Anthropic says Claude Mythos Preview cut AMD MI355X onboarding from up to 12 months to around three days

AI & Compute

Anthropic tells IPO investors Claude Mythos cut AMD MI355X onboarding to three days

Anthropic says Claude Mythos Preview brought up prompt processing on AMD MI355X in about three days, work it estimates once took six to 12 months and dozens of engineers.

By
Grace Kim
Filed
Channel
AI & Compute
Read
3 min read

By Grace Kim

Anthropic has told IPO investors that its Claude Mythos Preview model brought one of its AI models onto AMD MI355X GPUs in about three days, work the company estimates previously took six to 12 months. The claim, made in a Nov. 10 roadshow video, chips away at the software barrier that has protected Nvidia's CUDA ecosystem for nearly two decades.

The numbers carry caveats. A slide footnote limits the three-day timeline to prompt processing — the prefill step where a model reads input before generating a reply — for one Anthropic model on MI355X GPUs "at a useful level of performance." Anthropic has not said whether decode, the generation step, was included. It also has not disclosed how fast the model ran on AMD silicon versus Nvidia's.

What did Mythos actually do?

Anthropic said Claude "ported or wrote almost 100% of the kernels" with "basically no human intervention." Kernels are the low-level code that ties a model to a specific chip, and writing them has long been part of Nvidia's moat. The six-to-12-month estimate comes from Anthropic's "historical experience bringing up new accelerators using traditional engineering approaches."

Tom Brown, Anthropic's co-founder and chief compute officer, said the original effort required a team of "dozens of engineers." He had already shown a preview at AMD's Advancing AI event on July 23, describing one engineer who asked Claude to bring up an MI355X rack and left it running "over the weekend."

Anthropic runs its models on Nvidia and AMD GPUs, Google's TPUs and AWS Trainium, managed as "one fungible pool of compute." Brown called the multi-chip strategy "an unconventional bet." The company signed an agreement with AMD for up to 2 gigawatts and started working with AMD GPUs for the first time this year.

How have AMD and Nvidia responded?

AMD has argued for more than a year that AI can write this code. In August 2025, AMD researchers published GEAK, an agent that solved 19 of 30 expert-written kernels at 0.92 times the experts' average speed. CEO Lisa Su said on the Possible podcast in September 2025 that CUDA code needs to be "translated to AMD" and "a ton of that can now be done in software."

Anush Elangovan, AMD's vice president of AI software, told EE Times in April: "Claude is better than HIPify because it has web search built in." HIPify is AMD's tool for converting CUDA code to its ROCm stack. Vamsi Boppana, senior vice president of AI, said in July that AI-generated kernels were in some cases "better than the most manually tuned versions." Andrew Dieckman, corporate vice president of the data center GPU business, said he has had almost zero customer conversations about CUDA.

Nvidia CEO Jensen Huang pushed back in an April interview with Dwarkesh Patel: "We use a ton of AI to create the kernels that we have," and "nobody knows our architecture better than we do." He called Anthropic "a unique instance, not a trend." Nvidia and AMD are among seven companies buying $26.0 billion of Anthropic stock at the IPO price, so Nvidia is investing in the company Huang called the exception.

Does this settle the CUDA moat debate?

Not yet. Until this week, the boldest claims came from AMD's own stage. Anthropic buys chips rather than selling them, so its account carries more weight. Its prospectus states, without a figure: "We can onboard new chip platforms in a fraction of the time it once took."

SemiAnalysis, in a July article titled "Can AMD break the CUDA Moat?," concluded software creation is no longer the moat it was a year ago, but flagged a catch: "each agent requires GPUs to test their code against." The firm found a constant lack of GPUs inside AMD for this development, which may become a new bottleneck.

The moat is narrowing but not gone. Anthropic's first AMD MI450 chips, in Helios racks, are expected in the first half of 2027 — a new generation and a test of whether the three-day bring-up can be repeated.

Source: Tom's Hardware

Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,035 articles

Related articles

« Previous articleNext article »