Anthropic reports three-day Claude model bring-up on AMD MI355X
Anthropic said Claude Mythos Preview brought one of its AI models onto AMD MI355X GPUs in about three days, compared with an estimated six to 12 months using traditional engineering methods.
Key takeaways · 4
- 01
Anthropic said one model reached a useful level of performance on MI355X GPUs after about three days of work.
- 02
The estimate covers prompt processing, not the full inference process, and Anthropic had not said whether it included decode.
- 03
Anthropic said Claude ported or wrote almost 100% of the kernels, with basically no human intervention.
- 04
Anthropic did not disclose how quickly the model ran on AMD silicon compared with Nvidia’s.
What the three days covered
Anthropic made the claim in a Nov. 10 roadshow video, saying Claude Mythos Preview brought one of its AI models onto AMD MI355X GPUs in about three days.[1] The scope matters: the reported timeline covered prompt processing for one model, rather than the full inference process.[1] It had not said whether the three-day effort included decode, the generation step.[1] Teams assessing the claim should therefore distinguish the reported milestone from a complete account of serving performance.
A claimed change in engineering effort
Anthropic estimated that bringing the model up on the AMD GPUs would previously have taken six to 12 months.[1] The company based that estimate on its experience bringing up new accelerators with traditional engineering methods.[1] Anthropic co-founder and chief compute officer Tom Brown said the original effort required dozens of engineers.[1] Against that comparison, Anthropic said Claude ported or wrote almost 100% of the kernels for the work, with basically no human intervention.[1] Kernels are low-level code that connects a model to a specific chip.[1] The reported comparison is about this development task, not a general measure of engineering time saved across deployments.
Important performance limits
The announcement does not provide a direct comparison of how quickly the model ran on AMD silicon versus Nvidia GPUs; Anthropic had not disclosed that information.[1] Nor does the reported three-day timeline establish that the full inference process, including decode, was covered.[1] These omissions leave open practical questions for teams evaluating hardware: how the model performs during generation, and how its performance compares across chip platforms. The timeline alone is not enough to determine whether a particular workload is a fit.
Multi-chip context and outside caution
Anthropic said it runs models on Nvidia and AMD GPUs, Google TPUs and AWS Trainium, managing them as one pool of compute.[1] The company described its multi-chip strategy as unconventional.[1] A report characterized the result as “a unique instance, not a trend.”[1] SemiAnalysis found a constant lack of GPUs inside AMD for this development.[1] Separately, a report cautioned that Nvidia’s CUDA ecosystem includes libraries, tools, years of optimization and a large developer base.[2] Together, these points are context for interpreting one reported bring-up, rather than proof that the same timeline applies across chips or organizations.
For teams planning accelerator work, the claim is a reason to ask whether AI-assisted kernel development can shorten a specific bring-up task. The incomplete inference scope and missing AMD-versus-Nvidia speed comparison mean the timeline should not stand in for workload testing.
Why it matters
AI training for your team
kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.
How this developed
17 November 2026
Anthropic reports three-day Claude model bring-up on AMD MI355X
