Anthropic's AI Ported Itself to AMD Chips in 3 Days. The Old Way Took 6 to 12 Months.

Three days. That is how long Anthropic says its own AI model needed to move Claude onto a rival chipmaker's hardware, a job its own chart prices at six to twelve months of human engineering. Co-founder and Chief Compute Officer Tom Brown made the claim in Anthropic's IPO roadshow video, posted around November 10: the company's frontier model Mythos "got Claude running on AMD in just three days with basically no human intervention," and "Claude ported or wrote almost 100% of the kernels." Kernels are the small, low-level programs that tell a chip how to run a model's math, and writing them by hand is the most labor-intensive part of adopting any new accelerator. If the claim holds, one AI just erased the most expensive step in switching chip vendors.

Anthropic is not a disinterested narrator here. It filed its S-1 on October 26, 2026, ahead of a planned Nasdaq listing, and the roadshow video exists to sell shares. But the claim also arrives with a paper trail that makes it checkable in parts, and the parts that check out are the ones that matter for the chip industry. This story separates the two.

What actually happened

Mythos is Anthropic's frontier model, and the port targeted AMD's Instinct MI355X, a data-center accelerator built on AMD's CDNA4 architecture with 288 gigabytes of HBM3E memory, 8 terabytes per second of memory bandwidth, and a 1,400-watt board power rating. Anthropic only started working with AMD GPUs this year, according to Brown, and the roadshow slide lists four chip families Claude now trains and serves on: Amazon's Trainium, Google's TPUs, Nvidia's GPUs, and AMD's GPUs, described as one interchangeable pool rather than four isolated sources of compute. An earlier demo had Claude Code porting a CUDA backend to AMD's equivalent in about thirty minutes.

Context turns a demo into a strategy. On July 22, 2026, AMD and Anthropic announced a partnership under which Anthropic will deploy up to 2 gigawatts of AMD Instinct MI450 Series GPUs in Helios rack-scale systems, with the first gigawatt starting in the first half of 2027, while AMD committed to invest up to $5 billion in Anthropic as deployment milestones are hit. AMD's HIPIFY translator already rewrites the CUDA code that maps cleanly. Mythos did the rest. Anthropic has gigawatts of AMD hardware arriving and a financial reason to make AMD's ROCm software stack work, which is why a three-day port is a business event, not a lab curiosity.

The math nobody ran

Start with the speedup, because nobody in the coverage computed it. Six months is 180 days, and 180 divided by 3 is 60. Twelve months is 365 days, and 365 divided by 3 is about 122. So the claim, taken at face value, is a 60x to 120x compression of calendar time. Headlines say "months to days." Numbers say sixty to a hundred and twenty times faster.

Calendar time understates it, because the old way also burned people. Brown's baseline was "dozens of engineers months of work," and Anthropic's own footnote says the 6-to-12-month estimate reflects the company's historical experience bringing up new accelerators with traditional engineering. Take the midpoint: 24 engineers working 9 months is 216 engineer-months, or 18 engineer-years. Take the floor: a dozen engineers for 6 months is 72 engineer-months, or 6 engineer-years. A senior systems and kernel engineer in the Bay Area costs roughly $500,000 a year once salary, benefits, and overhead are loaded in, a standard industry rule of thumb we state as an assumption, not a fact. Multiply it out and the old port cost about $3 million at the floor and about $9 million at the midpoint, in human labor alone. Mythos compressed that into three days of model inference, a cost Anthropic has not disclosed but which sits, by any plausible accounting, three orders of magnitude lower. Porting a model to a new chip just went from a budget-line decision to a long-weekend experiment.

Anthropic's own chart supplies a third number, and it cuts against the hype. Its slide shows prefill running at 3,491 tokens per second and decode at 104.8 tokens per second on the MI355X. Prefill is prompt processing, the stage where a model reads its input. Decode is token generation, the stage where it writes its answers, and the stage users actually wait on. Divide the two and prefill runs 33.3 times faster than decode on the ported stack. "Useful level of performance" turns out to rest on the fast half of inference, while the slow half is the one the footnote does not cover.

The math, computed from Anthropic's roadshow figures, AMD's MI355X specifications, and the assumptions stated in the text.
Figure	Value	How it was computed
Calendar speedup	60x to 122x	180 days / 3 days; 365 days / 3 days
Engineer time, midpoint	18 engineer-years	24 engineers x 9 months
Engineer time, floor	6 engineer-years	12 engineers x 6 months
Labor cost, midpoint	~$9 million	18 engineer-years x ~$500,000 loaded cost
Labor cost, floor	~$3 million	6 engineer-years x ~$500,000 loaded cost
Prefill vs decode	33.3x	3,491 tok/s / 104.8 tok/s (Anthropic chart)
MI355X memory	288 GB HBM3E	AMD product brief
MI355X bandwidth	8 TB/s	AMD product brief
MI355X board power	1,400 W	AMD product brief
AMD deal size	Up to 2 GW	AMD press release, July 22, 2026
AMD equity in Anthropic	Up to $5 billion	AMD press release, July 22, 2026
CUDA developers	7.5 million+	Nvidia fiscal 2026 Form 10-K

The footnote is the story

Every careful write-up of this claim lands on the same sentence, because it is printed on Anthropic's own slide. "The ~3 day timeline reflects the approximate time for Claude Mythos Preview to bring up prompt processing for one Anthropic model on AMD MI355X GPUs at a useful level of performance." Read it slowly. One model, not the fleet. Prompt processing, not training and not full inference. Approximate time. Useful level of performance, a phrase Anthropic has never defined. Nobody outside the company knows what the run cost in compute, what "useful" meant in tokens per second beyond the two chart figures, or whether the ported model ever served a real user.

