TUE, NOVEMBER 17, 2026
Independent · In‑Depth · Practitioner‑Tested
✎ AI News

Three Days, Not Six Months - Anthropic's Claim About AMD and the CUDA Moat

In a roadshow video dated 10 November, Anthropic told IPO investors that Claude Mythos Preview did the software work to run one of its models on AMD MI355X GPUs in about three days, against a historical baseline of six to twelve months with dozens of engineers. The scope is prefill for one model, and the claim is unaudited.

By AIToolsRecap November 17, 2026 5 min read 13 views
Home › Articles › AI News › Claude › Anthropic Says Claude Mythos Ported a Model to ...
Straight answer
● The claim Anthropic told IPO investors that Claude Mythos Preview did the software work to run one of its models on AMD MI355X GPUs in about three days.
● The baseline Anthropic says the same job has historically taken six to twelve months with dozens of engineers.
● What it covers Prefill - prompt processing - for one model. Mythos ported or wrote close to 100% of the kernels.
● Where it came from A roadshow video by co-founder and chief compute officer Tom Brown, dated 10 November, alongside the amended prospectus filed the same day.

The number that matters is not three days. It is the gap between three days and six to twelve months, because that gap is the practical width of NVIDIA's CUDA moat. Porting a frontier model to non-NVIDIA silicon has been expensive in exactly one currency - senior engineer time - and Anthropic is telling investors that a model now does most of that work itself.

Read the scope carefully before drawing conclusions. This is prefill for a single model, not a full training and inference stack, and "close to 100% of the kernels" is a claim about code written rather than performance achieved. Anthropic has not published throughput comparisons against the equivalent NVIDIA deployment, and nobody outside the company has reproduced it. The context also matters: this appeared in a roadshow video and an amended S-1, which are documents written to sell shares. That does not make it untrue - statements in a prospectus carry legal weight - but it is a company describing its own result, not an audited benchmark.

It sits inside a larger AMD relationship. Anthropic has an agreement with AMD covering up to 2 gigawatts, with the first MI450-based Helios racks expected in the first half of 2027. If the porting cost really has collapsed, the strategic read is that frontier labs gain leverage over their silicon suppliers rather than that AMD beats NVIDIA - being able to move cheaply is worth more at the negotiating table than actually moving. NVIDIA is also one of seven companies buying $26 billion of Anthropic stock at the IPO price, which is its own comment on how the two companies see each other.

FAQ

How long did the AMD port actually take?
Anthropic says about three days for Claude Mythos Preview to do the software work, against a historical baseline it gives as six to twelve months with dozens of engineers. The three days cover prefill for one model.
Does this mean AMD now competes with NVIDIA for AI training?
Not from this alone. The claim is about the cost of porting software, not about the resulting performance. No throughput comparison against an equivalent NVIDIA deployment has been published.
Where was this announced?
A roadshow video by Tom Brown, Anthropic co-founder and chief compute officer, dated 10 November 2026, and the amended prospectus filed the same day ahead of the company's IPO.
Has anyone verified it independently?
No. It is a first-party claim in an offering document. Statements in a prospectus carry legal weight, but no external party has reproduced the result.
Tags
AI NewsAnthropicNvidia2026

Spot an inaccuracy?

We verify facts before publishing and correct errors promptly. If something in this article is wrong or outdated, let us know.

Report an error →