Run AI Models on Any Chip Free — No NVIDIA Lock-In | NCR

If you’ve been building with AI models and assumed you’re stuck with NVIDIA GPUs forever, a 20-person Parisian startup just changed that. ZML — endorsed by Turing Award winner Yann LeCun and backed by founders from Docker and Hugging Face — just dropped LLMD, a free inference server that runs open-source AI models at peak speed across NVIDIA, AMD, Google TPU, Apple Metal, and Intel Arc. No vendor lock-in, no licensing fees.

Here’s the citable version: LLMD is a free inference server from ZML, a 20-person Paris company that raised $20 million and whose founder (a former Zenly VP of Engineering, where Snapchat acquired the company for nine figures) says the server hits peak performance on NVIDIA, AMD, Google TPU, Apple Metal, and Intel Arc — sometimes faster than each chip’s native stack. Yann LeCun has endorsed the project.

Inference costs are quietly becoming the biggest expense for anyone running AI tools in production. We’ve covered free AI tools before — like how Google I/O 2026 dropped free AI tools for beginners — but those are consumer-facing. If you’re building anything that processes prompts repeatedly, the inference bill adds up fast. And if you’re locked into one hardware vendor, you’re paying whatever they decide to charge. The other free tools from startups we’ve covered suggest a pattern: serious infrastructure, released free, monetized later.

What is LLMD, in plain language?

LLMD is a free inference server from ZML that runs the same AI model on NVIDIA, AMD, Google TPU, Apple Metal, or Intel Arc chips without you rewriting anything. Inference is the step where a trained model processes your prompt and generates a response — every time you type into ChatGPT, Claude, or any LLM-powered app, that’s inference happening.

Most inference runs on NVIDIA GPUs because the software ecosystem, CUDA, is deeply entrenched. If you’re a company running AI models at scale, you’re probably on NVIDIA hardware, and switching means rewriting your entire software stack. LLMD breaks that: it translates model operations into whatever hardware you give it.

The claim that got my attention isn’t just “it works on different chips.” It works at full speed on different chips. ZML’s founder Morin told TechCrunch they’re hitting peak performance across hardware, and sometimes going faster than the native stack. Bold claim. But the credentials check out — Morin was VP of Engineering at Zenly before Snapchat bought it for nine figures.

Why should solo builders care about NVIDIA lock-in?

Because lock-in compounds, and so does the bill. Three things stack up against you if you run your own AI-powered tool, chatbot, or automation pipeline:

NVIDIA GPUs are expensive and scarce. Cloud providers charge premium rates for NVIDIA instances because demand outstrips supply, and if your app processes 10,000 prompts a day, you feel it. Once your stack is tuned for NVIDIA, switching gets riskier every month — providers know this and price accordingly.

Meanwhile, the alternatives got good. AMD’s latest GPUs, Google’s TPUs, and even Intel Arc handle inference well; the software barrier, not the hardware, kept most developers from trying them. LLMD removes that barrier.

Practically: if you’re building a side project that uses AI, you could run inference on cheaper AMD hardware or a local Apple Silicon Mac without rewriting anything. Running a small AI-powered business? Mix and match hardware based on price and availability instead of accepting whatever your cloud provider offers.

Is LLMD really free, and what’s the catch?

It’s free right now, but it’s not open source, and ZML plans to charge eventually. Morin’s reasoning, in his own words: “I’d rather measure and then generate revenue where it is most effective without hindering my growth stupidly because I have been too greedy from the get-go.”

It’s the playbook we’ve seen from this free cleaning robot training data startup: give away the core product, build adoption, learn how people use it, then charge for premium features or enterprise support. ZML is running the same play at the infrastructure layer.

The cap table tells you who’s paying attention. Backing from the founders of Docker and Hugging Face, plus Yann LeCun (now with AMI Labs). They raised $20 million with a team of 20 people, which means they can move without burning through runway.

How does ZML compare to the AI inference gold rush?

ZML is one of several startups chasing the same problem: making AI inference faster and cheaper. Baseten was recently valued at $13 billion in what The Next Web called the “inference gold rush”; Inferact came from the creators of vLLM, and RadixArk spun out of SGLang.

What separates ZML is the hardware-agnostic angle. Most competitors optimize for NVIDIA because that’s where the market is today. ZML bets the market fragments as more chip makers enter — European companies like Axelera, Fractile, Kalray, and SiPearl are all building AI-specific chips. If you can run inference on any of them at peak speed, you’re ready for a future where hardware choice matters.

This connects to our breakdown of OpenAI’s hardware ambitions: the companies building AI models increasingly think about hardware too, because the inference layer is where the money flows.

What should you actually do with your stack right now?

If you only use AI as a consumer — ChatGPT, Claude, Gemini — nothing changes today. If you process AI prompts in production, a few things are worth doing. First, audit your inference costs: if you’re spending more than $100/month through a single cloud provider, you’re a candidate for testing alternatives. Second, watch LLMD’s development; it just launched, so expect rapid iteration, and ZML’s track record suggests releases will land quickly.

Don’t rewrite anything yet, though. LLMD is free and worth testing, but it’s early, and the existing infrastructure works. This is about having options, not ripping out your stack. And if your costs drop because you can run on cheaper hardware, that’s margin you can reinvest or pass to customers.

If the pace of AI tool releases is overwhelming, you’re not alone — our guide on escaping AI tool overwhelm covers how to evaluate what actually matters for your situation.

FAQs

Is ZML’s LLMD open source? No. LLMD is free to use but not open source. ZML plans to monetize later, likely through premium features or enterprise support, once they’ve measured how developers actually use the server. The founder has said he’d rather grow adoption first than charge from day one.

Which chips does LLMD support? LLMD runs open-source AI models across NVIDIA, AMD, Google TPU, Apple Metal, and Intel Arc. ZML claims peak performance on each, sometimes exceeding what each chip’s native software stack achieves.

Do I need to rewrite my code to switch away from NVIDIA GPUs? That’s the point of LLMD: no. It translates model operations into instructions for whichever hardware you point it at, so the same model can run on AMD chips, TPUs, or Apple Silicon without changes to your stack.

How much does ZML’s LLMD cost? Nothing right now. ZML raised $20 million and is giving LLMD away free to build adoption, with monetization planned for later. Expect that to mean paid enterprise tiers or premium features once the user base is established.

Who is behind ZML? A 20-person team in Paris founded by Morin, previously VP of Engineering at Zenly (acquired by Snapchat for nine figures). Backers include the founders of Docker and Hugging Face, and Turing Award winner Yann LeCun, now with AMI Labs.

If you’re just getting started with AI tools, begin with /start-here/ — no hardware required.