Anthropic says it has moved Claude onto AMD GPUs in three days, with its Mythos model writing or porting nearly 100 percent of the required kernels. The company presents the work as evidence that AI could make it easier to run models across different chip platforms.
Why chip software matters
Nvidia’s strength comes from more than its GPUs. Its CUDA software ecosystem gives developers tools built around Nvidia hardware, while moving AI workloads to another architecture requires a different software and inference stack.
That work can take time and engineering resources. Anthropic says that, until now, bringing up a new chip platform kept dozens of engineers busy for months. The effort can discourage companies from trying alternatives, even when another supplier’s chips might suit their needs.
Anthropic says it chose a multi-chip strategy early, despite the costs of adapting software for each platform. Tom Brown, the company’s co-founder and Chief Compute Officer, says Anthropic has invested in this work for years and is now “incredibly good” at it.
Mythos takes on the porting work
The company says it now trains and runs Claude on all four major chip families, treating compute from different vendors, architectures, and locations as one pool. Its AMD test is presented as a particularly clear example of how AI might reduce the work involved in expanding that pool.
Anthropic says Mythos handled nearly all the kernel writing or porting when Claude was moved to AMD GPUs. The company reports that the process took three days, compared with the six to twelve months usually required for a new chip platform.
If the reported timeline and level of automation hold up, AI-assisted porting could lower one barrier to choosing hardware. A company might find it easier to shift work toward chips that are more affordable or available, rather than staying with one vendor because changing platforms is too burdensome.
What a more flexible compute pool could mean
Different chips may be better suited to different jobs, including training or inference. Anthropic argues that using multiple platforms lets it pursue a stronger price-performance ratio as compute supply remains tight.
The claim also speaks to Nvidia’s position in the market. The source article says Nvidia currently enjoys a margin of more than 70 percent, unusually high for hardware, because of its dominance in AI chips. If AI reduces the software effort required to switch, that advantage could face more competition.
Anthropic says it gets more out of every dollar of compute than any other frontier AI lab. That is the company’s own assessment, rather than an independently established comparison. Its broader strategy also includes “compute multipliers”: research breakthroughs intended to improve performance using the same physical computing capacity.
Claims still need independent verification
Anthropic says its compute multipliers build on one another, with each breakthrough amplifying earlier gains. The company eventually wants Claude to discover such improvements on its own. That possibility would amount to recursive self-improvement, a concept debated in AI research that has not been proven.
The company is also expanding its direct data center capacity through the “Anthropic Direct” program. Rather than relying only on cloud computing services, it leases chips and operates them in purpose-built data centers, with the stated aim of reducing total cost of ownership.
The porting figures come from Anthropic’s investor presentation and have not been independently verified. They show how the company sees AI as a way to loosen reliance on a single hardware ecosystem, but the reported three-day result remains a company claim.