Anthropic’s Claude Mythos Preview model reportedly achieved a significant milestone by bringing its AI to AMD GPUs in just three days, with “basically no human intervention,” compared to the usual 6-12 months. This dramatic reduction in time highlights the potential erosion of Nvidia’s CUDA programming moat as AI driven code generation becomes more prevalent.

At Anthropic’s IPO roadshow, co-founder and chief compute officer Tom Brown detailed the company’s achievements. During the keynote at AMD’s Advancing AI event on July 23, Brown explained that one engineer and Claude, Anthropic’s AI model, managed to get the Claude Mythos Preview running on AMD MI355X GPUs within a single weekend. According to the company, Mythos ported or wrote almost 100% of the necessary kernels, which typically take “dozens of engineers months” to complete using traditional methods.

The speed and efficiency of this process underscore Anthropic’s multi-chip strategy, which now includes GPUs from both Nvidia and AMD, Google’s TPUs, and AWS Trainium. Anthropic’s approach to managing its compute resources as a fungible pool aims to leverage the strengths of different hardware platforms, potentially reducing AI inference costs and improving overall performance.

While Nvidia has long emphasized the importance of its CUDA programming platform, the company’s CEO Jensen Huang acknowledged in an interview that AI models are already used to generate kernels internally. Huang noted that nobody knows Nvidia’s architecture better, but the use of AI in code generation could challenge this claim. He suggested that Anthropic’s approach is a unique exception rather than a trend.

AMD, on the other hand, has been actively promoting the use of AI in code generation through its ROCm platform and HIP programming layer. In April 2026, AMD’s vice president of AI software, Anush Elangovan, compared Anthropic’s Claude to AMD’s HIPify tool, stating that Claude was better because it included web search capabilities. AMD’s senior vice president of AI, Vamsi Boppana, also highlighted the potential of AI in generating GPU kernels, noting that in some cases, the resulting kernels were even “better than the most manually tuned versions.”

Despite these advances, the practical implementation of AI-driven code generation still requires significant hardware resources. Each AI agent needs access to GPUs to test and refine its code, which can create a bottleneck for development. According to SemiAnalysis, this requirement might limit the widespread adoption of agentic coding in the near future.

The implications for the broader AI ecosystem are significant. As AI models like Claude continue to streamline code generation, the traditional barriers to adopting new hardware platforms may diminish. This could lead to a more competitive market for AI inference chips, where cost and performance become more critical factors.

Anthropic’s prospectus notes that the company can onboard new chip platforms in a fraction of the time, compared to its historical experience. This claim is backed by the company’s recent success in bringing up its models on AMD GPUs in just three days. As Anthropic prepares for its IPO, this milestone is likely to attract more investors, given the potential to reduce AI inference costs and improve overall efficiency.

The first generation of Anthropic’s AMD MI450 chips, expected in the first half of 2027, will provide an opportunity to test whether the three day bring up can be repeated. This will be crucial for Anthropic’s strategy and could have broader implications for the AI hardware market.

In summary, Anthropic’s achievement marks a significant shift in how AI models are integrated onto different hardware platforms. As more companies explore AI-driven code generation, the traditional programming barriers may become less significant, potentially reshaping the AI inference landscape.

Source: https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropic-says-claude-mythos-slashed-software-work-to-onboard-amd-gpus-from-up-to-12-months-to-around-three-days-writing-almost-100-percent-of-kernels-autonomously-chips-away-at-nvidias-cuda-moat

Thinking about building an AI product?

Get in Touch