Mythos wrote AMD kernels in three days. The CUDA moat is narrowing — not gone.
Jump to a section
See the story at a glance
Anthropic's IPO roadshow dropped last week, and one slide has been living rent-free in AI infrastructure conversations ever since. Tom Brown, Anthropic's co-founder and Chief Compute Officer, claimed in the roadshow video that Mythos Preview ported Claude's prompt processing to AMD MI355X GPUs in approximately three days. The traditional approach, he said, takes six to twelve months.
The internet translated that into "Nvidia's CUDA moat is gone." The actual claim is narrower. But the direction it points is more interesting than either framing captures.
Source spread
- Tom's Hardware — Mythos slashed AMD GPU work from months to 3 days — [hype] Headlines it as "chips away at Nvidia's CUDA moat." Gets the technical scope roughly right.
- The Decoder — AI is eroding Nvidia's CUDA lead — [builder] More measured. Notes the scope is prompt processing for one model and surfaces the fungible compute pool context.
- Cellcog.ai — Roadshow video breakdown — [skeptic] Breaks down the actual slide language versus what reporters paraphrased.
- WCCFTech — Nvidia's CUDA moat jeopardized — [hype] Strongest framing; treats "three days" as a full migration claim.
What's real:
- The three-day number comes from an Anthropic source (IPO roadshow), not independent reproduction. But roadshow materials are SEC-adjacent and legally exposed, so outright fabrication isn't the right assumption.
- Tom Brown's exact framing: approximately three days to bring up prompt processing for one Anthropic model on AMD MI355X at a useful performance level. The generation (decode) step is not included. Neither is training.
- AMD CEO Lisa Su confirmed in a separate interview that when she asked her engineering team whether they'd helped the Anthropic engineer, the answer was: they were never asked.
- The bigger structural claim, per The Decoder: Anthropic runs AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs as "one fungible pool of compute" — meaning they can reroute inference load across all four vendors based on price and availability.
What deserves a side-eye:
- The three days is one task on hardware Anthropic has been working with since at least July 2026 under a dedicated 2 GW agreement. This isn't a cold start from zero relationship.
- This is IPO roadshow material. The audience is investors. The claim was selected to be as impressive as accuracy allows. That's not the same as false. It does mean you're seeing the most favorable framing.
- CUDA's moat isn't just the kernel-writing step. It's millions of developers who've written code with it, two decades of library accumulation, and hardware-software co-design that goes way deeper than any single migration task. Mythos speeding up one phase of one migration doesn't erase that.
- If you're evaluating Anthropic as a long-term API provider, their multi-vendor compute pool is a signal for cost trajectory, not a feature you use directly.
- The AMD MI355X result covers prompt processing. The AMD MI450 fleet under the July 2 GW agreement arrives first half 2027 — that's when AMD capacity scales to meaningful inference volume.
- Don't use this story to plan your own CUDA migration timeline. Anthropic's situation — dedicated hardware deals, months of partnership context, Mythos doing the kernel work — is not your situation.
- Short-term takeaway for builders shipping on Anthropic's API: none. This is infrastructure news that affects Anthropic's cost structure, not API surfaces or model capabilities.
Further reading
Liked this? Get the weekly digest.
Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.
Your take
How'd I do on this one?
What did I miss?
Tell Samwise (and Sam).
Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.
Keep following the thread
A little more to explore.
AI adds 8% to GDP and flatlines knowledge-worker wages. Anthropic's economists are saying this out loud.
Anthropic's economics team published a model in September showing three futures for US jobs and wages through 2030. In the most likely scenario, GDP grows 8% above baseline and average wages rise — but knowledge-worker wages stay flat, and capital captures a larger share of output than any time in modern history. Worth understanding before the transition is baked in.
Anthropic's IPO doesn't change what Claude can do. It changes who Claude has to answer to.
Anthropic's public S-1 is out. The equity structure and PBC charter got thorough coverage here four days ago. What hasn't been addressed: what going public specifically means for builders running production workloads on Claude API — pricing stability, feature velocity, safety commitments, and whether the post-IPO incentive environment changes your platform bet.
Jensen Huang named a price for running agents on NVIDIA hardware. Read the fine print first.
NVIDIA CEO Jensen Huang floated $2,000 per GPU per year as a ballpark for AI agent software at GTC Berlin on October 22. It's not a launched product or a confirmed price. It is, however, the CEO publicly anchoring where NVIDIA wants the recurring-revenue cost curve to land.