DEV Community

News
News

Posted on

Anthropic says AI ported Claude to AMD GPUs in three days, challenging Nvidia's CUDA moat

Infographic cover

๐Ÿ“… November 11, 2026 โ€ข Anthropic via The Decoder โ€” Anthropic says it now trains and serves Claude across four major chip families โ€” AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs โ€” treating them as one interchangeable pool. The claim sits in the approach-to-compute chapter of its investor presentation, quoted by Simon Willison and The Decoder, and comes from co-founder and Chief Compute Officer Tom Brown.

The striking part is AMD: this year Anthropic worked with AMD GPUs for the first time and used its own frontier model Mythos, and Claude was running on AMD Instinct MI355X GPUs in about three days, with almost no human intervention, per Anthropic claims reported by The Decoder. The company says Mythos ported or wrote nearly 100 percent of the kernels, the low-level code that runs a model on a chip. Previously, bringing up a new chip platform kept dozens of engineers busy for six to twelve months.

The balance matters here: the figures come from Anthropic's own investor presentation and have not been independently verified, The Decoder stresses. The prospectus itself uses guarded language, saying chip platforms now come into production in a fraction of the time. The counterpoint comes from Wccftech, which argues it is premature to declare the CUDA moat dead โ€” it remains a massive ecosystem of libraries, tools, years of optimization and developers.

The implication, as The Decoder frames it, is pricing power: Nvidia's CUDA software ecosystem has been the costliest barrier to switching chips, and Nvidia earns a hardware margin of more than 70 percent. If AI compresses porting from months to days, buyers can weigh price and availability more freely. For context, in July 2026 AMD and Anthropic announced a partnership to deploy up to 2 gigawatts of AMD Instinct MI450-series GPUs in Helios rack-scale systems, with the first gigawatt phase planned for the first half of 2027, per The Decoder and Cryptobriefing.

The presentation also introduced compute multipliers โ€” research breakthroughs that multiply effective compute and stack on each other, so a 4x multiplier means 25 compute units can do what needed 100, per the transcript quoted by Simon Willison. For infrastructure planners the numbers remain vendor-claimed, but the direction is explicit: chip choice is framed as a pooling decision across interchangeable families rather than a software rewrite tied to one ecosystem.

Sources:
the-decoder.com/anthropic-says-ai-is-eroding-nvidias-cuda-lead-cutting-an-amd-chip-switch-from-months-to-three-days
simonwillison.net/2026/Nov/11/anthropic-ipo-roadshow
wccftech.com/anthropic-jeopardizes-nvidias-cuda-moat-as-its-claude-mythos-model-took-only-3-days-to-figure-out-how-to-run-another-model-on-amds-mi355x-gpus
cryptobriefing.com/claude-mythos-amd-mi355x-three-days

Key figures

IMPACT checklist


Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky

AI Automation

Top comments (0)