AMD Is Catching Nvidia on Silicon. It's the Software It Can't Catch.

In 2022, a chip company hired a single freelance developer to do something strange: build a way to run its rival's software on its own hardware, without anyone noticing. The company was AMD. The project was called ZLUDA — a drop-in layer that lets unmodified programs written for Nvidia's CUDA run on AMD GPUs. AMD funded it quietly, then cut it loose, and in February 2024 it surfaced as open-source.⁶ Think about what that admission contains. AMD makes excellent chips. And to compete in AI, it concluded the fastest path was to make its hardware pretend to be Nvidia's.

The story everyone tells about this race is that Nvidia builds faster chips. That AMD is the perennial runner-up with weaker silicon. Almost every word of that is wrong. AMD's silicon is good — good enough to win the world's most demanding contracts. The thing it cannot easily buy, build, or out-engineer is the twenty years of software that runs on top of Nvidia's chips. The chip was never the moat. The code was.

AMD already proved it can build the hardware

Start with the part the popular story gets backwards. In the data-center CPU market — the one Nvidia doesn't even play in — AMD's EPYC reached roughly 32–33% of the x86 server market exiting 2024, with analysts projecting it would push toward 38–39% by the end of 2025.⁸ That is not a distant also-ran; that is a third of the world's servers, taken from Intel one design cycle at a time. And on the GPU side, AMD's Instinct accelerators power El Capitan at Lawrence Livermore — a machine built entirely on AMD chips, both GPU and CPU, that held the world's #1 ranking on the TOP500 list from its debut in November 2024.⁹ When the U.S. government wanted the most powerful supercomputer it could build, it did not buy Nvidia. So when AMD says its hardware is competitive, the receipts exist.

The money is moving too. AMD's Data Center segment doubled in a year — from $6.5 billion in 2023 to a record $12.6 billion in 2024, up 94%⁴⁵ — and CEO Lisa Su attributed more than $5 billion of that to Instinct GPUs alone.⁴ A company that could not build AI silicon does not generate that number. The hardware is not the problem. Which raises the obvious question: if AMD can match the chip, why is it still so far behind?

The moat is the code, and the code compounds

CUDA is Nvidia's software platform for running general computation on its GPUs, and it has been accumulating for the better part of two decades. That length matters more than any single feature. Every AI researcher learned on it. Every deep-learning framework was first optimized for it. Every tutorial, every Stack Overflow answer, every library of hand-tuned kernels assumes it. The asset is not a piece of software — it is the gravitational field that twenty years of developers built around the software. A chip is a thing you buy once. CUDA is a thing your entire team already knows, and the cost of un-knowing it is enormous.

This is why a faster AMD chip does not automatically win a sale. The buyer is not choosing a piece of silicon; they are choosing whether to rewrite, re-validate, and re-tune the code their business already runs. The switching cost lives in the customer's own codebase, which means Nvidia doesn't have to defend it — the customer defends it for them, every time they decline to port. AMD's challenge isn't to be better. It's to be enough better to justify the migration. That is a far higher bar than a benchmark.

And the gravity keeps the chip business spinning. Nvidia's revenue went from $130.5 billion in fiscal 2025, up 114%, to $215.9 billion in fiscal 2026, up another 65%, with Data Center revenue alone hitting $62.3 billion in a single quarter.¹² Each of those dollars funds more libraries, more developer tools, more reasons for the next engineer to learn CUDA and not ROCm. The moat doesn't sit still. It widens with every quarter the lead is spent.

How AMD is attacking a software lead with software

AMD understands this exactly, which is why its real campaign is not about transistors. ROCm — its open software stack — now ships PyTorch official packages as a first-class option, with upstream Linux kernel support, OpenAI Triton support, and Hugging Face integration, and AMD sits as a founding member of the PyTorch Foundation.⁷ That last point is the strategic one: if the framework most researchers actually write in supports your hardware natively, the developer never has to touch CUDA's lowest layers at all. AMD isn't trying to out-build CUDA feature-for-feature. It's trying to make the layer above CUDA — the one developers live in — vendor-neutral.

ZLUDA attacks from the other direction. Rather than asking developers to rewrite anything, it lets some unmodified CUDA binaries run directly on AMD GPUs.⁶ Pair that with AI-assisted code porting — tools that can now migrate a complete CUDA backend to ROCm without a single line written by hand¹⁰ — and the twenty-year head start starts to look less like a wall and more like a toll booth someone is quietly building a bypass around. The switching cost is exactly the thing both projects exist to drive toward zero.

That line is remarkable coming from the incumbent. Nvidia's own filing concedes that when policy locked it out of China, the absence handed rivals room to grow the one thing that actually matters — the ecosystem.³ The company knows precisely where its moat is, because it just watched a piece of it form on the other side.

Isn't a 'software moat' just a polite word for inevitable decline?

The fair objection is that switching costs always look unbreakable right up until they break — and the tools to break this one already exist. ZLUDA is open-source. PyTorch runs on AMD. AI can port code now. So isn't CUDA just the next dominant standard waiting for its disruption? Maybe. But notice the asymmetry. Every porting tool AMD ships is a layer of translation on top of an interface Nvidia controls and keeps moving. Building your competitive strategy on emulating a rival's API means you are forever one version behind, re-implementing their new features after they ship instead of inventing your own. ZLUDA's very existence is the tell: AMD couldn't beat the standard, so it tried to become compatible with it — and compatibility is a follower's position by definition.

The honest counter to Nvidia's confidence is subtler. Switching costs erode not when a rival matches them feature-for-feature, but when the layer customers actually touch moves somewhere neutral. If PyTorch and Triton become the real interface and CUDA becomes plumbing nobody sees, the moat doesn't fall — it gets abstracted away beneath a floor that no longer cares which chip is underneath. That is the genuine race. Not chip versus chip. Nvidia widening the lock-in through vertical integration, against an industry quietly agreeing to stop writing code that knows whose silicon it's running on.

AMD can build a chip that earns the world's top supercomputer ranking⁹ and still trail in AI, and that tells you everything about where the contest actually lives. Nvidia's genius was never only the GPU. It was getting an entire generation of engineers to write their work in a language it controlled, so that two decades later, even a rival with competitive silicon and a third of the server market has to begin by funding a project to speak that language back. The chips will keep trading benchmarks. The real war is over whose code the next model is written in — and the company that wins that war barely has to ship hardware at all.