Nvidia's most valuable asset was never really its chips. It was the switching cost — the months of painstaking work required to move an AI model off CUDA and onto anything else. Anthropic just claimed its own AI erased that cost in a long weekend.

In its IPO roadshow video, Anthropic co-founder and Chief Compute Officer Tom Brown said the company's frontier model, Mythos, got Claude running on AMD's MI355X GPUs in about three days, "with basically no human intervention," The Decoder reported. The model "ported or wrote almost 100% of the kernels" — the low-level code that makes a model actually run on a given chip. By Anthropic's own account, the same job used to take dozens of engineers six to twelve months.

Read the fine print before declaring the moat dead. The three-day figure covers prompt processing for one Anthropic model, at what Anthropic calls "a useful level of performance." It is not the whole model, and it is not training. And the numbers come from an investor presentation selling shares ahead of a November 18 pricing, not from an independent benchmark. AMD has not commented.

But the direction of travel is hard to ignore. Anthropic says it now trains and serves Claude across all four major chip families — AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs — treating them as one fungible compute pool and buying whichever offers the best price-performance. That is a deliberate bet against vendor lock-in, and it is backed by money: AMD agreed in July to invest up to $5 billion in Anthropic, with Anthropic set to deploy up to two gigawatts of AMD's MI450-series chips starting in the first half of 2027. If AI can do the porting, the cheapest chip wins — and Nvidia's hardware margin of more than 70 percent is exactly the kind of fat a fungible market eats first.

The irony is that Nvidia is on both sides of this trade. The chipmaker committed up to $10 billion to Anthropic last November and is participating in the $26 billion private placement alongside the IPO, according to Anthropic's S-1/A filing.

Nvidia reports earnings on November 17, five days after losing the CEO who built the company. Expect the CUDA question to come up. Anthropic's answer, for now, is a chart from a roadshow video and a claim nobody outside the company has verified. The test will be whether anyone else's model can do the same trick — and whether the next chip switch takes three days, or three days plus a year of asterisks.

— via The Decoder