Anthropic claims Mythos ported Claude to AMD GPUs in three days, versus six to twelve months

Anthropic says its frontier model Mythos wrote or ported nearly 100 percent of the kernels needed to run Claude on AMD GPUs, bringing the platform up in three days with almost no human intervention. According to the company, traditional engineering approaches need six to twelve months to onboard new accelerators. The claim comes from an Anthropic investor presentation and has not been independently verified, The Decoder reports.

What Anthropic claims

Tom Brown, Anthropic co-founder and Chief Compute Officer, says the company now trains and runs Claude on all four major chip families: AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs. Anthropic treats compute from different vendors, chip architectures and locations as one interchangeable pool. It allocates that pool across research, model development, customer demand and internal Claude usage.

Brown describes the multi-chip approach as an "unconventional bet" made early, when most of the industry considered the effort not worth it. Each chip needs its own custom software and inference stack. Anthropic has invested in this work for years and has become "incredibly good" at it, Brown says.

The AMD port is the most concrete example. Anthropic used AMD GPUs for the first time this year. Until now, bringing up a new chip platform kept dozens of engineers busy for months, according to the company. With Mythos doing the work, Claude was running on AMD after three days. The performance figures Anthropic shared refer to prompt processing of an Anthropic model on AMD MI355X GPUs.

Why CUDA matters here

Nvidia's competitive edge combines hardware with CUDA, the software ecosystem that locks developers into its chips. Porting models to other architectures takes enough effort that many companies do not try. Nvidia also holds a margin of more than 70 percent, which The Decoder describes as unusually high for hardware.

Compute multipliers and Anthropic Direct

Anthropic also points to what it calls compute multipliers: research breakthroughs that extract more performance from the same physical capacity. The company says these build on one another, so each new breakthrough amplifies the previous one. It says Claude should eventually discover such multipliers on its own, which would amount to recursive self-improvement. The Decoder notes that this concept is intensely debated in AI research and has not been proven.

The company is also building capacity through its "Anthropic Direct" program. Rather than buying compute only as a cloud service, Anthropic leases chips and operates them in purpose-built data centers, with the goal of reducing total cost of ownership. Anthropic claims it gets more out of every dollar of compute than any other frontier AI lab.

Financial ties to both GPU makers

• Nvidia: In November 2025, as part of a $45 billion deal with Microsoft, Nvidia committed to invest up to $10 billion in Anthropic. Anthropic's S-1/A registration statement, filed with the SEC on November 10, 2026, shows Nvidia also participating in a conditional private placement totaling $26 billion, in which seven technology companies subscribe to shares at the IPO price.
• AMD: AMD announced a strategic investment of up to $5 billion in July 2026. 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.

What this means

If the figures hold, the software overhead long considered Nvidia's strongest advantage shrinks when an AI model does the porting work. Companies could then choose between chip vendors on price or availability, which matters in a compute market where supply is tight. Anthropic's pooled approach also lets it match chips to tasks, since some chips suit training and others suit inference.

The evidence is limited. The numbers come from Anthropic's own investor presentation, and the three-day figure covers bringing hardware to a usable performance level, not matching Nvidia's results. The company also has financial relationships with both Nvidia and AMD. Still, the claim shows that at least one major AI lab no longer treats CUDA dominance as insurmountable, and that AI models could serve as a tool for breaking hardware dependencies.