Nvidia的CUDA护城河在推理领域受到挑战
Nvidia的CUDA护城河在推理领域受到挑战
Conclusion: JPMorgan is directionally right, but the CUDA moat is not disappearing uniformly. The evidence points to a split: CUDA remains close to indispensable for training and CUDA-first frontier libraries, while inference, which is the larger and more standardized workload over a model's life, is where the moat is leaking. A weaker moat could help AMD win more data-center accelerator demand, but mainly in inference and mainly through large, contracted deployments. Most of the AMD-side evidence is company-reported, and JPMorgan's share figures compare different bases, so the case is promising but unproven.
Data as of: 17 Nov 2026 (US Eastern) / 18 Nov 2026 (Beijing). I did not use live market quotes; this is a qualitative software and competitive-position analysis.
What supports JPMorgan's view
- Share of new contracted compute is falling. JPMorgan's Michael Cembalest says Nvidia GPUs are of the infrastructure frontier labs have contracted to secure going forward, versus an share of AI chip revenue in 2023–2025. Custom Broadcom chips rank second at , and AMD GPUs hold . Note the mismatch: the 39% is a share of contracted compute, while the 83% is a share of revenue. That is not a like-for-like share-loss number.
- The switching cost is shrinking. Cembalest says onboarding a model on AMD GPUs fell from a year earlier to now, because AI models help with the porting work. That directly undercuts Nvidia's argument that switching is uneconomical because of CUDA.
- Inference is where AMD is competitive. JPMorgan says AMD ranked ahead of Nvidia in an inference comparison that included memory and inter-chip communication. Epoch AI's modeling-based index also ranked AMD's MI455X highest per chip (Seoul Economic Daily, 16 Nov 2026).
- AMD has the hardware and customers to act on it. AMD's Q2 2026 data-center revenue was , more than double the prior year (The Next Web, 5 Aug 2026). AMD and Anthropic plan up to of Helios capacity, with the first 1 gigawatt in the first half of 2027. AMD's MI455X claims tokens per second per GPU on DeepSeek-R1, up to above Nvidia Vera Rubin's latest published MLPerf offline result. That is AMD's own testing.
What challenges it
- Nvidia's position is still very strong. JPMorgan notes Nvidia's gross margin and operating margin in 2023–2025, and says the company held 83% of AI chip revenue over that period. Its data-center business remains several times AMD's.
- Training is still CUDA-bound. One analysis puts Nvidia's training share above 90%, and says frontier libraries are CUDA-first (Stratrix, citing Nvidia's FY2025 10-K and other secondary sources). Stacks built on TensorRT-LLM or FlashAttention 3 have no viable non-CUDA path, according to Spheron's secondary blog post.
- Porting is easier than optimizing. AMD's own HIPIFY test preserved of CUDA source in one small, vendor-chosen application. AMD says this result does not generalize, because of CUDA library use, inline PTX, and architecture-specific code. An independent analysis says the real constraint is the library ecosystem, not the language. Its example was an MI300X running at of H100/H200 LLM inference performance, because kernels arrive on AMD months later or never.
- AMD's benchmark evidence is not a head-to-head with Nvidia. AMD's ROCm 10 post reports inference and training gains, but these compare a ROCm.AI preview with ROCm 7.0 on the same MI355X hardware. AMD says these results should not be presented as universal. Its MLPerf claim against Nvidia's B200 is also company-reported; Stratrix cites of B200 server performance.
- Concentration risk. AMD's data-center growth depends on a small number of very large buyers, according to The Next Web. The Anthropic deal is planned capacity, not delivered revenue.
- Nvidia's ecosystem is not standing still. The Next Web notes that Nvidia continues to deepen CUDA and partner lock-in.
Could a weaker CUDA moat help AMD win more demand?
Yes, under three conditions:
- The win is in inference first. Inference is standardized and high-volume, and the libraries that matter (vLLM, SGLang) can be validated per model. Training is far less likely to move.
- AMD's software makes deployment turnkey. ROCm 10 packages validated vLLM and SGLang containers, ROCm.AI, and Hyperloom for optimization. Those tools need to work on production codebases, not just AMD's demos.
