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, announced at GTC 2026, after roughly three months, with the replacement carrying the same name but delivering approximately half the performance due to packaging limits [1][2]. SemiAnalysis, which broke the story, simultaneously projects NVIDIA datacenter compute revenue running 20% above sell-side consensus for 2H FY2027, attributing the Rubin ramp delay primarily to HBM4 supply issues now resolved [6]. On the alternative-silicon front, Anthropic runs substantial Claude Code inference on AWS Trainium and trains Claude models on Google TPUs, backed by a $100 billion-plus, ten-year commitment to AWS technologies [7][9]. The Trainium narrative is contested: Enertuition argues the platform is a strategic failure requiring a reset [10].

NVIDIA faces simultaneous pressure on product execution and on the durability of its datacenter hardware dominance. A redesigned Rubin Ultra at half the original performance, combined with growing frontier-lab adoption of Trainium and TPUs, tests whether NVIDIA's supply-constrained revenue lead can persist into 2027 even as its product roadmap shrinks and alternatives gain credibility.

NVIDIA announced the 4-die Rubin Ultra GPU at GTC 2026 in March, then cancelled the original design roughly three months later. The replacement product retains the 'Rubin Ultra' name but is approximately half the size and delivers roughly half the real-world performance of the announced configuration, with TechTimes and SemiAnalysis attributing the cancellation to packaging limits that made the original design unmanufacturable at scale [1][2]. WCCFtech reports that both the Rubin and Rubin Ultra platforms face broader design and specification issues beyond the Ultra cancellation, and identifies AMD's MI500 as a competitor positioned for the second half of 2027 [3]. Tom's Hardware separately reported that Rubin CPX accelerators were removed from NVIDIA's roadmap [4], and NVIDIA cancelled some planned rack configurations, with SemiAnalysis reporting material downstream implications for HBM memory demand [5].

Despite these product setbacks, SemiAnalysis projects a strong NVIDIA second half overall. In a June 30, 2026 post, SemiAnalysis estimated NVIDIA datacenter compute revenue would run 20% above sell-side consensus for 2H FY2027, explaining that the Rubin ramp had been delayed by HBM4 supply issues now resolved and that front-end wafer supply has been built up to support the ramp [6]. This creates a dual position: SemiAnalysis frames the Rubin Ultra cancellation as a manufacturing execution failure while simultaneously arguing that demand exceeds even a reduced supply, enough to produce above-consensus revenue. Product-level setbacks and business-level outcomes are not the same thing when demand structurally exceeds capacity.

The competitive landscape on the alternatives side centers on Anthropic's departure from NVIDIA-centric infrastructure. SemiAnalysis reports that Anthropic runs a substantial share of Claude Code inference on AWS Trainium and trains Claude models on Google TPUs [7], characterizing this as something that would have been unimaginable a year earlier. Reddit discussion confirms Claude Opus was trained on AWS Trainium2 [8], consistent with Anthropic's April 2026 announcement committing over $100 billion over ten years to AWS technologies including Trainium generations 2 through 4, with nearly 1 GW of Trainium capacity expected online by end of 2026 [9]. Amazon deepened its financial stake simultaneously, committing up to $20 billion in additional investment on top of a prior $8 billion [9]. The Trainium narrative is actively contested: an analysis titled 'Amazon Trainium Is A Disaster; Strategy Reset Needed' argues the platform has not delivered on its promise [10], directly contradicting the framing from both SemiAnalysis and Anthropic's public commitments.