Anthropic tells investors Claude Mythos Preview got its AI running on AMD GPUs in ~3 days with "basically no human intervention", vs. the usual 6-12 months
First reported by Tomshardware ·
The cost of porting AI models to new hardware drops from months of engineer time to days, making it easier to use chips from competitors like AMD.
Anthropic informed its investors that its Claude Mythos Preview AI model successfully adapted to run on AMD GPUs in approximately three days, requiring minimal human intervention. This process, which previously took between six to twelve months, involved Mythos autonomously writing or porting nearly 100% of the necessary kernels. This achievement is significant because it directly challenges Nvidia's long-standing CUDA moat, the proprietary software ecosystem that has historically tied AI models to Nvidia's hardware. Anthropic's co-founder and chief compute officer, Tom Brown, highlighted that the prior six-to-twelve-month process involved dozens of engineers. Anthropic currently operates its AI models across a variety of hardware, including Nvidia GPUs, AMD GPUs, Google TPUs, and AWS Trainium, managing them as a single, interchangeable compute resource. This multi-chip strategy was an early, unconventional decision that now appears prescient, especially as platforms like Google's TPUs increasingly support native AI model execution. The company has a significant agreement with AMD for up to 2 gigawatts of power, indicating a deepening partnership.
Anthropic's breakthrough with Claude Mythos Preview directly attacks Nvidia's dominant position in the AI hardware market by automating the complex process of optimizing AI models for specific GPUs. For years, Nvidia's CUDA platform has acted as a significant barrier to entry for competitors, demanding specialized engineering effort to adapt models for their hardware. By autonomously generating the low-level code, or kernels, necessary for AMD GPUs, Anthropic is demonstrating a path to drastically reduce this friction, potentially democratizing AI hardware access and increasing competition.
This development signals a potential erosion of Nvidia's CUDA moat, a key competitive advantage. If other AI models can achieve similar rapid onboarding onto non-Nvidia hardware, it could lead to greater hardware choice and price competition for AI infrastructure. Companies like AMD, which have been investing heavily in their ROCm platform as a CUDA alternative, stand to benefit significantly if this trend accelerates, potentially reshaping the landscape of AI hardware development and deployment.
AI-written summary. May contain errors.