Anthropic Says Its AI Model Cut AMD Chip Switch to Three Days. Why it Could be a Challenge to Nvidia

Anthropic said its own artificial intelligence model helped get Claude running on AMD chips in three days, accelerating work that previously required months.

The claim, presented to investors ahead of Anthropic's proposed initial public offering, suggests AI coding tools could reduce the engineering burden that makes switching chip suppliers difficult. However, the three-day result covers a specific workload, rather than a complete migration of Claude's training and operations.

Tom Brown, Anthropic's co-founder and chief compute officer, said the company used its Mythos model to handle the transition after beginning work with AMD GPUs this year. According to the presentation, Mythos wrote or ported nearly all the necessary kernels, the specialized software routines that execute calculations on chips. "It got Claude running on AMD in just three days with basically no human intervention," Brown said in the investor presentation.

The approximately three-day timeline refers to bringing up prompt processing for one Anthropic model on AMD MI355X GPUs at a useful performance level. Prompt processing is the stage in which a model reads and processes input before generating its response.

The reported acceleration addresses a significant obstacle for Nvidia's competitors. CUDA provides the software infrastructure developers use to put Nvidia GPUs to work. Companies that have built their systems around that ecosystem can face substantial engineering costs when adapting their software to another supplier's hardware.

If AI can perform more of that adaptation, customers could gain flexibility to choose chips based on availability, performance and cost. Anthropic says it has spent years building that flexibility. Its IPO materials describe a computing architecture spanning four chip families: Nvidia GPUs, AMD GPUs, Amazon Web Services Trainium chips and Google's tensor processing units.

The company treats capacity across its partners as a shared pool, allocating computing resources to research, model development, customer services and internal AI use. Its investor fact sheet identifies both its multiple-chip architecture and research breakthroughs as foundations of that approach.

The strategy could become increasingly important as demand for AI computing grows. Different workloads have different requirements, and access to multiple suppliers gives Anthropic more options when adding capacity or deciding where to run particular tasks.

Anthropic also describes research advances as "compute multipliers," improvements intended to extract more capability from existing hardware. The company's longer-term ambition includes having Claude help discover further efficiency gains. That remains an objective, rather than proof of autonomous recursive self-improvement.

Alongside those efforts, Anthropic is pursuing its Anthropic Direct program, leasing chips and operating them in purpose-built data centers to lower ownership costs. Its relationship with AMD already extends beyond software experiments.

AMD announced in July that Anthropic would deploy up to two gigawatts of Instinct MI450-series GPUs in Helios systems, with the first gigawatt deployment beginning in the first half of 2027. AMD also committed to a strategic equity investment of up to $5 billion.