SemiAnalysis: Nvidia and Memory High Margins Face a 2028 Shift, AI Infrastructure Constrained by Financing and Supply Chain

BlockBeats news, October 25 - SemiAnalysis founder Dylan Patel and Bernstein analyst Stacy Rasgon recently pointed out that while there is still room for AI infrastructure investment, it is facing the dual challenges of rising financing costs and supply chain constraints. The U.S. 10-year Treasury yield has breached 5%, and the unsecured bond yields of some emerging cloud service providers are approaching 15%, while non-hyperscale cloud vendors that undertake nearly half of construction demand often lack sufficient balance sheet support. Projects lacking long-term offtake agreements are increasingly difficult to obtain funding, and financing capability is becoming an important dividing line for AI expansion. Meanwhile, substrates, HBM, optical interconnect components, and cleanroom capacity all face bottlenecks that are difficult to resolve in the short term. Producing the same capacity of HBM requires approximately 3 to 4 times the number of wafers as DDR5, intensifying capacity competition; new cleanrooms will not be gradually put into use until 2027. By comparison, power shortages can be partially alleviated through on-site gas turbines, fuel cells, and other means.

Regarding the core industrial chain, the two analysts warned of a potential turning point in 2028. Nvidia is still supporting high selling prices through continuous performance improvements, but self-developed chip layouts by Google TPU, Amazon Trainium, and companies such as OpenAI are providing large customers with alternative choices and strengthening their bargaining power. Even if Nvidia's revenue continues to grow, its market share and profit margins may not rise in tandem.

In terms of memory, current HBM gross margins are as high as 88% to 90%, but the market has not assigned an extremely high valuation, with the core concern being the concentrated release of new capacity after 2028. If expansion speed exceeds demand growth by then, combined with slowing AI demand growth, current high prices and high profits will face reversal. On the enterprise side, most enterprises will prefer to purchase tokens through cloud services rather than build their own computing power, which means future demand for AI infrastructure will highly depend on the usage of inference services.