Investor Research

AMD, NVDA · US

Can AMD break the CUDA moat? Advancing AI 2026 | SemiAnalysis note

SemiAnalysis: ROCm/CI/MI455 vs CUDA gates. https://newsletter.semianalysis.com/p/can-amd-break-the-cuda-moat-amd-advancing

Published Updated Open interactive reader

Cite a section with a deep link, e.g. /en/r/sa-amd-cuda-moat-2026#thesis

Snapshot

Source date
2026-07-25
Focus
ROCm / CI / MI455
Public CI timeline
~2026-10 (slip)
Peer anchor
NVDA CUDA
As of 2026-08-03

Structured research note summarizing a SemiAnalysis subscriber article (not a full reprint). Defer to the original for detail.

Thesis

Beyond silicon slides, SemiAnalysis focuses on ROCm/vLLM/SGLang CI and kernel quality: new MI455 ISA, Helios ramp pain, and unstable internal clusters mean wartime software still trails CUDA merge gates. Treat AMD as a #2-supplier option—software parity timing remains the core risk.

Analysis

Software is the moat battlefield

Unlike Rubin’s relative ISA continuity vs Blackwell, MI455 (gfx1250) is a new ISA vs MI355 (gfx950)—dual codepaths stress CI. A goal to land MI455 in public vLLM/SGLang nightlies by Advancing AI 2026 slipped toward ~Oct 2026 (with left-shift attempts).

Progress vs gaps

  • Progress: ROCm into upstream vLLM early 2026; more CI mirrors/gates
  • Gaps: public regression dashboards, accuracy gates, e2e disaggregation CI, automated perf gating still roadmap-ish; CUDA merge-gate parity unproven
  • Org: unstable internal clusters / under-planned CI capacity called out as drags

Read-through to AMD / NVDA

Split the share story into pricing/supply windows (winnable) vs large-scale MoE/rack software experience (still hard). See also CPU industry.

Implications

Name Read-through
AMD CI/nightly milestones > FLOPS slides
NVIDIA CUDA + rack software remains default

Outlook

Watch AMD nightlies on schedule, Helios/MI455 deployability, and whether large buyers still demand sweetheart pricing.

Risks

References

  1. Original (SemiAnalysis): Can AMD break the CUDA Moat? AMD Advancing AI 2026
  2. SemiAnalysis archive — https://newsletter.semianalysis.com/archive
  3. AMD investor relations — https://ir.amd.com/
  4. NVIDIA developer / CUDA — https://developer.nvidia.com/cuda-zone
  5. vLLM project — https://github.com/vllm-project/vllm

Not investment advice. Copyright remains with SemiAnalysis / authors; this package is an on-site research index.