Is Nvidia's CUDA moat gone? Anthropic says Mythos ported Claude to AMD chips in three days

"The Nvidia CUDA moat is gone." That one-liner from Marcelo P. Lima on Nov. 11 drew more than 5,600 likes and 510,000 views on X. It was set off by a single number Anthropic disclosed in its IPO roadshow video.

According to the roadshow remarks transcribed by Andrew Curran, Anthropic started working with AMD GPUs for the first time this year. "In the past, bringing on a new chip platform took dozens of engineers months of work," the company said in the video. "This time we had Mythos try it out. It got Claude running on AMD in just 3 days with basically no human intervention." Claude ported or wrote almost 100% of the kernels, the low-level code that runs on the chip. OfficeChai reported the target was AMD's MI355X and that the job used to take six to 12 months.

Anthropic now trains and serves Claude on four chip families: Nvidia GPUs, Amazon's Trainium, Google's TPUs and AMD. The idea is to treat every chip as one fungible pool of compute and chase the best price-performance. AMD has agreed to supply up to 2 gigawatts of MI450 chips and invest up to $5 billion in Anthropic as deployment milestones are hit. The number fed hopes that AI can paper over AMD's long-criticized software stack, ROCm.

Not everyone bought the framing. Critics noted the three days covered getting prompt processing for one model running at a useful level of performance. "The moat was never the onboarding time. CUDA simply comes with superior hardware," @klos replied, drawing 168 likes. Another user wrote they believe the three-day bring-up but wouldn't sign off on "as good as CUDA."