Anthropic has disclosed that its Claude Mythos model can move part of its AI workloads onto non-Nvidia chips in a few days. That job used to take six to 12 months, which puts pressure on the CUDA lock-in behind Nvidia's dominance.
Anthropic says Claude Mythos can move part of its AI workloads onto non-Nvidia chips in just a few days, with basically no human intervention. The same job used to take six to 12 months and dozens of engineers.
Porting to AMD chips takes about three days
The disclosures came as Anthropic prepares for its IPO. The headline capability is software that rewrites and tunes AI models to run on different hardware, largely on its own.
Its specific benchmark is AMD's MI355X GPUs. Claude Mythos can port an AI model to that chip in approximately three days, compared with the old norm of 6-12 months.
This matters because Nvidia's lead has never rested on silicon alone. Its CUDA software ecosystem is the real lock-in, since developers built their tools around it. When code can move between hardware platforms cheaply, chips start behaving less like proprietary platforms and more like interchangeable parts.
The AMD deal and Nvidia's shrinking share
Anthropic is not just talking about alternatives. Its deal for AMD's Instinct GPUs is projected to reach multi-gigawatt scale, with the first phase set for the first half of 2027.
Price is a big part of the appeal. AMD reportedly offers a per-gigawatt system cost that is roughly half that of some Nvidia platforms.
Broader market data points the same way. Nvidia's share of contracted future AI compute capacity for frontier labs has declined to around 39%, down from previous highs of 80%+.
Who gains and what limits the shift
For Nvidia, the risk is less about losing sales tomorrow and more about losing pricing power over time. Alphabet's TPUs could benefit when customers can move models onto them without a costly rewrite, and TSMC may gain as well because it manufactures chips for Nvidia, AMD, and Broadcom.
However, there are limits to the commoditization story. Anthropic's results cover part of its workloads, not all of them, and porting a model is different from running it at peak efficiency across every task. Nvidia also still holds a large installed base and deep relationships with nearly every major AI buyer.
Source: Crypto Briefing
Trading involves risk.