Clarity for AI buildersSponsored by AI Routing
availability

Anthropic says AI cut AMD chip porting to three days, challenging CUDA’s software advantage

Anthropic says AI cut AMD chip porting to three days, challenging CUDA’s software advantage
Anthropic says AI brought Claude onto AMD GPUs in three days, according to The Decoder; the claim is not independently verified. The Decoder reports that Anthropic described the work in an investor presentation, saying its frontier model, Mythos, ported or wrote nearly 100 percent of the required kernels with almost no human intervention. The presentation contrasts roughly three days to reach usable performance with six to twelve months under traditional engineering approaches for onboarding new accelerators. The reported figures concern prompt processing by an Anthropic model on AMD MI355X GPUs. That scope matters: the example is evidence of a claimed improvement in a particular workload, not proof that every stage of model training and inference can move between chip platforms in three days. According to the report, Anthropic co-founder and Chief Compute Officer Tom Brown said the company now trains and runs Claude across four major chip families: AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs. Brown described the multi-chip strategy as an early, unconventional investment that required years of work on platform-specific software and inference stacks. The distinction between that sustained investment and the three-day result is important. The account does not establish that a team without Anthropic’s existing infrastructure and experience could reproduce the same turnaround. AI-assisted porting may reduce the incremental work of adding hardware without eliminating the engineering needed to maintain a multi-platform operation. Nvidia’s CUDA software ecosystem is central to the competitive argument. As The Decoder explains, the cost and complexity of adapting AI software to alternative hardware can discourage companies from switching suppliers. If Anthropic’s figures hold up, automated kernel development could reduce one component of that switching cost. That would not, by itself, establish that AMD hardware is faster or cheaper for every workload. Porting speed, achieved performance and total operating cost are different measures. The supplied report describes a usable-performance milestone but does not provide enough benchmark detail to compare end-to-end economics across the four platforms. Anthropic says it treats compute across vendors, architectures and locations as a fungible pool serving research, model development, customer demand and internal Claude use. The practical aim is greater flexibility: the ability to assign work across available hardware rather than depend on a single supplier. The reported AMD example supports that strategy’s software-portability argument, but does not demonstrate that every chip is interchangeable for every task. The presentation also describes “compute multipliers”—research advances intended to extract more capability from the same physical computing resources. Anthropic’s longer-term ambition is for Claude to discover such improvements itself. The Decoder notes that recursive self-improvement remains unproven; this ambition should therefore be distinguished from the reported porting result. Separately, Anthropic is pursuing an “Anthropic Direct” program under which it leases chips and operates them in purpose-built data centers, according to the report. Its stated goal is to reduce total cost of ownership rather than purchase all computing capacity as a cloud service. Taken together, the claims describe an effort to make hardware sourcing more flexible and software adaptation less labor-intensive. They do not establish that Nvidia’s CUDA advantage has disappeared. The central unresolved question is how broadly the reported three-day result applies beyond the specified AMD prompt-processing workload, and whether comparable results can be independently verified.