Claude cracks open Nvidia's 20-year CUDA moat in three days? Anthropic has Mythos port GPU low-level code on its own.

Beating AI News Flash: Anthropic disclosed during its IPO roadshow that Claude Mythos Preview took only 3 days to get a Claude model running key computational steps on AMD GPUs. Nearly all the required GPU kernels were ported or written by Claude itself, with minimal human intervention. Anthropic said that in the past, onboarding a new chip platform typically required dozens of engineers and 6 to 12 months.

One of Nvidia's most irreplaceable advantages is CUDA.

CUDA is the GPU computing platform Nvidia launched in 2006. By 2018, mainstream deep learning frameworks such as PyTorch and TensorFlow had already widely adopted CUDA and its accompanying computing libraries. Nvidia also progressively refined multi-GPU communication, debugging, and performance optimization tools. Developers could efficiently utilize GPUs without writing large amounts of low-level code from scratch.

As more and more AI training and inference programs were developed around CUDA, Nvidia accumulated a vast amount of ready-made code and tools. Developers often choose Nvidia because the software has already been adapted. Even if AMD builds high-performance GPUs, customers wanting to switch would still face code migration, compatibility testing, and performance optimization challenges.

AMD's MI355X packs 288GB of high-bandwidth memory, and its hardware specs are no slouch. The trouble is that GPU kernels written for Nvidia may not run directly on AMD, and even if they do, performance may not be ideal. Engineers still need to adjust code for AMD's chip architecture and repeatedly test computational results and running speed.

AMD has long offered tools like ROCm and HIPIFY to help developers migrate CUDA programs. Cross-platform programming tools such as Triton are also lowering migration barriers, but complex low-level computations still require substantial manual optimization.

This time, Anthropic had Mythos take over part of the work, porting or rewriting GPU kernels on its own. According to the roadshow chart notes, it achieved usable performance for a Claude model's prompt processing on AMD MI355X within 3 days.

Anthropic has in fact long been reducing its dependence on a single chip vendor. Its prospectus discloses that the company spent years building a compute scheduling system spanning Nvidia GPUs, AMD GPUs, Google TPUs, and Amazon Trainium, allowing it to select different chips based on price and performance.

In July of this year, Anthropic also reached a partnership with AMD, planning to deploy AMD's next-generation GPUs starting in the first half of 2027, with a scale of up to 2GW. The two sides will also use Claude to accelerate ROCm development and GPU performance optimization. AMD has committed to investing up to $5 billion in Anthropic in the future.

Investor Marcelo P. Lima believes that Nvidia's CUDA moat has disappeared. But analyst Ben Pouladian offered a different view: getting part of a model's computation to run on AMD is very different from getting an entire system to achieve stable, cost-effective, and competitive performance in large-scale deployment.