AI Compute Sold Out Past 2028; Supply Chain Strain Spreads to Adjacent Industries

Amazon CEO Andy Jassy told a16z co-founder Ben Horowitz that Amazon has no available compute capacity through at least 2028, a disclosure corroborated separately with the framing that even a hyperscaler with enormous capital access cannot secure capacity. Google CEO Sundar Pichai confirmed a similar picture, saying customers seeking 2027 compute find it incredibly difficult to secure, that Google leased short-term compute in the prior year to accommodate customers, and that Google Cloud is at any given moment foregoing short-term revenues to make compute tradeoffs supporting frontier AI development. Pichai also acknowledged that AI labs are living with compute bets made approximately two years ago, and that no company accurately forecast the scale of demand that materialized.

The hardware supply chain data reinforces the executive statements. TrendForce data shows HDD lead times at 50 weeks versus a normal 16, DRAM at 20 weeks versus 8, enterprise NAND at 16 weeks versus 8, and ABF substrates at 48 to 56 weeks versus a normal 12, with hyperscalers securing purchasing commitments extending to 2028 while still anticipating shortages. TSMC CoWoS packaging capacity is fully allocated through at least mid-2027, pushing NVIDIA H100 and H200 lead times to 36 to 52 weeks. Chinese firms including Alibaba, Tencent, and ByteDance ordered roughly 6 million Nvidia RTX Pro 5500 GPUs totaling approximately $70 billion; Nvidia can supply roughly 1% within four months and reportedly told partners it can produce at most 3 million cards worldwide in 2027 as TSMC capacity shifts toward Rubin data center chips.

SemiAnalysis analyst Dylan Patel argues that while power is a real bottleneck, substrates, memory, lasers, and optics are more limiting, and that creative on-site power solutions such as turbines, reciprocating engines, and fuel cells are helping alleviate power constraints. Nearly 50 GW of data centers are under construction globally, including 35 GW in the US. Morgan Stanley estimates a potential 76.5 GW power shortfall in US data centers by 2029 even after accounting for all planned capacity.

The buildout is now drawing resources beyond the core semiconductor supply chain. Paint manufacturer PPG expects data center paint sales to triple in the current year, driven by AI construction demand for specialized coatings that resist fire, rust, and static electricity. Chip War author Chris Miller argued that data centers are consuming memory chips, workers, and resources at a scale that raises costs for industries that see no AI benefit. Fujifilm is investing approximately 10 billion yen to increase LTO tape production capacity fivefold relative to fiscal year 2025, citing generative AI as driving exponential growth in data archiving needs.