Listen to this post: Anthropic Says Claude Ported Itself to AMD Chips in Three Days

Last updated: 17 November 2026. Figures below are sourced to primary publications, linked inline.
The 60-second version
- In its IPO roadshow video, published 10–11 November 2026, Anthropic says an internal system called “Mythos” ported Claude onto AMD Instinct MI355X GPUs in “about three days, with basically no human intervention.”
- The claim comes from Tom Brown, Anthropic’s co-founder and chief compute officer, speaking directly to camera for investors — not from a technical paper, benchmark or engineering blog post.
- It builds on a real, already-public deal: AMD’s 22 July 2026 announcement of up to 2 gigawatts of AMD Instinct MI450-series GPUs for Anthropic, alongside an AMD equity investment of up to $5 billion.
- Coverage has run from “NVIDIA’s CUDA moat is dead” to “this proves almost nothing.” Neither is quite right — the claim covers one inference workload on hardware Anthropic already used, not a full migration.
- No independent benchmark, repro steps or kernel-level detail has been published by Anthropic, AMD or anyone else to verify the headline figure.
Key dates and figures
| Date / figure | What it is | Source |
|---|---|---|
| 22 July 2026 | AMD and Anthropic announce a strategic partnership: up to 2GW of AMD Instinct MI450-series GPUs in Helios systems, plus an AMD equity investment of up to $5bn in Anthropic | AMD press release (primary) |
| H1 2027 | Target start date for the first gigawatt of that MI450 deployment | AMD press release |
| Undated, pre-November 2026 | Anthropic’s “existing use” of AMD Instinct MI355X GPUs, which the July deal explicitly builds on | AMD press release |
| 10–11 November 2026 | Anthropic’s IPO roadshow video goes live, containing the Mythos/AMD claim | Anthropic IPO roadshow video, via Simon Willison’s transcript |
| “About three days” | Claimed time for Mythos to get Claude running on MI355X GPUs at what Anthropic calls “a useful level of performance,” versus a typically cited 6–12 months | Self-reported by Anthropic; 6–12 month baseline appears only in secondary press coverage, not confirmed in the roadshow transcript itself |
What Anthropic actually said
The claim surfaces in the “Approach to compute” section of Anthropic’s IPO roadshow video, which the company published on its own investor site the same week it launched Claude Meetings, Whiteboards and Sheets to coincide with the roadshow’s opening. Tom Brown, describing Anthropic’s multi-chip strategy, says: “This time, we had Mythos try it out. It got Claude running on AMD in just three days, with basically no human intervention,” adding that “Claude ported or wrote almost 100% of the kernels.” The company has already drawn attention for using Claude on unusual internal engineering work — in September it had Claude spend 11 days formalising Fermat’s Last Theorem in the Lean proof language, a similarly striking, similarly hard-to-independently-verify claim about what its models can do unsupervised.
What’s notably absent is any accompanying technical detail. There is no engineering blog post, no benchmark table, no kernel-level breakdown, and — as far as we can establish — no video transcript published officially by Anthropic itself; the most complete public transcript currently circulating was produced independently by developer and commentator Simon Willison using transcription software, not released by Anthropic. That matters, because the roadshow video is a sales document aimed at institutional investors ahead of a share sale, not a peer-reviewed or independently audited result.
The deal underneath the claim is real and well documented, though. AMD’s own July announcement confirms the hardware relationship: up to 2 gigawatts of Instinct MI450-series GPUs in AMD’s Helios rack-scale systems, EPYC “Venice” CPUs, Pensando networking and ROCm software, with the first gigawatt targeted for the first half of 2027. AMD chief executive Lisa Su called it a deepening of the companies’ partnership “at gigawatt scale,” and Brown said in the same release that “access to compute is central to keeping Claude at the frontier.” Separately, the release notes the new deployment “builds on Anthropic’s existing use of AMD Instinct MI355X GPUs” — the same chip family Brown says Mythos ported Claude onto in three days.
