Did Claude breach NVIDIA's 20-year CUDA moat in three days? Anthropic lets Mythos port GPU low-level code by itself

Beating AI News: In just three days, the Claude Mythos Preview enabled a Claude model to run critical computations on AMD GPUs. Almost all required GPU kernels were either ported or written by Claude itself, with minimal human intervention. It was noted that in the past, integrating new chip platforms typically required dozens of engineers spending 6 to 12 months.

One of the hardest advantages to replace is CUDA.

CUDA is a GPU computing platform launched in 2006. By 2018, mainstream deep learning frameworks like PyTorch and TensorFlow had widely adopted CUDA and its associated computing libraries. It has also continuously improved multi-card communication, debugging, and performance optimization tools. Developers can efficiently utilize GPUs without having to write a large amount of low-level code from scratch.

As more and more AI training and inference programs are developed around CUDA, a vast amount of ready-made code and tools have been accumulated. Developers often choose based on the fact that software is already adapted. Even if AMD produces high-performance GPUs, customers face challenges such as code migration, compatibility testing, and performance optimization if they want to switch.

AMD's MI355X is equipped with 288GB of high-bandwidth memory, and its hardware configuration is not weak. The trouble lies in the fact that the 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 the code according to AMD's chip architecture and repeatedly test the computation results and running speeds.

AMD has long launched tools like ROCm and HIPIFY to help developers migrate CUDA programs. Cross-platform programming tools like Triton are also reducing migration difficulties, but complex low-level computations still require a lot of manual optimization.

This time, Mythos took on part of the work, porting or rewriting GPU kernels by itself. According to the presentation charts, it achieved usable performance for a Claude model's prompt processing on the AMD MI355X in just three days.

In fact, it has long been reducing dependence on a single chip manufacturer. The prospectus disclosed that the company spent years establishing a computing power scheduling system that spans GPUs, AMD GPUs, Google TPUs, and Amazon Trainium, allowing for the selection of different chips based on price and performance.

In July this year, it also reached a cooperation with AMD, planning to deploy AMD's next-generation GPUs starting in the first half of 2027, with a scale of up to 2GW. Both parties will also leverage Claude to accelerate ROCm development and GPU performance optimization, with a commitment to invest up to $5 billion in the future.

Investor Marcelo P. Lima believes that CUDA's moat has already disappeared. However, analyst Ben Pouladian offered a different view: getting part of the model's computations to run on AMD is very different from achieving stable, cost-effective, and competitive performance for the entire system in large-scale deployments.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.