cd /news/ai-infrastructure/anthropic-mythos-ported-claude-to-am… · home › topics › ai-infrastructure › article
[ARTICLE · art-170361] src=byteiota.com ↗ pub= topic=ai-infrastructure verified=true sentiment=· neutral

Anthropic Mythos Ported Claude to AMD in 3 Days — CUDA Moat Update

Anthropic co-founder Tom Brown said during the company's IPO roadshow that Anthropic's frontier model Claude Mythos 5 ported Claude's prompt processing to AMD MI355X GPUs in roughly three days, autonomously writing or adapting nearly 100% of the required GPU kernels — work that previously took dozens of engineers six to twelve months. The claim covers only prefill for one model, not full inference, training or production-grade optimization, and no independent kernel benchmarks have been published. Anthropic has committed to 2 GW of AMD MI450 capacity and taken a $1 billion AMD equity stake, while Meta is deploying approximately 6 GW of MI450-series GPUs.

read4 min views2 publishedNov 16, 2026
Anthropic Mythos Ported Claude to AMD in 3 Days — CUDA Moat Update
Image: Byteiota (auto-discovered)

During Anthropic’s IPO roadshow this week, co-founder Tom Brown dropped a claim that lit up developer Twitter: Anthropic’s internal frontier model, Mythos, ported Claude’s prompt processing to AMD MI355X GPUs in roughly three days — autonomously writing or adapting nearly 100% of the GPU kernels required. A job that previously took dozens of engineers six to twelve months. If you have ever tried to run ML code on non-NVIDIA hardware and hit the CUDA wall, that number lands differently.

What Mythos Actually Did #

Mythos — Claude Mythos 5, Anthropic’s most capable frontier model — was given access to AMD’s MI355X hardware and told to get going. With minimal human guidance, it wrote or adapted the chip-specific kernels needed for Claude’s prompt processing phase: the stage where a model reads its input before it generates a response. Three days later, it was running at a useful performance level on hardware built on AMD’s CDNA 4 architecture rather than NVIDIA’s.

Keep the footnote in view: the three-day figure covers prompt processing for one model — not full inference (token generation), not training, and not production-grade optimization. Anthropic is currently on its IPO roadshow, a context in which companies tell their most compelling story. Independent kernel benchmarks have not been published. GPU code that runs is not necessarily GPU code that runs fast.

That said, even a qualified version of this claim is significant. Prefill is a real bottleneck in long-context inference stacks. And Mythos writing nearly 100% of the kernels autonomously — even for one phase — represents a new mode of chip bring-up that did not exist a year ago.

Why CUDA’s Moat Existed #

NVIDIA’s lock-in was never about the hardware. It was about two decades of software that made switching feel impossible. The dependency chain alone — your code, then PyTorch, then cuDNN, then CUDA, then the GPU — means swapping the chip requires validating every layer above it. Add six million CUDA-trained developers, hand-tuned libraries like cuBLAS and NCCL, the Nsight profiler ecosystem, and curricula at 350 universities, and you get a moat that AMD’s ROCm project has been chipping away at for years without fully closing.

The practical cost: anyone writing custom CUDA kernels — the kind used in production inference serving, attention mechanisms, quantization — faced months of porting work and an AMD debugging ecosystem that lagged NVIDIA’s. Most teams stayed on NVIDIA even when AMD hardware was price-competitive. The math never justified the migration pain.

Where ROCm Actually Stands in 2026 #

ROCm has gotten meaningfully better. ROCm 7.2.4 now supports PyTorch 2.7.0 as a first-class backend. vLLM and SGLang both have official ROCm support. AMD is a core PyTorch contributor. The bigger signal is production: Meta is deploying approximately 6 gigawatts of AMD’s MI450-series GPUs. Anthropic committed to 2 GW of MI450 capacity and took a billion AMD equity stake. ROCm is validated at hyperscaler scale in 2026 in a way it was not in 2024.

Cloud-side, AMD MI355X instances start at roughly .95 per GPU-hour at TensorWave — meaningfully below NVIDIA pricing for comparable inference workloads. For teams running standard PyTorch or vLLM stacks where custom kernel work is minimal, the case for benchmarking AMD is stronger than it has ever been.

What Developers Should Do Now #

  • If your stack is PyTorch, vLLM, or SGLang: ROCm first-class support is real. Run an inference benchmark on AMD cloud against your current NVIDIA spend. The gap has closed enough to be worth the hour it takes.
  • If you write custom CUDA kernels: AMD’s HIP API mirrors CUDA closely. AI-assisted porting via Claude Code or similar is now viable for non-trivial kernel work — the Mythos story is partly a proof of concept for this workflow.
  • Watch OpenAI’s Triton: Theopen-source compiler writes kernels that run on CUDA, ROCm, and TPUs . If Triton continues to mature, the per-chip kernel problem becomes someone else’s problem.
  • At the API level: Use an OpenAI or Anthropic wire-compatible abstraction for inference routing. Tools like OpenRelay already route across NVIDIA GPUs, AMD GPUs, TPUs, and Trainium through a single API. That is the cheapest hedge against chip lock-in.

The Bigger Picture #

The story beneath the story is not “CUDA is dead.” NVIDIA’s ecosystem will not be replaced by a three-day chip port on a roadshow video. The story is that AI is now doing the work of chip porting, and that changes the economic calculus of the CUDA moat. If Mythos can bring up a new chip platform in days rather than months, the switching cost that has kept teams locked to NVIDIA is no longer a fixed engineering expense — it is a task that can be delegated to a model. That is a different kind of moat erosion, and worth taking seriously regardless of the IPO footnotes.

── more in #ai-infrastructure 4 stories · sorted by recency
── more on @anthropic 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/anthropic-mythos-por…] indexed:0 read:4min 2026-11-16 · —