Written by Ralph Sun
The binding constraints on the AI build-out are no longer leading-edge wafers. They are memory, advanced packaging and electricity — and the evidence now shows up in device prices, subcontracting agreements and construction figures rather than in chip announcements.
Ask what is scarce in AI and the answer has changed. Two years ago the constraint was leading-edge logic capacity. Today that is largely a solved problem, and the shortage has moved downstream into three layers that are harder to expand quickly: memory, advanced packaging and power.
Memory is the most acute. Samsung, SK Hynix and Micron have shifted capacity toward AI customers, and IDC analysts expect the shortage to persist well into 2027. The scale of the squeeze is measurable in corporate language: the terms “memory prices” and “memory shortage” appeared in 473 company transcripts last quarter, according to AlphaSense data. Qualcomm’s quarterly earnings presentation attributed a 20% year-over-year drop in handset revenue to “unprecedented increases in memory pricing and supply constraints.” AMD’s chief executive told investors the company would “substantially increase our supply in 2027,” a timetable that concedes the present.
The downstream evidence is unambiguous. Nvidia raised the price of its Shield TV Pro by $100 on 2 October. Amazon increased prices on Echo, Kindle, Fire TV and Eero products by up to 60%, citing memory and storage costs. Samsung lifted US prices on the Galaxy S26 line by $100 or more. Raspberry Pi has raised prices repeatedly. Valve said its new Steam Frame headset was supposed to cost far less than $1,059. Acer’s chief executive expects PC prices to rise another 20% before easing in the second half of 2027, and sales of sub-$100 phones fell 64% year over year in the second quarter as manufacturers simply stopped shipping them.
Advanced packaging is the second constraint, and the clearest sign of how tight it has become is that TSMC is now subcontracting. Under a five-year agreement valued at $2 billion, GlobalFoundries will produce silicon interposers for CoWoS-S packaging at its Malta, New York facility, with volume production scheduled for the first half of 2028. Interposers are the silicon bridges that connect logic chiplets to high-bandwidth memory stacks.
The caveat matters more than the headline. Interposers are only one step in CoWoS. Attaching chiplets and memory stacks to them, assembling the result onto organic substrates and testing it all require separate capacity that the agreement does not address. Subcontracting interposers frees TSMC resources but does not by itself produce more finished AI processors. Meanwhile HBM stacks will continue to come from Asia for years, even as Micron and SK Hynix plan US assembly.
Power is the third and probably deepest constraint. Of 16 gigawatts of data-centre capacity announced for delivery in 2026, only about 5 gigawatts was under construction as of this summer, according to TD Economics. Transformers, switchgear and gas turbines carry multi-year lead times, and grid interconnection queues are measured in years rather than quarters. A gigawatt that cannot be energised is a gigawatt that cannot host accelerators, however many are available.
Secondary pressure is building in CPUs and analogue parts. Intel has reportedly planned a roughly 10% price increase on PC processors, with AMD expected to follow, partly reflecting higher foundry pricing.
It is worth stating plainly what is not constrained. Leading-edge logic wafers are not the bottleneck. Sony has said it secured the memory needed for PlayStation 5 volumes in its current fiscal year. And not every category is in crisis simultaneously; the constraint set rotates as capacity comes online.
The practical implication is that the AI build-out should now be read as a chain rather than a single input. Memory gates accelerator output, packaging gates memory integration, and power gates everything. Watch DRAM and HBM contract pricing, CoWoS assembly and test capacity announcements, and interconnection queue data — those three series will say more about the pace of AI deployment over the next year than any model release.
