SemiAnalysis: Vera Rubin IRR jumps to 38%

SemiAnalysis argued in an April 30 post that Nvidia and Taiwan Semiconductor Manufacturing Co. have more room to raise prices on AI hardware because model labs and cloud operators are capturing a larger share of the economics. The post, listed in SemiAnalysis’ archive as “TSMC, Vera Rubin VR NVL72: V for Value,” paired that claim with a Vera Rubin financial model showing how higher system pricing could change project returns. The firm’s example is the number that has circulated most widely: a hypothetical 40% increase in GPU pricing raised a Vera Rubin project’s internal rate of return to 38% from 15.3%, according to the social-media summary of the piece and secondary coverage citing the report. SemiAnalysis also published a separate AI cloud TCO model describing project IRR, equity IRR and cash-generation analysis for GPU clusters. (newsletter.semianalysis.com) ### Why is SemiAnalysis focusing on Vera Rubin? Nvidia’s Vera Rubin NVL72 is the company’s next rack-scale AI platform, and Nvidia says it combines 72 Rubin GPUs, 36 Vera CPUs, ConnectX-9 SuperNICs and BlueField-4 DPUs in a single system aimed at “AI factories.” Nvidia has said the platform is designed for lower cost per token at scale, which makes it a useful reference point for any argument about who captures value in the AI stack. (macrostream.ai) SemiAnalysis’ choice of Vera Rubin matters because the system sits near the top of the current AI infrastructure spending cycle. Its April 30 archive entry appears alongside other SemiAnalysis pieces on GPU rental pricing, AI silicon shortages and cluster economics, showing the firm is treating pricing as part of a broader supply-and-demand model rather than a one-off product call. ### What does “underpricing AI value” mean here? (developer.nvidia.com) SemiAnalysis’ broader April 30 article, “AI Value Capture - The Shift To Model Labs,” said value in AI has been moving toward model labs and inference providers even as hardware remains supply-constrained. The post said some parts of the hardware stack have already repriced, citing memory prices and rental GPU contracts, but said two companies with “incredible pricing power” had not moved much. Search snippets and secondary summaries identify those companies as Nvidia and TSMC. (newsletter.semianalysis.com) That is an argument about bargaining power, not a reported company plan. SemiAnalysis is effectively saying that if downstream customers are earning strong margins on tokens and inference, upstream suppliers could charge more without destroying demand. That interpretation comes from SemiAnalysis’ framing in the April 30 posts and not from any announced Nvidia or TSMC pricing change. ### How does TSMC fit into the pricing case? SemiAnalysis has separately reported that TSMC’s N3 capacity is being pulled toward AI accelerators. (t.co) A March 12 SemiAnalysis archive entry was titled “The Great AI Silicon Shortage — TSMC N3 Wafer Shortages, Memory Constraints, Datacenter Bottlenecks,” and outside summaries of that work said AI accelerators could take roughly the mid-80% range of N3 capacity by 2027. TSMC itself has reported that AI demand is driving results. The company posted first-quarter 2026 revenue of NT$1.13 trillion, up 35% year over year, according to its April earnings coverage, with analysts and the company pointing to sustained AI-chip demand. ### Why are memory and utilization figures part of the story? S&P Global Market Intelligence reported in January that AI-driven demand for high-bandwidth memory was diverting DRAM capacity and tightening supply for conventional memory. (newsletter.semianalysis.com) That supports the supply-side backdrop behind SemiAnalysis’ claim that multiple chokepoints — logic, packaging and memory — are all getting tighter at once. (cnbc.com) The next concrete markers will come from company disclosures rather than commentary. Nvidia has said Vera Rubin systems are aimed at deployment in 2026, while TSMC’s future quarterly results and SemiAnalysis’ own cluster-economics updates will show whether tighter capacity translates into actual price increases. (developer.nvidia.com) (spglobal.com)