Anthropic’s AMD Bring-Up Took 3 Days vs 6-12 Months Before

A slide in Anthropic’s IPO roadshow video puts the time to bring up a new chip at “~3 days with Mythos” against “~6-12 months with traditional approach”, for AMD GPUs. The slide was among four frames posted on X by Tanay Jaipuria on 11 November; the video is on Anthropic’s site for US investors.

Taking a month as 30 days, that is about 60 to 120 times faster, by our arithmetic. The slide’s footnote narrows the claim: the three days cover “prompt processing for one Anthropic model on AMD MI355X GPUs at a useful level of performance”.

Business Pill · GPUs Explained

An evergreen explainer on why AI runs on GPUs. It teaches the general idea only and does not discuss Anthropic or AMD.

What the Video Says About AMD

In the video’s official captions, which do not name speakers, the passage reads: “this year we started working with AMD GPUs for the first time. In the past, bringing on a new chip platform took dozens of engineers months of work. This time we had Mythos try it out. It got Claude running on AMD in just three days with basically no human intervention. Claude ported or wrote almost 100% of the kernels”.

The slide’s footnote says the six-to-12-month figure “is based on the Company’s historical experience bringing up new accelerators using traditional engineering approaches”. It describes the three days as “the approximate time for Claude Mythos Preview” to do that one task, which is narrower than the captions’ “running on AMD”.

The captions set the claim inside a cost argument: “we believe a dollar of compute at Anthropic goes further than a dollar at any other frontier lab”. Our IPO timing piece carries an earlier update box on the same passage.

What Was Posted on X

Tanay Jaipuria posted on 11 November at 01:37 UTC four images the post calls slides from Anthropic’s roadshow video, noting a “big emphasis on fungible compute, ability to use compute efficiently and onboard new chips quickly”. One image compares about three days with Mythos to about six to 12 months by the traditional approach for AMD GPUs; this piece reads that slide’s footnote and checks it against the video’s captions and Anthropic’s S-1/A.

One Pool, Four Chip Families

A second slide shows a “Fungible compute pool” serving four uses, “Frontier research”, “Model development”, “Customer demand” and “Internal Claude use”, drawn from AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs. The captions add: “Our multi-silicon chip strategy means we can treat all of our chips across providers, architectures and locations as one fungible pool of compute”.

The S-1/A filed on 10 November uses the same four names: “We train and serve Claude natively across four major chip families (AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs) using a proprietary orchestration layer that enables us to optimize workloads for price and performance”.

It says that foundation, “paired with Claude’s increasing understanding of chip and model architectures, allows us to onboard new chip platforms in a fraction of the time this once required”. The filing does not give the three-day figure in the text we searched; the slide does.

The AMD Contract Behind the Test

The filing describes a Hardware Supply Agreement signed with AMD in July 2026 for “AMD-based Helios rack-scale systems featuring AMD Instinct™ MI450 series GPUs”. Under it, Anthropic “committed to obtain an initial tranche of compute capacity in calendar year 2027”.

The agreement “does not specify a fixed dollar commitment”. The filing says: “If we procure the full amount of this initial committed capacity, we expect aggregate expenditures to exceed $20.0 billion over the term of the agreement”. It also allows for additional capacity of up to about the amount of the initial commitment.

AMD has committed to buy up to $5.0 billion of Anthropic’s Class N stock, unlocked by GPU delivery and deployment milestones, the S-1/A says. AMD’s $2.5 billion purchase in the concurrent private placement counts against it, leaving up to $2.5 billion, with the first milestones expected in 2027.

The chips differ. The three-day example names MI355X GPUs; the supply agreement covers the MI450 series. Our count of Anthropic’s named compute deals puts Google TPUs at 44% of the dollar total.

The Structural Read

As we read the video and the filing together, the three-day figure is offered as evidence for a cost claim, that a dollar of compute goes further at Anthropic, while the filing’s own wording stays general: new chip platforms in “a fraction of the time this once required”.

As we read the slide, the comparison is between one narrowly defined task and Anthropic’s historical estimate for full bring-ups, so the 60 to 120 times is an upper reading of the speed-up, not a measured like-for-like test.

Three Implications

FOR AMD The S-1/A expects aggregate expenditures to exceed $20.0 billion if Anthropic procures the full initial MI450 tranche, with AMD’s remaining equity commitment of up to $2.5 billion tied to delivery milestones from 2027.

FOR THE CHIP MIX The filing names four chip families in one pool: AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs; the slide shows the same four.

FOR THE CLAIM The three days cover prompt processing for one model on MI355X GPUs, the footnote says; the supply agreement is for the MI450 series.

The Business Engineer Lens

This story maps onto the Business Engineer framework The AI Supercycle.

The framework maps the chip layer as: “Three competitive categories within one layer: generalist substrate (NVIDIA), hyperscaler custom (TPU, Trainium, MTIA), merchant custom challengers (Cerebras, Groq, SambaNova).”

As we read it, Anthropic’s pool now spans the first two categories plus AMD, and the roadshow presents the speed of adding a chip family as part of its cost case. The framework does not discuss the video.

What Is Not Established

The three days measure one task, prompt processing for one model at a useful level of performance, as the footnote defines it; neither the video nor the S-1/A text we searched says how long full training or serving on AMD took. The six-to-12-month figure is Anthropic’s estimate from its own history, not a measured comparison. We read the slides as frames posted on X by Tanay Jaipuria, not from the video file itself. AMD and Anthropic were not contacted for this piece.

The Bottom Line

Anthropic’s roadshow video says Claude Mythos Preview brought up prompt processing for one model on AMD MI355X GPUs in about three days, against six to 12 months by its traditional approach, about 60 to 120 times faster by our arithmetic. Its S-1/A ties AMD to a 2027 MI450 tranche that it expects to exceed $20.0 billion if fully procured.