ZML raises $20M to cut AI inference costs across rival chips
Paris startup ZML has launched a free inference server that runs across Nvidia, AMD, Google and Apple chips, and raised $20M to break the industry's costly hardware lock-in.
The News
Paris-based startup ZML has released LLMD, a free inference server built to run large language models across competing brands of AI chips, and disclosed that it has raised 20 million dollars to fund the effort. The company was founded by Steeve Morin, previously vice-president of engineering at Zenly, the location-sharing app Snapchat bought for a nine-figure sum in 2017.
LLMD is designed to squeeze faster, cheaper inference out of whatever silicon a customer already owns. It supports Nvidia, AMD, Google TPU, Apple Metal and Intel Arc hardware, an unusually broad spread for a single serving layer. The product is free at launch but not open source, a deliberate choice Morin says lets the 20-person team measure real adoption before settling on a business model.
The raise drew an eye-catching cap table. Backers and angels include Yann LeCun, the Turing Award winner now at AMI Labs, Docker and Dagger founder Solomon Hykes, and Hugging Face co-founders Clement Delangue and Julien Chaumond. Institutional cheques came from 20VC, Kima Ventures, Kindred Capital, LocalGlobe and others.
Why It Matters
The economics of running a model, not training one, is where most companies now feel the pain. Inference is the recurring bill that never stops, and it is overwhelmingly paid to a single vendor. By making one serving layer speak fluently to Nvidia, AMD, Google and Apple chips alike, ZML is selling optionality: the freedom to route workloads to whatever hardware is cheapest or merely available.
That pitch lands in a crowded field. Open-source projects such as vLLM and SGLang already command developer mindshare, and well-funded rivals like Baseten carry valuations reported around 13 billion dollars. The last time infrastructure this low in the stack attracted this calibre of angel, it was the container wars of the mid-2010s, when a thin abstraction layer over messy hardware reshaped how software shipped. ZML is betting inference is having its Docker moment, and it has recruited Docker's own founder to make the point.
Indian Angle
For Indian AI teams, chip-agnostic inference is not a luxury feature but a survival strategy. High-end Nvidia accelerators are scarce and dollar-priced, which punishes rupee-earning startups twice over: once on availability, once on the exchange rate. A serving layer that runs equally well on AMD or Google TPU capacity widens the pool of usable, and affordable, hardware.
This matters directly for the IndiaAI Mission, whose subsidised compute programme has been assembling a mixed fleet of GPUs from multiple vendors rather than a uniform Nvidia estate. Software that treats that heterogeneity as normal, instead of a problem to engineer around, lowers the barrier for firms drawing on that shared pool. Model builders such as Sarvam and Krutrim, both wrestling with inference costs at scale, stand to gain from any tool that decouples their software from one supplier.
There is a talent dimension too. Indian engineers dominate the ranks of global inference and systems teams, and open, hackable tooling of this kind tends to spread fastest through exactly those communities. If LLMD gains traction, expect Bengaluru and Hyderabad infrastructure teams to be among its earliest power users.
FAQ
Is LLMD open source?
No. ZML has released LLMD free of charge but has not open-sourced the code. The company says the free tier is a way to gauge adoption before deciding how to monetise, distinguishing it from community projects like vLLM and SGLang.
Which chips does it support?
At launch LLMD runs on Nvidia, AMD, Google TPU, Apple Metal and Intel Arc, an unusually wide range for one inference server. That breadth is the core selling point for buyers wanting to avoid single-vendor lock-in.
How much did ZML raise?
The company disclosed 20 million dollars in funding, backed by investors including 20VC, Kima Ventures and Kindred Capital, plus angels such as Yann LeCun and Solomon Hykes. ZML currently employs 20 people.
Where can I read the original announcement?
TechCrunch reported the launch and funding. The source link appears at the end of this article.
This story was reported by TechCrunch. Read the full original coverage at TechCrunch.
Sources & Citations
Frequently Asked Questions
Is LLMD open source?
No. ZML has released LLMD free of charge but has not open-sourced the code. The company says the free tier is a way to gauge adoption before deciding how to monetise, distinguishing it from community projects like vLLM and SGLang.
Which chips does it support?
At launch LLMD runs on Nvidia, AMD, Google TPU, Apple Metal and Intel Arc, an unusually wide range for one inference server. That breadth is the core selling point for buyers wanting to avoid single-vendor lock-in.
How much did ZML raise?
The company disclosed 20 million dollars in funding, backed by investors including 20VC, Kima Ventures and Kindred Capital, plus angels such as Yann LeCun and Solomon Hykes. ZML currently employs 20 people.
Where can I read the original announcement?
TechCrunch reported the launch and funding, with details on the product, backers and the founder's background. The full original coverage is linked at the end of this article.