Company Updates

Anthropic ports Claude to AMD chips in 3 days with its own AI

Share
Anthropic ports Claude to AMD chips in 3 days with its own AI

Anthropic's internal frontier model, Claude Mythos Preview, ported its Claude system to AMD's MI355X accelerators in three days—with almost no human involvement. The feat, revealed in the company's IPO roadshow materials, marks the first time an AI wrote nearly all the low-level code needed to run a production model on non-Nvidia hardware.

TL;DR

  • Anthropic's AI model ported its Claude system to AMD GPUs in three days, a process that typically takes teams of engineers six to twelve months.
  • The company used its internal Claude Mythos Preview to write nearly all the kernel-level code for AMD's ROCm platform, bypassing Nvidia's CUDA dominance.
  • Anthropic's move strengthens its negotiating position with chipmakers ahead of its $26B IPO and $5B investment from AMD.

What happened

Anthropic disclosed in its IPO roadshow that its internal model, Claude Mythos Preview, ported a production-ready Claude system to AMD's MI355X accelerators in approximately 72 hours. The company stated that Mythos wrote or ported nearly all the kernel-level code itself, a task that traditionally requires teams of dozens of engineers working for six to twelve months.

The test focused on prompt processing for a single model, achieving what Anthropic calls a "usable performance level" on AMD hardware. Mythos Preview is not publicly available; Anthropic released its capabilities as Claude Fable 5 only after adding safeguards for cybersecurity.

Anthropic runs its workloads across four chip families—Amazon Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs—treating them as a shared compute pool. Faster hardware readiness directly improves the company's leverage in chipmaker negotiations.

In July 2025, Anthropic signed a deal with AMD for up to 2 gigawatts of MI450 chips, with the first gigawatt expected online in the first half of 2027. In return, AMD committed to investing up to $5 billion in Anthropic as part of the company's $26 billion private placement tied to its IPO.

Why it matters

For AI startups and developers, this demonstrates that AI-native tooling can dramatically accelerate hardware porting, reducing reliance on Nvidia's CUDA ecosystem. It signals a shift toward multi-vendor hardware strategies, which could lower costs and increase competition in the AI chip market.

For investors, Anthropic's ability to rapidly deploy on AMD hardware strengthens its negotiating position with chipmakers ahead of its public listing. The company's multi-chip strategy diversifies its supply chain, mitigating risks tied to Nvidia's dominance.

The breakthrough is limited to prompt processing and does not address training workloads, which remain far more complex. Nvidia's CUDA ecosystem, with over 7.5 million developers, remains entrenched, supported by decades of tooling, libraries, and networking infrastructure.

Open questions include whether Mythos can port training workloads to AMD and how performance compares to Nvidia GPUs. Independent verification of Anthropic's claims is also pending, as noted by industry analysts.

Key facts

  • Anthropic ported its Claude system to AMD's MI355X accelerators in three days using its internal Claude Mythos Preview model.
  • The porting task focused on prompt processing and achieved a "usable performance level" on AMD hardware.
  • Mythos Preview wrote or ported nearly 100% of the kernel-level code for AMD's ROCm platform, a task typically requiring teams of dozens of engineers for six to twelve months.
  • Anthropic runs workloads across four chip families: Amazon Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs.
  • Anthropic signed a deal with AMD in July 2025 for up to 2 gigawatts of MI450 chips, with the first gigawatt due in the first half of 2027.
  • AMD committed to investing up to $5 billion in Anthropic as part of the company's $26 billion private placement tied to its IPO.
  • Nvidia reports over 7.5 million CUDA developers, according to its annual report.
  • AMD's stock rose from about $515 to $585 in the two trading days after the roadshow slide became public, before retracing to around $504.

Context

The move underscores a broader industry trend toward AI-native software development, where models assist in writing the low-level code needed to run on diverse hardware. AMD, OpenAI, and Google have all invested in similar tooling to reduce dependence on Nvidia's CUDA ecosystem.

AMD's ROCm.AI initiative, introduced in 2025, aims to optimize AI models for AMD hardware automatically, aligning with Anthropic's approach. OpenAI has also designed its own AI chip, Jalapeño, with AI assistance, while Google has long operated its TPU line with proprietary software outside CUDA.

Nvidia's CUDA remains the de facto standard for AI development, supported by a vast ecosystem of libraries, networking technology, and developer tools. Despite Anthropic's breakthrough, analysts caution that Nvidia's lead is not easily replicated, given the depth and maturity of its platform.

Topics

Related coverage

Join the discussion

Have a take on this story? Weigh in with our community on Facebook.

💬 Discuss on Facebook →