Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days

Anthropic Says Claude Mythos Ported Its Stack to AMD's MI355X in About Three Days, Work It Says Once Took Up to 12 Months

Anthropic Says Claude Mythos Ported Its Stack to AMD's MI355X in About Three Days, Work It Says Once Took Up to 12 Months

Anthropic's IPO roadshow materials say Claude Mythos Preview brought one model's prompt-processing stage to usable performance on AMD's MI355X in about three days, writing or porting nearly 100 percent of the kernels, work the company says once took dozens of engineers up to 12 months. The figures are self-reported and cover only that stage, not a full training migration.

Anthropic has disclosed to IPO investors that its Claude Mythos Preview model completed the software adaptation needed to run one of its models on AMD's MI355X GPU in about three days, with almost no human intervention. The disclosure, reported by Tom's Hardware and attributed to co-founder and Chief Compute Officer Tom Brown, describes work that the company says historically required dozens of engineers and six to twelve months.

The scope of the claim matters as much as the headline. Anthropic says the three days covered bringing the prompt-processing stage of a single model to usable performance on the MI355X, and that Mythos ported or wrote nearly 100 percent of the underlying kernel code for that stage. Prompt processing is the front half of inference, where a model reads and digests its input before generating output. It is a demanding, performance-critical workload, but the disclosure does not mean Anthropic moved a full training and inference stack to AMD silicon in a weekend, and the company has not offered independent verification of the timeline.

The reason this lands with force is that the code in question has long been treated as Nvidia's deepest moat. Low-level kernels bind model math to specific silicon, and writing them for a new accelerator has been slow, specialist work measured in engineer-years. If a frontier model can genuinely compress that cycle from months to days, the software switching cost that keeps AI labs on CUDA weakens, and chip choice becomes a negotiation about price, supply and performance rather than a lock-in decision.

Anthropic's own infrastructure strategy explains why it invested in this capability early. The company runs its models across Nvidia GPUs, AMD accelerators, Google TPUs and AWS Trainium, and manages the whole fleet as a fungible compute pool. Brown has described the early multi-chip decision as an unconventional bet; the roadshow disclosure is effectively evidence that the bet paid off, at least for one stage of one workload on one accelerator.

The competitive backdrop sharpens the story. Bernstein analysts noted in the same coverage that AMD's share of the x86 server CPU market rose about one point quarter-over-quarter to 45.6 percent in the third quarter, with server CPU average selling prices up 12 percent sequentially and 34 percent year over year, and argued that falling software migration barriers could help AMD's server business roughly double revenue by 2027. Separately, Google has confirmed TPUs now natively run PyTorch-built models in Meta's production environment, which is the same erosion of the porting barrier from another direction.

For the industry, the claim is a signal rather than a settled fact. Anthropic has every incentive to impress IPO investors, and the numbers come from the company's own materials without third-party benchmarking. But the direction of travel is consistent with what AMD, Google and the open-weight ecosystem have all been arguing: if AI systems can increasingly write their own hardware adaptation layer, the last structural advantage of the incumbent chip vendor becomes the next thing a model automates.

Comments (0)

Log in to join the discussion

Log In

No comments yet