Anthropic's Mythos Ported Prefill to AMD GPUs in Three Days
Anthropic said Claude Mythos Preview ported one model’s prefill stage to AMD MI355X GPUs in about three days, but released no decode or comparative performance data.
By Olivia Hart
4 min read

What's News
- Anthropic said one model’s prompt-processing stage moved to AMD MI355X GPUs in about three days.
- The claimed result covered prefill and almost all kernels, but excluded decode-stage results.
- Anthropic’s IPO priced at $83 before November 19, putting its fully diluted valuation near $1.5 trillion.
- AMD’s Helios system costs around $20 billion per gigawatt, roughly half the reported cost of Nvidia’s Vera Rubin.
- Nvidia’s contracted frontier-lab compute share is cited at around 39%, down from previous highs above 80%.
Anthropic told investors during its November roadshow that Claude Mythos Preview adapted the prompt-processing stage of one model to AMD MI355X GPUs in about three days, with “almost no human intervention.” Anthropic said comparable accelerator work had required six to 12 months and dozens of engineers. The technical claim formed part of its broader effort to diversify AI computing.
What did the three-day test actually show?
The disclosure came as Anthropic marketed its IPO at $83 per share before a November 19 listing, putting its fully diluted valuation near $1.5 trillion. The 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” and reached a “good enough” performance level on MI355X for this stage. The company explicitly called the three days a software-onboarding timeline, not a measure of GPU performance. That narrower description matters because software portability does not establish faster silicon.
The result does not cover decode, where the model generates its response. Anthropic supplied no end-to-end throughput or latency figures comparing AMD with Nvidia. The evidence shows rapid adaptation of one stage, not a complete hardware win.
How strong is the historical comparison?
The six-to-12-month estimate came from Anthropic’s own history of bringing software to new accelerators. It was not a side-by-side test between the Mythos-assisted effort and a conventional engineering team. Tom Brown, Anthropic’s co-founder and chief compute officer, said the traditional approach required a team of “dozens of engineers.”
An investor presentation showed two lines both hitting “~100%” without specifying what the percentages measured. Anthropic has not published comparative model speeds on AMD and Nvidia. The caveat is material: the headline speedup lacks a controlled historical benchmark.
Why does Anthropic want multiple chip suppliers?
Anthropic already runs models on Nvidia GPUs, AMD GPUs, Google TPUs and AWS Trainium, managing them as a “fungible pool of compute.” Brown called the early multi-chip decision “an unusual bet.” The strategy seeks to shorten platform onboarding and let Anthropic buy on price-performance rather than accept vendor lock-in.
The company’s AMD agreement is projected to reach multi-gigawatt scale. Anthropic expects MI450 chips in Helios racks in the first half of 2027 for the first phase. It has yet to repeat the three-day port on MI450. The method may shorten adoption, but its largest test remains ahead.
What can AMD gain from Claude?
AMD’s Helios system is priced around $20 billion per gigawatt, roughly half the reported cost of Nvidia’s Vera Rubin platform. AMD has argued that AI can write GPU kernels; its engineers use models to generate kernels that in some cases are “better than hand-tuned versions.”
Under a July agreement, AMD also uses Claude to optimize workloads for Instinct GPUs and accelerate ROCm development. AMD vice president of AI software Anush Elangovan said: “Claude is better than HIPify because it has built-in web search.”
That endorsement supports AMD’s software strategy, but it does not establish full-workload superiority or production economics. Cheaper hardware creates an opportunity, not proof of a lower total cost.
How much does Nvidia’s position weaken?
Nvidia CEO Jensen Huang, who resigned on November 11 amid a separate controversy, said in April that Nvidia also uses AI for its own kernels. He argued “no one understands our architecture better than we do” and called Anthropic “a unique case, not a trend.”
Cited estimates put Nvidia’s share of contracted future AI compute capacity for frontier labs at around 39%, down from previous highs above 80%. The reported commitments totaled $26 billion across seven companies, including Nvidia and AMD, all at the $83 IPO price.
The commitments give both chip vendors exposure to Anthropic’s valuation, but they do not resolve the missing decode and efficiency data. Nvidia’s software moat looks thinner, even if optimized hardware still protects a performance edge.
What did SemiAnalysis conclude?
SemiAnalysis concluded in late July that software no longer has the moat it held a year earlier. The firm identified a different constraint: agentic coding still requires GPUs to test the code it writes, and AMD lacks internal GPUs for that work.
SemiAnalysis said the milestones to watch were performance against Nvidia, decode-stage results and whether the three-day process repeats with MI450. Its assessment exposes a circularity: AI can port software to a chip, but models doing the porting still need accelerated hardware to test their output. The commercial case will depend on measured workload economics, not onboarding time alone.
What remains unproven?
Anthropic has not released decode-stage results, comparative performance data or a replication of the effort on the upcoming MI450. The three-day claim applies to one model’s prefill stage on one AMD GPU generation. Anthropic has not demonstrated sustained, peak-efficiency performance across complete workloads and successive hardware generations.
The next test is specific: Anthropic must repeat the port on MI450 and publish decode, throughput and latency comparisons. Until then, the evidence supports faster software onboarding, not interchangeable AI accelerators.
Source: epyhia.com
More from Olivia Hart
Show full bio
Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,538 articles