Separating AI hype from reality.

Anthropic Says Claude Mythos Ported Prompt Processing to AMD GPUs in Three Days

November 22, 2026 · Anthropic · AMD · Nvidia

Anthropic told investors during its November roadshow that its Claude Mythos Preview model adapted the prompt-processing stage of one of its models to run on AMD MI355X GPUs in about three days, with "almost no human intervention." The company said the same work previously took six to 12 months and dozens of engineers [1][2]. The disclosure came as Anthropic marketed its initial public offering, which priced shares at $83 ahead of its November 19 listing, putting its fully diluted valuation near $1.5 trillion [4].

What the three-day figure covers

The reported task was prompt processing, also called prefill: handling input tokens before the model generates a response. Anthropic said Mythos wrote or ported "almost 100% of kernels" on its own and reached a "good enough" performance level on the MI355X for this stage [2]. The company explicitly characterized the three-day figure as a software-onboarding timeline, not a measure of GPU performance [1]. It does not cover the decode stage—generating the response—nor does it provide end-to-end throughput or latency numbers comparing AMD and Nvidia hardware [1][2].

Historical baseline, not a controlled test

Anthropic's six-to-12-month estimate comes from its own past experience bringing software to new accelerators, not from a side-by-side trial against the Mythos-assisted effort [1]. Tom Brown, Anthropic's co-founder and chief compute officer, said the traditional approach required a team of "dozens of engineers" [2]. The investor presentation showed two lines both hitting "~100%" without specifying 100% of what, and the company has not published how fast models run on AMD versus Nvidia GPUs [2].

The fungible-compute strategy

Anthropic already runs models on Nvidia GPUs, AMD GPUs, Google TPUs, and AWS Trainium, managing them as "a fungible pool of compute" [2]. Brown called the early decision to go multi-chip "an unusual bet." The goal is to bring any new chip platform into service in a fraction of the previous time, letting Anthropic buy based on price-performance rather than vendor lock-in [3]. The company's deal with AMD is projected to reach multi-gigawatt scale, with the first phase—MI450 chips in Helios racks—expected in the first half of 2027 [2][6].

AMD's cost advantage and software friction

AMD's Helios system is priced around $20 billion per gigawatt, roughly half the reported cost of Nvidia's Vera Rubin platform [3][6]. AMD has long argued that AI can write GPU kernels; its engineers use models to generate kernels that in some cases are "better than hand-tuned versions" [2]. Under a July agreement, AMD also uses Claude to optimize workloads for Instinct GPUs and accelerate ROCm development [2]. Anush Elangovan, AMD's vice president of AI software, said "Claude is better than HIPify because it has built-in web search" [2].

Nvidia's position

Nvidia CEO Jensen Huang, who resigned on November 11 amid a separate controversy [4], said in April that Nvidia also uses AI for its own kernels and argued "no one understands our architecture better than we do." He called Anthropic "a unique case, not a trend" [2]. Nvidia's share of contracted future AI compute capacity for frontier labs has declined to around 39%, down from previous highs above 80% [6]. Nvidia and AMD are among seven companies that committed to buy $26 billion of Anthropic stock at the IPO price [2].

Independent assessment

SemiAnalysis concluded in late July that software is no longer the moat it was a year ago, but noted a new bottleneck: agentic coding still needs GPUs to test the code it writes, and AMD lacks internal GPUs for that work [2]. The firm said the key milestones to watch are performance numbers against Nvidia, results for the decode stage, and whether the three-day process repeats with the MI450 [2].

What remains unknown

Anthropic has not released decode-stage results, comparative performance data, or a replication of the effort on the upcoming MI450. The three-day figure applies to one model's prefill stage on one AMD GPU generation. Whether AI-assisted porting can deliver sustained, peak-efficiency performance across full workloads—and across successive hardware generations—has not been demonstrated.

Sources

  1. Claude Mythos on AMD MI355X: what three days means (neoteo.com)
  2. Claude Mythos Preview ports AI to AMD GPUs in 3 days - migovi (migovi.com)
  3. Anthropic Just Delivered Great News for AMD Investors and Bad ... (finance.yahoo.com)
  4. Jensen Huang's Nvidia legacy, Claude Fable 5.2 and Anthropic's ... (deeplearning.ai)
  5. Not Nvidia. Not Broadcom. If I Could Buy and Hold Only 1 AI Chip Stock Through 2030, It Would Be This One (finance.yahoo.com)
  6. Nvidia's moat looks shallower as Anthropic shows AI chips are becoming interchangeable (cryptobriefing.com)
  7. Introducing Claude Fable 5.2 and Claude Mythos 5.2 \ Anthropic (anthropic.com)