Why “three days” matters more than the three days themselves
The number is less interesting than what it’s being used to argue. Switching a large language model’s training and inference stack between GPU vendors has historically been one of the most painful, specialised jobs in AI infrastructure — not because the chips are radically different, but because Nvidia’s CUDA software ecosystem has spent nearly two decades accumulating libraries, compiler toolchains, debugging tools and millions of lines of hand-tuned code that assume Nvidia hardware underneath. AMD’s equivalent, ROCm, has closed much of that gap on paper but has lagged in practical maturity and the sheer weight of existing software written against CUDA. Re-platforming a model like Claude onto AMD chips has typically meant rewriting or re-validating custom kernels, debugging numerical mismatches, and tuning for an unfamiliar memory and networking topology — the “6–12 months” figure that circulated in press coverage, even though we could not find it stated directly in Anthropic’s own roadshow remarks.
If an AI system really can do a meaningful slice of that work largely unsupervised in days rather than months, the implication isn’t just “Anthropic likes AMD now.” It’s that the single biggest practical barrier to multi-vendor AI infrastructure — engineering cost — might be falling fast, for reasons that have nothing to do with AMD’s chips becoming better and everything to do with AI models becoming better at writing and validating low-level systems code. That’s a genuinely new argument, distinct from the usual “AMD vs Nvidia” chip-spec comparisons that dominate most coverage of this story.
It also lands at a pointed moment for Nvidia. The company is still absorbing the fallout from Jensen Huang’s resignation as chief executive amid espionage allegations, and investors parsing every signal about whether Nvidia’s dominance is structural (the CUDA ecosystem) or circumstantial (first-mover advantage and supply) have reason to pay close attention to any claim that erodes the former.
The CUDA moat, briefly explained
“Moat” in this context means switching costs, not raw chip performance. Nvidia’s advantage is less about any single GPU’s specifications and more about the surrounding software: CUDA’s libraries for linear algebra and communication, the debugging and profiling tools built around it, and — critically — a developer base numbering in the millions who have learned to write and optimise code for it. A faster AMD chip is only useful if the software running on it is equally fast, and getting software there has traditionally required specialist engineers rewriting code by hand. An AI system capable of doing that work itself, even partially, is a direct attack on switching costs rather than on chip specifications — which is exactly why this claim, even unverified, unsettled commentators who otherwise agree on little else.
What the coverage is getting wrong
Some outlets have run with headlines declaring Nvidia’s CUDA moat effectively dead on the strength of one investor-presentation anecdote. That overstates it. By Anthropic’s own account, as reported in press coverage of the roadshow, the three-day port covered “prompt processing for one Anthropic model… at a useful level of performance” — inference, not training, on hardware (MI355X) Anthropic was already running, not a cold start on unfamiliar silicon. It says nothing about multi-node training clusters, Nvidia’s NVLink-dependent networking assumptions, or the thousands of custom operators in a full production stack. Nvidia’s own published figure of more than 7.5 million developers in its Developer Program — a figure broader than CUDA specifically, and not an independently audited count — did not shrink because of one weekend’s inference benchmark on one rack.
The opposite overcorrection is also wrong: dismissing this as meaningless marketing misses that Anthropic is simultaneously backing the claim with real capital — its AMD deal, its continuing Nvidia-based $35bn Lambda compute arrangement, and a multi-chip strategy disclosed across S-1 filings that now totals roughly $895bn in compute commitments. Companies rarely back an unsupported marketing claim with multi-billion-dollar, multi-year contracts with the vendor the claim is meant to make irrelevant. The honest reading sits between the extremes: a real, strategically significant data point about AI-assisted infrastructure work, wrapped in investor-presentation framing that nobody has yet backed with publishable technical evidence.
