The binding constraint on AI capacity is no longer chips — it's the electrical grid. Nearly half of the U.S. data centers planned for 2026 (about 7 GW of 12 GW announced) have been canceled or delayed; only ~5 GW is under active construction, and the shortage has shifted onto transformers, switchgear, and batteries rather than compute silicon. Power transformers now run ~128 weeks and generator step-up units ~144 weeks, with high-capacity units quoted four to five years out — versus roughly two years before 2020. The delay isn't capital: the four largest hyperscalers are expected to spend more than $650B on AI infrastructure in 2026; the money is committed, but the power to energize it is years behind, with interconnection queues in Northern Virginia, Phoenix, and Dallas now running four to seven years.
The demand side is the other half of the equation — and it's extraordinarily concentrated. Per Ramp's June 2026 AI Index (70,000+ U.S. firms):
The whole thesis is the "when the rest catch up" question. Today's median is $11.38 against a top-1% of $7,450; if agentic AI keeps expanding what companies automate and token spend becomes the third major cost center after people and software, the top 1% may look like the median within a few years — a demand wave measured in orders of magnitude, landing against a supply side that can't add gigawatts on the same timescale. Falling unit costs don't relieve the pressure: token prices have dropped 90%+ since 2023, yet total corporate AI spending doubled since late 2025 — Jevons paradox in action, which is why Anthropic and others have shifted to usage-based billing. The honest counterweight: Goldman Sachs argues falling per-token costs drive hyperscaler margin inflection rather than scarcity pricing, and some analysts contend demand is inflated by "tokenmaxxing" vanity metrics. A defensible thesis, not a settled call.