The $725 Billion Squeeze

In Texas, the AI buildout has found its simplest expression: a licensed electrician can now earn double wages wiring a data center instead of a house. The state's residential homebuilders are losing skilled workers to hyperscaler construction sites, and with 20,000 electricians retiring annually and no pipeline to replace them, Texas has begun importing electricians from Alabama, Arkansas, and Iowa through reciprocity agreements ^[1]. The capital is there. The labor is not. And the money, far from solving the problem, is what set the competition in motion. This is the pattern in miniature, and it is replicating across every physical layer the AI industry touches. In power markets, PJM Interconnection — the grid operator serving 65 million people across 13 states — saw capacity prices jump more than 1,000 percent from 2024 to late 2025. The transmission reached manufacturers directly: industrial electricity prices rose 31 percent in Pennsylvania and 26 percent in Ohio over the same period ^[2]. The mechanism is not mysterious. When data center demand hits a grid whose supply is fixed by transformer lead times and turbine backlogs, the price clears higher for everyone. In semiconductors, the same dynamic is playing out in memory chips. Nine US trade groups — representing automakers, retailers, and medical device manufacturers — warned the Treasury and Commerce secretaries in June that AI data centers are consuming a disproportionate share of memory chip capacity, threatening supply chains for cars, electronics, and medical devices ^[3]. Micron's high-bandwidth memory is sold out through 2027 and into 2028, with non-cancellable customer agreements extending to 2030 ^[4]. The CEO's assessment of when supply might catch up was blunt. ^[4][5]

We expect tight conditions to persist beyond calendar 2027 as a result of AI-driven demand across all segments coupled with structural supply constraints. — Sanjay Mehrotra

SK hynix's chair estimates the shortage could persist until 2030, because new fabrication plants take years to build ^[3]. The AI industry is not just buying chips — it is repricing them for every other industry that needs them. The grid itself is the deepest constraint. Substation transformer lead times now exceed 160 weeks — more than three years ^[6]. In Denmark, the grid operator Energinet imposed a moratorium after receiving connection requests totaling 60 gigawatts against peak demand of 7 GW ^[6]. In Ottawa, the utility's CEO told the city council that data centers now account for 60 percent of the connection queue and that the industry is asking the utility to build in two to three years what took 110 years to build originally ^[7].

We are being asked to build … in the next two to three years, what has taken us the better part of 110 years to build. — Bryce Conrad

Jefferies reported that in 2025 only 8.9 gigawatts of data center capacity came online against 21.1 GW of demand — a 12 GW structural deficit — and attributed the gap to supply chain bottlenecks in engineering labor, cooling equipment, and electrical components, not to a shortage of investment ^[8]. A single AI facility can require up to 50,000 tons of copper. The constraint is physical inputs and labor, not capital. And yet the spending projections are not converging on a stable number — they are accelerating. In April, Nvidia projected $3 to $4 trillion in global data center capital expenditure by 2030 ^[9]. In May, the top four hyperscalers were projected at roughly $600 billion for 2026 ^[10]. By July, the figure had climbed to $725 billion: Amazon at $200 billion, Microsoft and Alphabet up to $190 billion each, and Meta at $125 to $145 billion ^[11]. The numbers rise even as every physical constraint hardens. The brute-force strategy is now fracturing the hyperscalers themselves unequally. Alphabet's Google Cloud revenue surged 63 percent, and its shares rose. Meta raised its spending forecast without a precise product scaling plan, its shares fell, and the company was forced to sell $25 billion in bonds while preparing workforce reductions of 20 percent or more ^[11]. The same capital deployment that looks rational against a $2 trillion cloud backlog is straining the balance sheets of the companies deploying it — and the strain is not distributed evenly. The demand is real. Micron's supply is committed through non-cancellable contracts extending to 2030. Cloud backlogs total roughly $2 trillion ^[8]. Jefferies was direct about the constraint.

Demand for data centers continues to outpace supply, with hyperscaler capex accelerating and chip volume forecasts implying GWs of capacity ahead of feasible data center delivery. — Jefferies Group

Each actor, looking at its own backlog, makes a rational decision to spend. Collectively, they intensify the scramble for finite physical inputs — and the cost of that scramble lands on the hospital buying memory chips for imaging equipment, the homebuyer waiting months for an electrician, and the manufacturer in Pennsylvania paying 31 percent more for power.