Anthropic says Mythos ported Claude to AMD chips in three days: Nvidia's CUDA moat under pressure

By Yahoo Finance · November 11, 2026.

Anthropic says in its pre-IPO investor presentation that its Mythos model got Claude running on AMD GPUs in about three days and “with basically no human intervention.” According to the company, Claude ported or wrote nearly 100% of the “kernels,” the small low-level programs that carry out mathematical operations on the chip. Anthropic maintains that, until now, bringing up a new hardware platform took dozens of engineers months of work; the figure Yahoo Finance cites is 6 to 12 months.

The news comes at a bad time for Nvidia. Yahoo Finance reports that its board has opened an independent investigation into its CEO, Jensen Huang, and that there are shareholder lawsuits. On November 11, according to the page's own data, Nvidia shares were down 5.76% while AMD's were up 11.12%. Those moves can't be attributed solely to this announcement, but the narrative fits them.

To understand why it matters, you need to understand CUDA. It is the software platform developers use to program Nvidia GPUs. For years, companies have built their code around it, so switching to another manufacturer's chips means months of rewriting and optimization. That switching cost is the “moat” that keeps customers locked in and partly supports Nvidia's prices.

There is an important caveat that the source itself highlights: those three days cover one part of the process for a specific model, reading and processing a prompt with useful performance. That is not the same as porting training, full inference and the entire software stack at CUDA's production quality. In addition, the figure comes from a promotional Anthropic document, not an independent audit.

Our take: the relevant point is not that CUDA is dying, but that its value erodes if AI turns porting between chips into a matter of compute hours rather than engineer-years. Until now, Nvidia's advantage lay largely in software and friction, not just silicon. If that friction drops, bargaining power shifts toward customers and toward in-house chip designers (the custom ASICs from Google, Amazon, Microsoft or OpenAI, which Yahoo Finance cites as mounting pressure). The question the original analysis raises is how much more customers will be willing to pay to stay in the Nvidia ecosystem.

There is also a takeaway about Anthropic. Broadly speaking, a lab using its own model to accelerate its infrastructure illustrates the automation loop behind the optimism about AI: models improve the tools that make models better. But it should be treated as an announced capability, not a proven one, until there are independent benchmarks of performance per dollar on those AMD chips. A roadshow is a showcase, and the incentive to present the best possible figure is obvious.

In the short term, this adds volatility and pressure on Nvidia's margins, and strengthens the case for diversifying compute suppliers. In the long term, it is good news at heart: more competition in hardware means cheaper compute, and cheap compute is one of the levers toward the abundance that would speed up biomedical research and access to AI. For industry context, dependence on a single supplier has been one of this industry's most expensive bottlenecks; if AI itself helps break it, that would be, if confirmed, a rather instructive irony.