Cranes are still moving somewhere inside a data center facility that is being built outside of Phoenix. Steel is rising. Infrastructure for cooling is being run thru structures that were nonexistent eighteen months ago. The speed is very impressive, yet the parts that those structures are being constructed to house are frequently still waiting months to be delivered in allocation queues. For the better part of two years, Wall Street analysts have been attempting to explain this disparity between actual chip supply and physical fabrication in a single graphic.
Major investment banks are currently projecting that there won’t be a short-cycle supply shortage. It’s more structural in nature, a mismatch between the speed at which the biggest IT businesses in the world are investing capital and the speed at which the semiconductor industry can actually manufacture the chips needed to use it. Both sides of the equation have led analysts to conclude that conditions won’t significantly stabilize until late 2027 or possibly 2028. In a sector where cycles are typically measured in quarters, that is a long period.
It’s difficult to take in the spending figures without pausing. It is anticipated that between 2025 and 2027, the combined capital expenditure of the main cloud providers and AI platform businesses would exceed one trillion dollars, with a significant amount of that money going directly into AI-optimized infrastructure. Demand for servers as well as the specialized memory and connection components that enable AI workloads to function is increased with each new data center order. The component that is now receiving the most attention is high-bandwidth memory, which is the dense, costly, power-hungry memory that is stacked directly next to AI accelerators because it is the one with the most severe supply limitations.
Micron, SK Hynix, and Samsung have all moved substantial production capacity away from the consumer and conventional enterprise memory markets in favor of HBM. Although HBM commands far greater margins, this reallocation makes financial sense. However, it also concentrates manufacturing in a product category where capacity is already limited and adding new output requires years of retooling rather than months. The industry may have misjudged the rate at which AI workloads will transition from training-only to full-scale inference deployment, and this miscalculation is now manifesting as a supply gap that cannot be filled by purchase agreements alone.
The hyperscalers have reacted by locking in, which is what big buyers always do when they’re concerned about access. Companies now routinely enter into multi-year supply agreements with chip manufacturers, agreeing to future production runs well in advance of such runs starting. Contracts are signed and capacity is reserved, creating a sort of fictitious stability on paper, but the fundamental production limitation remains unresolved. Agreements do not alter the factory’s quarterly production cap if it is limited.
TSMC is making significant investments to put sophisticated packaging and logic production closer to its major clients, and its expansions in Arizona and Japan are genuine. However, those who anticipate that those facilities will alleviate the present shortfall are misguided. From groundbreaking to significant volume output, major fabrication expansions take three to five years. Although the Arizona factories are making headway, they won’t be operating at maximum capacity in time to significantly change the near-term supply scenario.
Watching this unfold gives me the impression that the AI infrastructure cycle is truly different from previous semiconductor booms—not because demand is more irrational, but rather because the deployment velocity has shortened the typical adjustment window. Prior to the shortfall being structural, a spike in demand would eventually generate enough price signals to attract additional capacity. This time, the typical feedback loop hasn’t had time to function because the demand came in so swiftly and in such large quantities. It’s still unknown if 2027 is the correct forecast or if the shortage lasts longer. However, it doesn’t appear like the experts who have been monitoring this the most closely are changing their timelines sooner.
