Anthropic says its Mythos AI model helped migrate Claude to AMD graphics processing units (GPUs) in around three days, highlighting how advanced AI could speed up the engineering work required to run large language models on different chips.
The reported demonstration is significant because AI companies are competing not only to build more capable models but also to secure enough computing power to train and operate them. Moving software between hardware platforms can require extensive engineering, testing and performance tuning.
How Did Mythos Help Run Claude on AMD GPUs?
According to reports, Anthropic used Mythos to help adapt Claude for AMD’s Instinct MI355X GPUs, with limited human intervention. The work reportedly took about three days.
The process involved adapting software and optimising performance for AMD hardware. Traditionally, migrating complex AI workloads between chip architectures can take months, although timelines vary significantly depending on the software, hardware and existing infrastructure.
The demonstration suggests that AI models may increasingly assist with the engineering needed to make their own deployment more flexible. However, the reported three-day timeline should not be interpreted as proof that every AI model can be transferred between processors in the same period.
Why Anthropic Wants Multiple AI Chip Suppliers
Anthropic is expanding beyond reliance on any single chipmaker. Its computing strategy spans Nvidia GPUs, AMD GPUs, Google’s Tensor Processing Units (TPUs) and Amazon’s Trainium chips.
In its November 2026 regulatory filing, Anthropic described managing computing resources as a shared pool and directing workloads to different hardware based on price and performance. The company says its software orchestration capabilities help it allocate computing capacity more efficiently.
This flexibility matters for both AI training and inference. Training builds a model’s capabilities, while inference is the computing process used to generate responses for users. Different chips can offer different advantages depending on workload, availability, cost and software compatibility.
Anthropic’s $5 Billion AMD Investment Deal
Anthropic’s work with AMD also sits within a much larger infrastructure agreement announced on July 22, 2026.
Under the partnership, Anthropic plans to deploy up to 2 gigawatts of AMD Instinct MI450-series GPU capacity using AMD Helios rack-scale systems. Deployment of the first gigawatt is scheduled to begin in the first half of 2027.
AMD has also committed to a strategic equity investment of up to $5 billion in Anthropic. The companies will collaborate on improving AI workloads for AMD chips and accelerating development of ROCm, AMD’s software platform for GPU computing.
The planned MI450 deployment is separate from the reported Mythos-assisted migration involving existing MI355X hardware.
Can AMD Challenge Nvidia in AI Computing?
The development could strengthen AMD’s position in the AI infrastructure market, where Nvidia remains a major supplier. More flexible software could make it easier for AI developers to use alternative chips, potentially widening their options for capacity and pricing.
But switching hardware is not simply a matter of making a model run. Performance at scale, reliability, networking, software maturity and the cost of operating large systems all influence whether an alternative platform is competitive.
The bigger takeaway: Anthropic’s reported experiment points to a future in which AI helps reduce the engineering barriers between chip platforms. If the approach scales reliably, it could help AI companies expand computing capacity faster while reducing dependence on individual hardware suppliers.