- Large buyers choose AMD for cost or supply reasons. This is the pattern in the Anthropic and OpenAI agreements. Cembalest argues that AI labs and cloud providers have partnered with Broadcom or chosen AMD for cost-effective processors. But custom silicon from Broadcom, Google TPU, and Amazon Trainium are also taking share, so AMD is competing with other non-Nvidia options, not only with Nvidia.
My read: a weaker moat helps AMD more than it helps the broader market, because AMD's realistic gains are in inference capacity bought by a few large labs and clouds. It does not yet show AMD displacing CUDA in the mass market of smaller developers and enterprise training.
Signposts that would change this view
- Helios systems come online on the 2027 schedule, and the customer's production workloads run on ROCm without bespoke engineering.
- Independent (non-AMD) inference benchmarks on identical workloads and software baselines, including the August ROCm.AI release.
- Frontier training workloads start running on AMD at scale. This is the clearest sign the moat is breaking in its hardest segment.
- Nvidia's contracted share in the next JPMorgan or other allocation updates, measured on a consistent basis.
- Evidence that CUDA-first libraries (FlashAttention, TensorRT-LLM) ship on ROCm within weeks rather than months.
Caveats
- I could not access the full JPMorgan note. The JPMorgan Asset Management page contains only the summary. The figures above come from NDTV Profit's and Seoul Economic Daily's reporting of the note, so they are secondary and should be checked against the full report.
- The SemiAnalysis report on the CUDA moat is paywalled; I relied only on its headline and summary.
Sources
- J.P. Morgan Asset Management, "The Icarus Files: What's next for Nvidia and AI" (Michael Cembalest; summary page, estimated 14 Nov 2026): https://am.jpmorgan.com/us/en/asset-management/liq/insights/market-insights/eye-on-the-market/the-icarus-files-whats-next-for-nvidia-and-ai/
- NDTV Profit, report on JPMorgan's "Icarus Files" (secondary summary of the note): https://www.ndtvprofit.com/markets/nvidia-faces-a-new-challenge-from-the-very-companies-driving-the-ai-boom-jpmorgan-says-earnings-nvidia-stock-valuations-wall-street-ai-12332210
- Seoul Economic Daily, "Nvidia's Share of AI Compute Falls to 39% From 83%" (16 Nov 2026): https://en.sedaily.com/world/2026/11/16/nvidias-share-of-ai-compute-falls-to-39-percent-from-80
- AMD, "AMD ROCm 10: A Simpler Path to Production AI on AMD Instinct GPUs" (27 Aug 2026): https://www.amd.com/en/blogs/2026/amd-rocm-10-a-simpler-path-to-production-ai-on-amd.html
- AMD Newsroom, "AMD Advances the Instinct MI400 Series for AI and Scientific Computing" (12 Nov 2026): https://newsroom.amd.com/news/amd-advances-instinct-mi400-series-ai-scientific-computing/
- AMD, "From CUDA to AMD ROCm Software Without Starting Over" (23 Jul 2026): https://www.amd.com/en/blogs/2026/from-cuda-to-amd-rocm-software-without-starting-over.html
- Magica, "AMD's ROCm.AI pitch is a workflow, not CUDA compatibility" (23 Jul 2026): https://magica.com/news/amd-rocm-ai-cuda-migration-workflow
- SDxCentral, "Beyond CUDA: Inside the push to loosen Nvidia's grip on AI computing" (19 Jan 2026): https://www.sdxcentral.com/analysis/beyond-cuda-inside-the-push-to-loosen-nvidias-grip-on-ai-computing/
- The Next Web, "AMD's record quarter is really a bet that open beats Nvidia's moat" (5 Aug 2026): https://thenextweb.com/news/amd-q2-2026-earnings-rocm-open-source-nvidia
- Stratrix, "5.9 Million Developers Keep Nvidia Winning Against Chips With 1.5x the Compute" (18 Sep 2026; secondary, cites Nvidia FY2025 10-K and other sources): https://www.stratrix.com/moat-anatomy/cuda-as-a-moat-why-developers
- SemiAnalysis, "Can AMD break the CUDA Moat? AMD Advancing AI 2026" (25 Jul 2026; paywalled, headline and summary only): https://newsletter.semianalysis.com/p/can-amd-break-the-cuda-moat-amd-advancing