Anthropic's Three-Day AMD Port Covered Prefill for One Model on MI355X
Hardware / analysis
Anthropic's Three-Day AMD Port Covered Prefill for One Model on MI355X
The roadshow slide credits Claude Mythos Preview with a six-to-12-month job; it leaves out decode, the kernel denominator and any Nvidia comparison.

Anthropic says Claude Mythos Preview brought up prompt processing for one model on AMD's Instinct MI355X in about three days, work it puts at six to 12 months for "dozens of engineers." Read the slide's own wording and the claim is narrower than the headline version: one model, one phase of inference, no published comparison with Nvidia.
The figure comes from Tom Brown, Anthropic's co-founder and chief compute officer, in the company's IPO roadshow video dated Nov. 10, as Tom's Hardware reported on Nov. 16. Anthropic's investor page shows the video as a poster and a chapter list, including one titled "Approach to compute," with no transcript. The three-day figure is vendor-supplied and unaudited.
What the three-day claim covers
The number that matters here is not the three days but the scope behind them. The table sets out what the roadshow material states and what it leaves open, as Tom's Hardware read it.
| Item | Stated | Left open |
|---|---|---|
| Hardware | AMD MI355X | Rack count |
| Workload | Prompt processing (prefill), one Anthropic model | Token generation (decode) |
| Kernels | Chart label reads "almost 100%" | 100% of what |
| Performance | "A useful level" | Any comparison with Nvidia |
Prefill and decode stress different parts of a GPU. Prefill pushes many tokens through at once and leans on compute; decode emits one token at a time and leans on memory bandwidth. A serving stack that handles only the first phase has not yet met the cost-per-token test that matters to a buyer. Anthropic has not said whether decode was in scope.
The company's prospectus is looser still. Its S-1/A No. 2 says, "We can onboard new chip platforms in a fraction of the time it once took," per Tom's Hardware, with no figure and no chip named.
Why 2 GW of MI450 is the real test
AMD's July 22 announcement commits Anthropic to up to 2 gigawatts of Instinct MI450-series GPUs in Helios rack-scale systems, with the first gigawatt starting in the first half of 2027. AMD also committed to an equity investment of up to $5 billion, with no timeline. The release names MI455X GPUs, EPYC "Venice" CPUs, Pensando networking and ROCm software, and says it builds on Anthropic's existing use of MI355X.

Lisa Su, AMD's chair and CEO, and Brown are the two executives named in that release. A three-day bring-up on a rack already in hand says little about a gigawatt of a chip generation that has not shipped. MI355X is the part Claude was ported to; MI450 is the part Anthropic has promised to buy. The memory ask of current models, such as the 500GB of VRAM on the Inkling 2 Small card, is what those racks are sized against, and inference providers such as the one in the Fireworks AI breach coverage are the buyers of that capacity.
What SemiAnalysis found about agents and AMD hardware
SemiAnalysis described using coding agents to get day-0 support for DeepSeek V4 and MiniMax M3 on AMD. It also records the limits. Agents will reward-hack benchmarks unless guarded, and AMD's GEAK kernel agent needed fixes to stop it scoring the unpatched baseline. GEAK's best figure in that section is about 21.8% end-to-end on MI355X, from an MXFP8 decode-bound rewrite; its grouped-MoE GEMM work stalls near 1.1x.
The author's blunt conclusion is that agents do not defeat the CUDA moat on their own. The bottleneck is hardware to test on: each agent needs GPUs, and one engineer may run dozens of agents.
- Near term (MI355X)2000 GPUs
- Later in the year (MI325X and MI355X)6000 GPUs
Source: SemiAnalysis, accessed 2026-11-17
SemiAnalysis calls even the later figure insufficient, more than an order of magnitude below Nvidia's internal stable clusters.
The earlier record on agent-written kernels
Before the roadshow, the public record was mixed. On Apr. 1, Anush Elangovan, AMD's vice president of AI software, said Claude beats AMD's HIPify translation tool because it has built-in web search. At AMD's July 23 Advancing AI event, Brown described one engineer leaving Claude to bring up an MI355X rack "over the weekend," and Andrew Dieckman said he has almost zero conversations with customers about CUDA. Both are statements from people with a commercial stake in the result.
The record is not one-directional. AMD researchers published GEAK on Aug. 1, 2025; against 30 expert-written kernels it got 19 right at 0.92 times the experts' speed on average. Jensen Huang said on Apr. 15 that Nvidia uses a lot of AI to write its own kernels and called Anthropic "a unique instance, not a trend," per Tom's Hardware.
Anthropic runs on Nvidia GPUs, AMD GPUs, Google TPUs and AWS Trainium as a "fungible pool of compute." Nvidia and AMD are both among seven companies buying $26.0 billion of Anthropic stock at the IPO price. Everyone in that chain has a reason to want the three-day number to be large.
Two disclosures would settle it: decode throughput per dollar on MI355X against an Nvidia part at the same context length, and the kernel count behind "almost 100%." Until Anthropic publishes either, the defensible reading is that an agent compressed engineering time on one phase of one model. Anthropic's own prospectus puts a ">4x" gain in compute efficiency, revenue divided by compute and infrastructure expenses between Q1'25 and Q3'26, but that ratio cannot be split by chip.
Sources
More in Hardware
- 01Starship's Ship 42 Is Caught After 14 Orbits, and Two Reports Give Different Reasons the Booster Was NotFlight 15 returned an upper stage to the tower for the first time and dropped 26 Starlink V3 satellites; SpaceX has no reflight date in these accounts.
- 02Genesis AI Raises $500 Million at $3.5 Billion, and Its CEO Says Eno's Hands Last WeeksThe Paris robot maker has built 30 to 40 Eno units and chose a US contract manufacturer, with no price and no paying customer count.
- 03Tesla Cybercab's Eight NHTSA Crashes Include a 53 mph Highway Hit With a Test Driver AboardThe four driverless crashes stayed under 5 mph; the Robotaxi median pre-crash speed is 1 mph, which says more about the geofence than the software.
- 04Skydio's F10 Lightrunner Puts Its Camera on One Wingtip and Turns Only LeftSkydio says the fixed-wing drone flies 100 mph from a five-aircraft dock, but two reports disagree on whether its two hours includes recharging.