Practical takeaways for people who build and ship software
For engineering teams actually making GPU-vendor decisions today, a few things follow from this that don’t depend on resolving the “is CUDA dead” argument:
First, don’t plan infrastructure strategy around a single investor-deck anecdote. Until someone publishes benchmark data, repro steps or kernel-level detail, treat the three-day figure as a claim worth tracking, not a capability to rely on for your own migration. Your own CUDA-to-ROCm bring-up time will depend on your specific custom kernels, testing rigour, and how much of your stack already sits on vendor-agnostic frameworks.
Second, it’s a reasonable signal to start experimenting, cautiously, with AI coding agents on lower-risk porting and compatibility work — not frontier training infrastructure, but internal tooling or non-critical services where a faster, AI-assisted port is worth trying even if it only gets you most of the way there. The direction of travel — AI doing more of the grunt work of systems-level portability — looks durable even if this particular figure proves generous.
Third, treat vendor lock-in as something that may be falling, not something that has already fallen. If your infrastructure planning assumes you’re stuck with one GPU vendor for multi-year horizons, revisit that periodically rather than treat it as settled — while still budgeting for the real engineering work a migration requires today.
What we still don’t know
- Whether “about three days” is a precisely measured duration or a rounded figure chosen for an investor audience.
- What “a useful level of performance” means in concrete throughput, latency or cost-per-token terms relative to Anthropic’s Nvidia-based baseline.
- Whether the port covered only single-rack inference, as several analysts assume, or something closer to production-representative load.
- How much human engineering review, debugging or correction happened around the “basically no human intervention” claim — the phrase leaves room for substantial human oversight that simply wasn’t hands-on coding.
- Whether this generalises beyond the specific model and workload tested, including to training workloads, which are far more sensitive to interconnect and numerical precision issues than inference.
- Whether Anthropic intends to publish supporting technical detail, given that the claim currently exists only in investor-facing material rather than in any engineering-facing venue.
FAQ
What is “Mythos”?
Claude Mythos is a real, named Anthropic model family, first disclosed in April 2026 — a restricted, cybersecurity-focused model given to vetted partners through Anthropic’s “Project Glasswing” vulnerability-scanning program, and briefly subject to US export controls in June 2026 before access was restored. What’s unconfirmed is narrower than the FAQs on this story sometimes suggest: whether the system Tom Brown describes using to port Claude onto AMD chips is that same restricted model, a newer version of it, or a differently scoped internal tool that merely shares the name. Anthropic’s roadshow remarks don’t spell out the connection.
Did an AI really port Claude to AMD chips with no human help?
That is what Anthropic’s chief compute officer said in an investor presentation: “basically no human intervention.” It’s a self-reported claim from a sales-oriented roadshow video, not an independently verified or peer-reviewed result, and no technical writeup has accompanied it.
Does this mean Nvidia is in trouble?
Not on this evidence alone. The claim covers one inference workload on one model, using GPUs Anthropic already had access to. It doesn’t address training workloads, multi-node scaling, or the depth of Nvidia’s existing software and developer ecosystem. It’s a data point, not proof of a shift already underway.
Is Anthropic actually moving away from Nvidia hardware?
There’s no sign of that. Anthropic continues to disclose large Nvidia-linked compute commitments in its public S-1 filings alongside its AMD, AWS Trainium and Google TPU arrangements. AMD itself frames the Anthropic deal as additive capacity rather than a replacement for other chip suppliers.
Sources
- AMD and Anthropic Announce Strategic Partnership — AMD newsroom, 22 July 2026 (primary)
- Anthropic’s IPO roadshow video — Simon Willison, 11 November 2026, with independently produced transcript
- Anthropic Says Its AI Model Cut AMD Chip Switch to Three Days — International Business Times, 12 November 2026
- Anthropic Jeopardizes NVIDIA’s CUDA Moat — Wccftech, 11 November 2026
- Claude Mythos / AMD MI355X Port: Claimed vs Verified — Explainx, 11 November 2026
- Anthropic’s Model Taught Itself to Run on AMD in “Three Days” — Yahoo Finance / TIKR, 12 November 2026
