Anthropic says AI ported Claude to AMD GPUs in three days

Anthropic says it ported Claude to AMD GPUs in just three days, with almost no human intervention, cutting a process that normally takes six to twelve months down to a long weekend. The claim comes from an investor presentation, and if the numbers hold up, they strike at the heart of Nvidia's most durable competitive advantage: the CUDA software ecosystem.

Nvidia's dominance in AI chips rests on more than fast hardware. CUDA, the company's parallel computing platform, has locked developers into its chips for nearly two decades. Porting AI models to other architectures demands so much engineering effort that many companies don't even attempt it. Anthropic now says it used AI itself to break that lock-in, and the implications for a compute-hungry industry are significant.

What did Anthropic actually do?

According to Tom Brown, co-founder and Chief Compute Officer at Anthropic, the company now trains and runs Claude on all four major chip families, treating them as a single, interchangeable compute pool. Anthropic made what Brown calls an "unconventional bet" on a multi-chip strategy early on, at a time when most of the industry considered the effort not worth it.

The strategy is expensive by design. Each chip family needs its own custom software and inference stack, which costs time and money to build and maintain. Anthropic has been investing in this work for years and has since become "incredibly good" at it, Brown says.

The most concrete test of that investment came with AMD GPUs. Until now, bringing up a new chip platform kept dozens of engineers busy for months. This year, Anthropic used AMD GPUs for the first time and handed the porting work to its own frontier model, Mythos.

The result, according to Anthropic: Claude was running on AMD after three days, with almost no human intervention, and Mythos wrote or ported nearly 100 percent of the kernels itself. The takeaway in plain terms: Anthropic claims an AI model did in days what has historically taken a team of engineers half a year or more.

Why does this matter for Nvidia?

Nvidia currently enjoys a gross margin of more than 70 percent on its AI chips, which is unusually high for a hardware business. That pricing power exists largely because customers can't easily leave. CUDA's ecosystem of libraries, tools and optimized code makes switching to a rival chip painful enough that many buyers simply pay the premium.

If Anthropic's figures are accurate, that software overhead shrinks dramatically when AI can write the porting code itself. Companies could switch between chip providers and pick whichever option is the most affordable or readily available at a given moment. In a compute market where supply is tight, that flexibility turns chip purchases into a price-performance decision rather than an ecosystem commitment.

It also changes the calculus by workload. Some chips are better suited to training, while others perform better on inference, the stage where a trained model actually answers requests. A multi-chip approach lets a lab match each workload to the best hardware. Anthropic says treating compute from different vendors, chip architectures and locations as one interchangeable pool means it gets more out of every dollar of compute than any other frontier AI lab.

The stakes for the market are straightforward: Nvidia's margin cushion depends on switching costs, and AI-written porting code attacks those costs directly. That's why one lab's investor deck is being read as a signal about the entire AI hardware market.

What are compute multipliers?

Beyond the multi-chip strategy, Anthropic relies on what it calls compute multipliers. These are research breakthroughs that squeeze more performance out of the same physical computing capacity. According to the company, these multipliers build on one another, so each new breakthrough amplifies the effect of the previous one.

Anthropic's stated endgame is for Claude to discover such compute multipliers on its own. That would amount to recursive self-improvement, a concept that remains intensely debated in AI research and has not been proven in practice. Anthropic isn't claiming it has achieved this, only that it's the direction the research is heading.

The company is also building its own data center capacity through a program called "Anthropic Direct." Instead of buying computing power only as a cloud service, Anthropic leases chips and operates them in purpose-built data centers. The stated goal is to reduce total cost of ownership, which compounds with the compute multipliers: cheaper infrastructure plus more efficient use of it.

In short, the strategy is to attack the cost of AI compute from both the hardware side and the research side at once.

How are the GPU makers tied to Anthropic?

Both major GPU manufacturers now have deep financial ties to Anthropic, which adds a layer of complexity to the company's claims of independence from any single vendor.

In November 2025, as part of a $45 billion deal with Microsoft, Nvidia committed to investing up to $10 billion in Anthropic. According to Anthropic's S-1/A registration statement filed with the SEC on November 10, 2026, Nvidia is also participating in a conditional private placement totaling $26 billion, in which seven technology companies are subscribing to shares at the IPO price.

AMD took its own step in July 2026, announcing a strategic investment of up to $5 billion in Anthropic. In return, Anthropic plans to deploy up to two gigawatts of AMD Instinct MI455X GPUs in Helios rack-scale systems, with the first gigawatt phase beginning in the first half of 2027.

That means the company arguing CUDA lock-in is dead counts both Nvidia and AMD among its investors and suppliers. It doesn't invalidate the technical claims, but it shapes how the market reads them.

What should readers take with a grain of salt?

The figures come from Anthropic's investor presentation and haven't been independently verified. A company preparing for a public listing has clear incentives to present its infrastructure as efficient, flexible and cheap relative to competitors. The three-day AMD port, the near-100 percent kernel automation and the claim of best-in-class compute efficiency all come from the same source: Anthropic itself.

Even with that caveat, the presentation shows something notable. At least one major AI lab no longer considers Nvidia's CUDA dominance insurmountable, and it's demonstrating that AI models themselves can serve as the tool for breaking hardware dependencies. If other labs replicate the approach, the bargaining power in the AI chip market shifts, and Nvidia's record margins become harder to defend.