLM project, and RadixArk commercializes SGLang. Both vLLM and SGLang partially compete with LLMD, but Morin argues ZML covers a broader spectrum, including co-designing silicon with chipmakers. He cited European AI chip startups Axelera, Fractile, Kalray, OLIX, Q.ANT, SiPearl, SpiNNcloud, and VSORA as potential collaborators.
Morin's track record helps. He was VP of engineering at Zenly, which Snapchat acquired for nine figures in 2017. ZML's lean team of 20 people raised $20 million from investors including Harry Stebbings' 20VC, Kima Ventures, LocalGlobe, and Puzzle Ventures.
What this means for the market
The key takeaway is not that ZML will topple Nvidia — Morin explicitly says he's not bearish on the chip giant, noting ZML's good relationship with Nvidia. Rather, LLMD represents a growing trend: inference software is becoming hardware-agnostic, which commoditizes the chip layer. If enterprises can freely switch between Nvidia, AMD, Google TPU, and Intel Arc, the pricing power of any single vendor diminishes.
This is particularly relevant as Google booked Intel to package 3 million TPUs by 2028 [per recent history], and Nvidia's next-gen AI rack system was delayed to 2028 due to manufacturing snags. The window for alternative chips is widening.
ZML has not disclosed LLMD's benchmark performance on specific models or chips, but the company plans to release more details in the coming weeks.
What to watch
Watch for ZML's next release: Morin hinted at co-designing silicon with European AI chip startups. Also track whether enterprise adoption of LLMD leads to measurable shifts in GPU procurement patterns away from Nvidia, particularly in cloud deployments.
Source: techcrunch.com
