Anthropic Says Mythos Made AMD MI355X Prompt Processing Usable in 3 Days

Anthropic said in an investor presentation that its Mythos model brought prompt processing of one Anthropic model on AMD MI355X GPUs to a usable level in about 3 days, against 6 to 12 months by conventional engineering, The Decoder reported on Nov. 11.

The figure comes from a chart in Anthropic's investor presentation, and The Decoder says it has not been independently verified. According to The Decoder, Anthropic says Mythos wrote or converted close to 100 percent of the kernels on its own, with almost no human help. The Decoder's caption on the chart says the figures refer to prompt processing of an Anthropic model on AMD MI355X GPUs, and the report gives no figure for token generation or production serving.

What Anthropic's S-1/A Says About New Chips

Anthropic's amended S-1 registration statement, filed with the SEC on Nov. 10, is more guarded. It says bringing a new chip platform into production once took intense engineering effort over many months, and that Claude now lets its engineers do it more quickly, in a fraction of the time. The filing names no day count, no MI355X and no Mythos in that passage. It credits part of the speed to years of multi-chip work, including a proprietary orchestration layer that measures performance on Anthropic's own workloads.

The filing describes one compute pool across AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs. It also reports that revenue grew 1,127% year over year in the quarter to Sept. 30, while compute and infrastructure expenses grew 382%.

Why Kernels Are the CUDA Moat

Kernels are the small programs that run a model's maths on a specific chip. NVIDIA's CUDA has years of tuned kernels behind it, and rewriting them for AMD's ROCm stack is a major cost of switching. If a model can write most of them in days, then the price of moving a workload between vendors falls, and buyers can chase whichever accelerator is cheaper or available. That is an argument, not a measurement: a usable level for one phase of one model is a long way from matching NVIDIA in production.

The claim lands on hardware Anthropic is about to buy at scale. AMD said on July 22 that Anthropic will deploy up to 2 GW of Instinct MI450 Series GPUs, the MI455X, in Helios racks, with deployment of the first gigawatt starting in the first half of 2027. AMD said that deal builds on Anthropic's existing MI355X use, and that the two would use Claude to optimise Instinct workloads and speed up ROCm development. AMD also committed to invest up to $5 billion in Anthropic equity in the future. MI355X capacity is already spreading, as with Vultr's 20 MW cluster in Chennai.

Both GPU makers stay tied in. The S-1/A lists AMD and NVIDIA among seven companies that agreed to buy a combined $26.0 billion of stock in a private placement at the IPO price, alongside Broadcom, Google, Micron, Microsoft and TSMC Partners. The placement depends on the offering closing.

What Prompt Processing Covers, and What It Leaves Out

Serving a model has two phases. Prompt processing, also called prefill, reads the whole prompt at once and is compute-heavy. Token generation, or decode, produces one token at a time and re-reads the model's weights and cache for each, so it leans on memory bandwidth.

Anthropic's claim covers only the first phase, on one model. That makes it a statement about how fast a compute-bound workload can be brought up on new silicon, not about how a full Claude deployment performs on AMD hardware.

What Is Not Confirmed Yet

The 3-day and 6-to-12-month figures come only from Anthropic's presentation as reported by The Decoder; the S-1/A gives no number. Voltgrid found no AMD comment and no independent measurement of MI355X performance on Anthropic's model as of Nov. 12, and Anthropic has not defined what a usable level means. NVIDIA reports third-quarter fiscal 2027 results on Nov. 17, the next scheduled chance for its management to address whether AI-written kernels erode CUDA's hold.