That narrowness does not make the claim false. It makes it specific, and specificity is what separates a demonstration from a press release. Bringing up prefill for one model in three days is genuinely impressive work, the kind of thing that used to take a team a quarter or more, and it is the right first step: prefill is where a new accelerator either works or does not. Training on AMD hardware, serving full inference at production quality, and porting the long tail of framework integration are separate jobs, and none of them is covered by the three days. Treat the claim as a beachhead, not a conquest.

What it means for Nvidia's moat

Nvidia's defense was never the silicon. It was the software: CUDA, launched in 2006, now counts more than 7.5 million developers according to Nvidia's fiscal 2026 Form 10-K, with roughly 250 widely used libraries built over two decades. Switching costs that high are why Nvidia has held 75 percent gross margins while rivals shipped faster chips. Yet the moat has two layers, and they are not equally thick. One layer is the libraries and the developers, which no model ports in a weekend. Another layer is the cost of porting a workload off CUDA, and that layer just got 60 to 120 times cheaper for the hardest part of the job, which is writing kernels.

Think of it as a negotiating lever, not a kill shot. When porting costs dozens of engineers and most of a year, a second-source chip is a science project, and the incumbent can price accordingly. When porting costs three days of model time, a second source becomes a credible threat, and credible threats change prices even before anyone switches. Anthropic's four-chip pool is the proof of concept: Trainium, TPUs, Nvidia GPUs, and AMD GPUs treated as one interchangeable supply, with the buyer, not the vendor, choosing where each workload lands. If that pattern spreads to other large buyers, Nvidia keeps the ecosystem and loses the pricing power that the ecosystem used to guarantee.

Limitations

Several numbers here carry asterisks that belong in the open. No independent verification exists for the three-day timeline, the autonomy level, or the performance figures; every number in the claim is Anthropic-attributed, from a roadshow video whose job is to sell shares, and the prospectus itself does not repeat the three-day figure. AMD has not commented. Anthropic has not disclosed the compute cost of the three-day run, the size of the model ported, or decode performance at production quality, so the cost comparison in this article covers human labor only. Engineer counts and the $500,000 loaded cost are stated estimates (12 to 24 engineers, 6 to 9 months), which makes the $3 to $9 million range an order-of-magnitude estimate, not an audit; the arithmetic is shown so readers can redo it with their own inputs. Prefill and decode figures come from Anthropic's chart as reproduced in secondary reporting, not from a published benchmark. Mythos is not a public model, so no reader can rent this capability today.

Strongest case against

Steelman: the skeptic's case starts with the source and ends with the long tail. Anthropic's claim comes from IPO marketing, not a benchmark, and it has not been independently verified, AMD has not commented, and the prospectus does not repeat it. Its own footnote confines the result to prompt processing for one model, not training, not full inference, not production quality, and decode ran at 104.8 tokens per second on Anthropic's own chart while nobody outside the company knows what "useful level of performance" meant or what the run cost. CUDA's moat was never just the kernels: 7.5 million developers and roughly 250 libraries built over twenty years do not get ported by a model in a weekend, and every lab that has tried to leave CUDA has discovered the long tail of framework integration, profiling tools, and edge-case kernels. One port of one stage of one model is a demonstration, not a dead moat, and treating it as the end of Nvidia's pricing power is the kind of extrapolation that has embarrassed tech journalism before. Every clause of that case is true, and it is also compatible with the central finding here: the most expensive step in switching chips just got sixty to a hundred and twenty times cheaper, and that changes the negotiation even if nobody switches tomorrow.

What you can do

If you run AI infrastructure, price a second-source pilot this quarter instead of next year, because the experiment that used to cost a team and most of a year now costs a long weekend, and the only way to learn whether your workloads port cleanly is to try one. If you invest in chips, stop watching benchmark charts and start watching porting stories, since the moat that matters is no longer measured in teraflops but in how many days it takes a model to move, and Anthropic just published the first data point. If you build with CUDA, audit which of your dependencies are truly CUDA-only and which are habit, because HIPIFY already handles the parts that map cleanly and models are now doing the parts that do not. And if you are simply watching the AI buildout, remember the number 33.3, because prefill running thirty-three times faster than decode is the shape of every porting claim you will hear next, and the question to ask is always the same: which half did they port?

The Bottom Line

Anthropic says its Mythos model moved Claude onto AMD chips in about three days, work its own chart prices at six to twelve months of human engineering, and our math puts the old labor bill at roughly $3 to $9 million against three days of model time. The footnote narrows the feat to prompt processing for one model, the decode half ran 33 times slower on Anthropic's own chart, and nothing about the claim has been independently verified. Both things are true at once, and that is the point: a demonstration this narrow should not kill a moat built by 7.5 million developers over twenty years, but a porting cost that falls sixty to a hundred and twenty times is the kind of change that rewrites who gets to negotiate, and the next port will tell us whether three days was a stunt or a new normal.