Anthropic Ported Claude to AMD in 3 Days Instead of 6–12 Months, Using AI

If you have ever tried to run code written for CUDA on anything else, you know the feeling: in theory everything is possible, but in practice it means six months of work. For years, that barrier has protected Nvidia more effectively than its chips themselves. Now Anthropic says it has learned to get around it using its own model.

What exactly Anthropic claimed

According to co-founder and chief compute officer Tom Brown, the company trains and runs Claude on all four major chip families: AWS Trainium, Google TPU, Nvidia GPUs and AMD GPUs. Each one requires its own software and its own inference stack, which is expensive and slow. Anthropic bet on a multi-chip strategy early, when most of the industry considered it not worth the effort. The company has been investing in this engineering for years and, in Brown's words, has become "incredibly good" at it.

The most concrete example is AMD. This year Anthropic used AMD GPUs for the first time and handed the work to its frontier model, Mythos. According to the company, Claude was up and running on AMD in three days with almost no human involvement. The model wrote or ported almost 100% of the kernels. Kernels are low-level programs that do the math directly on the GPU, and they are what usually ties code to specific hardware. Traditionally, dozens of engineers need six to twelve months for such a launch. The figures in the presentation refer to prompt processing by an Anthropic model on AMD MI355X GPUs.

Why this matters beyond Anthropic

Imagine an application tightly bound to a single cloud's SDK. Every migration means rewriting hundreds of integrations, so nobody migrates. If AI can do such a migration over a weekend, vendor lock-in stops being a wall and becomes an ordinary line item. CUDA works the same way: it holds developers not through quality, but through the cost of leaving.

If the picture is accurate, companies will be able to choose chips based on price and availability rather than compatibility. In a compute shortage, that is a serious advantage. In addition, some chips suit training better, others inference. Nvidia's margin exceeds 70%, unusually high for hardware, and a lower switching cost hits it directly.

Anthropic also talks about "compute multipliers": research techniques that squeeze more out of the same physical hardware. The company is also developing its Anthropic Direct program: it rents chips and operates them itself in purpose-built data centers to lower total cost of ownership. In the future, Claude is supposed to find such multipliers on its own. That would be the recursive self-improvement that the industry argues about a great deal but that no one has yet proven.

Caveats and money

The main caveat: all the figures come from an investor presentation and have not been independently verified. This is material from a company preparing for an IPO. Its S-1/A filing was submitted to the SEC on November 10, 2026. Also, "three days" means reaching a working level of performance, not necessarily parity with hand-optimized CUDA kernels. And this concerns Mythos, a closed frontier model, in the hands of a team that has spent years building multi-chip infrastructure. No one promises ordinary teams such a result in three days.

There is also context worth keeping in mind: both GPU companies are financially tied to Anthropic. In November 2025, as part of a $45 billion deal with Microsoft, Nvidia pledged to invest up to $10 billion. It is also taking part in a conditional private placement of $26 billion, in which seven tech companies buy shares at the IPO price. In July 2026, AMD announced strategic investments of up to $5 billion. In return, Anthropic plans to deploy up to two gigawatts of Instinct MI455X GPUs in Helios racks, with the first gigawatt due to start in the first half of 2027.

What to do with this in practice

For engineers working on inference, the main takeaway is this: porting kernels with an AI agent is worth trying on at least one or two of your hottest operators. Such an experiment once looked like madness, but it may now pay off even in a pilot.

Product teams and founders should build multi-vendor options into their plans and contracts, especially if compute costs are growing faster than revenue. Your negotiating position with hardware and cloud providers is noticeably better when you have a plausible alternative.

Still, don't expect miracles: "it runs" and "it runs as fast and cheaply as on CUDA" are different things. Until there are independent measurements, this is a strong claim, not a proven fact. But the shift itself is already visible: AI is starting to break the hardware dependence on which part of Nvidia's market power rested.