AI growth outlook stays strong as risk-free returns come under pressure

Vera Rubin's real strategic variable isn't raw compute—it's the migration of the industry bottleneck from transistor process to advanced packaging and interconnect. CoWoS-L and SoIC capacity have effectively become the 'throttle valve' for global AI compute, and the bargaining power of packaging operations in Taiwan, China is now rivaling that of wafer foundries themselves. Sovereign AI builds in France and Saudi Arabia open a second demand curve beyond hyperscalers. These buyers face lower export-control sensitivity but demand extreme supply-chain continuity, forcing Nvidia into a 'dual-track delivery' architecture while concentrating advanced-packaging geographic risk. On competition, AMD's MI400 and custom ASICs (TPU, Trainium) keep eroding general-purpose GPU share, yet 2027 capex expansion signals the market is large enough for multipolar supply. The structural risk isn't demand—it's the HBM4 yield-ramp versus packaging-capacity timing mismatch. 12–24-month call: 'packaging is compute' becomes the industry pricing anchor, and OSAT players gain, for the first time, bargaining parity with IDMs in the value chain.