Hot French startup ZML releases free product to speed inference across lots of AI chips

What has ZML actually released?

ZML, a Paris startup founded by ex-Zenly engineering lead Steeve Morin, has released LLMD: a free, open-source tool that lets one AI model run efficiently across different types of chip, including Nvidia, AMD, Google TPU, Apple Metal and Intel Arc, instead of being locked to a single vendor's hardware.

The company has raised $20 million, led by 20VC, with backing from Yann LeCun and Hugging Face co-founders Clément Delangue and Julien Chaumond. LLMD is written in Zig rather than the usual Python and CUDA combination, which means a model can be compiled once and deployed across multiple chip types from the same codebase.

Why should a UK small business owner care about an AI chip tool they'll never install?

You won't touch LLMD directly, and you don't need to. What matters is the direction it points in: AI inference is getting cheaper and less tied to expensive specialist hardware. That trend eventually shows up in the tools you already use, not in a server you manage yourself.

Every AI feature you interact with, a chatbot, a booking assistant, an AI-generated summary in Google, runs on inference happening somewhere on someone's chips. When that gets cheaper to run, the businesses that build on top of it (your booking software, your CRM, your website platform) can offer more AI features for less, or the same features at lower cost. That's the practical chain reaction worth watching.

Is this something my business should adopt right now?

No, not directly. LLMD is infrastructure for developers running AI models at scale, not a product for a salon, plumber or restaurant to install. Trying to use it yourself would be like a café owner trying to use a coffee roaster's industrial machinery instead of just buying the beans.

What's actually relevant to you is timing. UK government research (DSIT) shows only 16% of UK businesses have deliberately deployed a recognised AI technology, dropping to just 14% for micro businesses, versus 36% for large firms. Tools like LLMD are part of why that gap should narrow over the next year or two: cheaper inference means the AI features bundled into ordinary small business software get better without the price going up.

What's actually stopping small businesses from using AI, if not the technology?

Cost and skills, not the underlying technology. The same DSIT research found 76% of UK businesses rate cost as a major barrier to AI adoption, and 60% cite a lack of in-house AI skills. Developments like LLMD chip away at the cost side of that equation, but they do nothing for the skills side.

That's the gap Braynex Services exists to close. We don't ask small business owners to understand inference servers or chip architecture. We set up the practical layer on top: a Google Business Profile that actually converts, a booking system you own, follow-up automation that catches leads you're currently losing. The 65% of UK AI-adopting businesses who cite efficiency and productivity as their main motivation, per the same DSIT data, are proving the point: you don't need to understand the engine to benefit from the car running better.

What should I actually do with this news?

• Don't chase the technology. Watch for AI features appearing in tools you already pay for (booking systems, website builders, review management) and ask whether they're now included free or at lower cost.
• Prioritise owning your data and platform before adding AI on top. An AI feature bolted onto a rented platform like Linktree or Fresha still leaves you exposed if that platform changes its terms.
• Ask any software supplier directly: "is this AI feature actually included, or an extra charge?" Falling inference costs mean the honest answer should increasingly be "included."

Braynex Services' view

Cheaper AI inference is good news, but it solves the wrong problem for most small businesses we work with. The businesses losing money aren't losing it because AI is expensive to run. They're losing it because they don't own their booking system, their website is a PDF, or their Google Business Profile was never verified.

We saw this with a nail salon paying around £1,800 a month in commission to Fresha before moving to its own booking system at £35 a month, saving roughly £21,000 a year. No AI chip breakthrough was needed there, just ownership of the infrastructure. Developments like LLMD will make the AI layer on top cheaper over time. Get the foundations right first and you'll be ready to benefit when it does.

If you're not sure whether your current setup is costing you customers or money, book a free audit at braynexservices.com and we'll show you exactly where.