Anthropic says Claude ported itself to AMD GPUs in three days

Anthropic says it used its own model to move Claude onto a new chip vendor in days. The Decoder reported on 11 November 2026 that the company's investor roadshow presentation claims its frontier model Mythos got Claude running on AMD GPUs in three days, with almost no human intervention, against the six to twelve months such a port usually takes. Anthropic's preliminary prospectus, dated 10 November, makes the same point without the number: by using Claude, it says, its engineers can now bring new chip platforms into production more quickly.

What Anthropic put in writing

The prospectus says Anthropic trains and serves Claude natively across four chip families: AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs. A proprietary orchestration layer routes each workload to whichever hardware gives the best price and performance, and the company manages all of it as what it calls a single, fungible pool. The same pool serves customers, research and Anthropic's internal use of Claude.

On porting, the filing says that bringing a new chip platform into production historically required intense engineering effort over many months. It credits the speed-up to two things: Claude, and a foundation built over years that includes tooling to test and measure performance against Anthropic's real workloads. The filing does not say how long a port takes now.

What the roadshow adds

The Decoder's account of the presentation supplies the detail. It attributes the remarks to Tom Brown, Anthropic's co-founder and Chief Compute Officer, and reports four claims:

• Anthropic used AMD GPUs for the first time this year.
• Bringing up a new chip platform previously occupied dozens of engineers for months.
• Mythos ported or wrote nearly 100% of the kernels, the low-level routines that run a model on a given chip.
• The figures refer to prompt processing of an Anthropic model on AMD MI355X GPUs.

The Decoder notes the figures come from the investor presentation and have not been independently verified.

Why a team outside a lab should care

NVIDIA's hold on AI computing rests partly on CUDA, its software layer: code written for it does not run elsewhere without rework, and that rework has been expensive enough that most companies never attempt it. If a coding agent can do most of the rewrite, switching chips becomes a question of price and supply. That matters to anyone who self-hosts models or negotiates GPU capacity, and it is one more data point on whether coding agents can take on low-level migration work.

It also bears on what Claude costs. Anthropic's investor page reports a more than 4x gain in compute efficiency since the first quarter of 2025, which it measures as revenue divided by compute and infrastructure expenses. The prospectus says revenue grew 1,127% year over year in the quarter to 30 September 2026 while those expenses grew 382%.

The catch

This is a company describing itself to prospective shareholders ahead of a listing it expects on 19 November. The sources do not give throughput on AMD relative to NVIDIA, the share of production traffic AMD hardware carries, how much human verification the port needed, or what the three days cost in compute. The prospectus itself ties the result to years of earlier investment in test infrastructure, which a team starting from nothing does not have.

Anthropic also has commercial ties to the vendor involved. The prospectus describes a hardware supply agreement signed with AMD in July 2026 under which AMD committed to buy up to $5.0 billion of Anthropic stock, and says AMD is buying $2.5 billion of shares in a private placement alongside the IPO. And one figure cuts the other way: the filing says $5.0 billion of short-term compute, which costs significantly more than multi-year capacity, reduced revenue per dollar of compute in the latest quarter.

If you have a port you have been avoiding, the practical test is small: give an agent one hot path, a benchmark from your real workload and a correctness check, and time it.