Anthropic says its Claude Mythos Preview model ported and optimized Claude to run on AMD's MI355 GPUs in three days. The disclosure could weaken Nvidia's CUDA software advantage and help AMD, though Nvidia keeps ecosystem and supply strengths.
Anthropic's Claude Mythos Preview model took just three days to port and optimize Claude to run on AMD's MI355 GPUs, according to a slide in the AI lab's latest investor deck ahead of its initial public offering. Claude ported or wrote 100% of the kernels. The result potentially weakens Nvidia's CUDA moat significantly, writes The Motley Fool.
Three days instead of months
A kernel is AI code that tells a GPU how to execute a mathematical operation. Anthropic said onboarding new AI accelerators would normally take a dozen engineers between six and 12 months.
According to The Motley Fool, Anthropic's goal was a "fungible compute pool" where it could bring any AI chip up to speed within days. That covers GPUs from Nvidia or AMD, as well as custom silicon such as Amazon's Trainium chips or Alphabet's Tensor Processing Units. The approach would let it buy chips based on the best price performance regardless of vendor.
Why CUDA mattered to Nvidia
Nvidia's GPUs have become the main chips powering AI workloads, and Nvidia's CUDA platform made them easy for developers to program. Most original AI code was written on CUDA and optimized for Nvidia's chips. One of Nvidia's biggest advantages has been that software written on CUDA did not run on competitors' chips and was difficult to modify for them.
If Claude Mythos can greatly reduce the time, cost, and complexity of running AI models on alternative chips, the Motley Fool writes that this meaningfully levels the playing field.
Nvidia shares were down 2.01% at $162.85, as quoted on the page.
AMD's cost edge, Nvidia's exposure
The Motley Fool's Geoffrey Seiler says AMD's advantage is cost. Nvidia's Vera Rubin platform can cost around $40 billion per gigawatt, while AMD's Helios sells for roughly $20 billion per gigawatt. Even if Helios lags Vera Rubin on overall performance, AMD could take a significant share on price-to-performance.
Nvidia, by contrast, could face pressure to lower prices if its CUDA moat significantly shrinks and it starts losing share. That would hit sales, gross margin and profitability. Seiler still would not write Nvidia off, because it has the premier AI infrastructure solution, the best overall ecosystem, and has locked up a lot of supply-constrained high-bandwidth memory that AI chips still need.
Source: The Motley Fool
Trading involves risk.
Risk Disclaimer
