Anthropic says its AI model wrote nearly all the kernels needed to run Claude on AMD GPUs. Until now, porting a model to a new chip platform typically took six to twelve months.
Anthropic says it used AI to port its Claude model to AMD graphics processing units in just three days. The achievement is described in an investor presentation prepared for the company’s planned IPO, as reported by The Decoder and other media outlets.
According to the presentation, Anthropic’s own frontier model, Mythos, ported or rewrote nearly 100 percent of the required kernels with minimal human intervention. Using conventional methods, bringing a new chip platform online previously took six to twelve months and required dozens of engineers. The figures refer to running prompts from an Anthropic model on AMD MI355X GPUs. They have not been independently verified.
CUDA as a Competitive Advantage
Nvidia’s biggest competitive advantage is widely considered to be the combination of its hardware and CUDA software platform. Because adapting models to other chips can be a complex and time-consuming process, many companies stick with Nvidia. If AI can dramatically reduce the effort involved, businesses could switch between chip vendors more easily and choose hardware based on price or availability. Nvidia currently reports gross margins of more than 70 percent.
Four Chip Families as a Single Compute Pool
According to Tom Brown, Anthropic co-founder and Chief Compute Officer, the company now trains and runs Claude across four chip families: AWS Trainium, Google TPUs, and GPUs from Nvidia and AMD. Anthropic made what Brown describes as an “unconventional bet” on this multi-chip strategy early on, when the industry generally considered the effort unjustifiable.
The company treats computing capacity across different providers, architectures, and locations as a single, interchangeable pool. This capacity serves research, model development, customer demand, and internal workloads. Anthropic also focuses on so-called compute multipliers: research breakthroughs that extract more performance from the same hardware. Through its Anthropic Direct program, the company also leases chips and operates them in its own data centers.
Investments From Both GPU Manufacturers
Both major GPU manufacturers have close financial ties to Anthropic. In November 2025, Nvidia pledged to invest up to USD 10 billion. According to Anthropic’s IPO prospectus, filed on November 10, Nvidia is also participating in a conditional private placement worth USD 26 billion. AMD announced an investment of up to USD 5 billion in July. In return, Anthropic plans to deploy up to two gigawatts of AMD Instinct MI455X GPUs in Helios rack systems. The first phase is scheduled to begin in the first half of 2027. Market research firm Forrester expects that by 2027, one-quarter of AI infrastructure tenders will require accelerators from vendors other than Nvidia.
(Editorial Team)