A surge in token demand driven by agentic AI workflows has created a severe compute shortage, with on-demand GPU prices rising even for older Hopper-generation chips and every neocloud cluster reportedly locked up, according to SemiAnalysis. The shortage is concentrated at TSMC's N3 process node, where AI accelerators from Nvidia, AMD, Google, AWS, and Meta are all converging in 2026; AI-related N3 demand is projected to account for just under 60 percent of total N3 output this year. Nvidia is transitioning from 4NP-based Blackwell to N3P-based Rubin, Google's TPU v7 is already in production on N3E, and AWS's Trainium3 is ramping on N3P, while TSMC's capacity expansion has not kept pace. SemiAnalysis notes that Google's 2026 capital expenditure expectations have roughly doubled versus prior consensus estimates, primarily driven by datacenter and server spending.
The concentration of every major AI accelerator on a single TSMC process node creates a systemic supply bottleneck that constrains how fast hyperscalers can expand data center compute capacity, regardless of how much capital they commit. Capacity pressure at TSMC N3 is already redirecting some customer interest toward Intel Foundry and Samsung Foundry, which could shift long-term semiconductor supply chain dynamics for the entire data center industry.
SemiAnalysis is a highly credible, closely followed publication on chip supply dynamics; an analysis piece on AI silicon shortages is directly relevant to data center capacity planning and hyperscaler buildout timelines.