NVIDIA Cancels 4-Die Rubin Ultra and Faces Structural Market Share Erosion from Trainium, TPUs, and AMD

NVIDIA cancelled the original 4-die Rubin Ultra GPU just three months after announcing it at GTC 2026, citing manufacturing execution concerns; the replacement product carries the same name but delivers roughly half the performance of the original design [1]. Separately, SemiAnalysis reports that Anthropic already runs a substantial share of Claude Code inference on AWS Trainium and trains Claude on Google TPUs rather than NVIDIA hardware [4]. This is consistent with a major Anthropic-Amazon deal announced in April 2026, committing over $100 billion to AWS technologies including Trainium generations 2 through 4 [5]. SemiAnalysis frames both developments as concrete evidence that NVIDIA's CUDA moat is eroding in practice, compounded now by a roadmap setback [6].

The Rubin Ultra cancellation is the first concrete sign that NVIDIA's next-generation roadmap has slipped, and it arrives at a moment when a leading frontier AI lab has already moved substantial production workloads to non-NVIDIA silicon. If manufacturing difficulties continue, the window for alternatives to expand their foothold widens before NVIDIA can reassert hardware leadership.

Three months after NVIDIA announced the 4-die Rubin Ultra at GTC 2026, the company cancelled the original design due to manufacturing execution concerns [1]. A replacement product retains the 'Rubin Ultra' name but is approximately half the size and delivers roughly half the real-world performance of what was announced [1]. SemiAnalysis, which reported the cancellation on June 29, 2026, also noted that NVIDIA has cancelled some planned future rack configurations, with downstream implications for HBM memory demand quantified in a separate research report [2]. Tom's Hardware separately reported that Rubin CPX accelerators were also removed from NVIDIA's roadmap [3].

Concurrent with these hardware setbacks, SemiAnalysis reports that Anthropic — the operator of Claude Code, which it describes as the most successful AI agent — runs a substantial share of Claude Code inference on AWS Trainium chips and performs Claude model training on Google TPUs, not NVIDIA hardware [4]. SemiAnalysis characterized this level of Trainium and TPU adoption as something that 'would have been unimaginable' just a year ago [4]. The infrastructure choices align with a large partnership Anthropic announced with Amazon in April 2026: Anthropic committed more than $100 billion over ten years to AWS technologies spanning Graviton and Trainium generations 2 through 4, with nearly 1 GW of Trainium2 and Trainium3 capacity expected online by end of 2026 [5]. Amazon simultaneously deepened its financial stake, investing an additional $5 billion immediately with up to $20 billion more committed on top of a prior $8 billion [5]. Anthropic's annualized run-rate revenue exceeded $30 billion at announcement, up from approximately $9 billion at end of 2025 [5].

SemiAnalysis ties these threads together with an explicit competitive argument: NVIDIA's market share is being eroded by AWS Trainium, Google TPUs, and AMD chips, and manufacturing execution failures will only accelerate those losses [6]. The counterpoint comes from a Substack analysis titled 'Amazon Trainium Is A Disaster; Strategy Reset Needed,' which argues Trainium has not delivered on its promise [7] — a direct challenge to the bullish framing from SemiAnalysis and from Anthropic's own public statements about its AWS commitment. Neither view has been independently validated with performance benchmarks in this set of sources.

The full picture remains incomplete. The Anthropic-Amazon announcement does not specify what fraction of Claude workloads run on Trainium versus other hardware, and Anthropic's simultaneous multi-cloud positioning — Claude is available on AWS Bedrock, Google Cloud Vertex AI, and Microsoft Azure Foundry [5] — implies some inference runs on NVIDIA-backed infrastructure on other clouds. The scope of customer impact from the Rubin Ultra cancellation and any knock-on effects on data center build-out plans are similarly not yet public beyond early media and analyst coverage.