If you invest in AI infrastructure, judge how long upstream compute can keep earning, or allocate across data centers, power, and chips — here’s a judgment that traditional-utility P/E badly underrates: the power supply curve is the master switch on the duration of the entire AI upstream profit pool.

The book so far covers how the market pushes profit to the two ends; this part covers how three exogenous forces price the “exchange rate” of those ends — physics, politics, law. Physics sets the depth of upstream scarcity. Start with the hardest, most underrated one: power.

AI’s Expansion Limit Is Shifting From “How Smart the Algorithm Is” to “How Much Headroom the Grid Has”

Take it literally. The IEA’s Energy and AI gives the bedrock numbers: global data-center electricity demand is projected to double by 2030 to about 945 TWh — slightly more than all of Japan’s national consumption today. The AI-optimized data-center portion is projected to quadruple. The US is the epicenter: data centers will account for nearly half of US electricity-demand growth from 2024 to 2030.

In one line: a brand-new electrical load equal to a developed industrial nation has to be crammed, within five years, mostly into the US grid. And the structural trouble the IEA flags: data centers don’t spread evenly like EVs — they’re highly geographically concentrated, thousands of megawatts piled onto a few parcels, hitting local grids far harder than the aggregate number implies.

Interconnection Queues: Expansion Speed Hits the Grid’s Physical Speed

The demand numbers are scary, but what actually stalls AI is supply-side physical lag. Building a data center takes a year or two; getting enough power to it takes far longer. Per grid operators like PJM (as reported), AI-infrastructure projects that went live in 2025 averaged over seven years to reach operational status. In hubs like Northern Virginia, Phoenix, Dallas, interconnection queues run four to seven years. Worse, the bottleneck is migrating from “the queue” down to harder links — high-voltage substations have 3-to-5-year lead times; transformers and switchgear are short across the board.

The consequence is concrete and brutal: a campus that only joined the Northern Virginia queue in mid-2026 realistically won’t draw real grid power until 2030–2033. Those loudly announced multi-gigawatt pipelines get translated by this queue curve into a reality curve that fully energizes in the 2030s, far behind the announcement cadence. Model cost curves iterate monthly; building the data center takes a year or two; powering it takes five to ten years. The industry’s expansion speed is hitting the grid’s physical speed — and the grid obeys none of AI’s exponential curves.

Why Power Is “the Wall”

Now the central judgment. Upstream scarcity comes in three kinds: engineering scarcity (chip design — caught up by big-customer in-housing and second sources), cyclical scarcity (HBM — bottomed out by capacity cycles), and physical scarcity (power). Only power has no erosion mechanism — there’s no “open-source generation,” no “in-house physics.” You can route around a chip supplier; you can’t route around thermodynamics.

But power’s key role isn’t just being scarce itself — it’s holding up the other scarcities. Imagine a world with no power constraint: compute can be piled infinitely, so “capacity-type scarcities” like chip output and data-center capacity get matched by capital sooner or later — upstream excess profit gets competed thin like any capital-chased capacity, and compute slides toward memory-like commoditization.

Power blocks that slide. Because energizing takes five to ten years and interconnection capacity is physically finite, compute can’t be piled infinitely fast — so capacity-type scarcity stays tight and upstream excess profit keeps its duration. Power is the wall: it pins a capacity scarcity that capital would otherwise erase fast onto a level set by physical lag.

Hence the line that matters most for valuation: the power supply curve is the master switch on the duration of the entire upstream profit pool. The deeper the gap and slower the interconnection, the harder upstream scarcity (chips included) and the longer the excess-profit duration; the day power opens up (nuclear restarts, SMRs land, storage and grid upgrades bite), upstream scarcity gets “released” and the slide to commoditization accelerates. Watching AI upstream duration is, at bottom, watching the power supply curve.

(Heading off one rebuttal: efficiency is soaring, per-unit-intelligence power use is plunging — doesn’t that save us? Not in the IEA baseline — Jevons: efficiency makes intelligence cheaper, usage explodes, and each efficiency gain is more than eaten by added usage. Efficiency shrinks the denominator; the numerator grows faster. After fully accounting for efficiency, the IEA still gives “demand doubles.”)

Three closing lines. One: track the power supply curve as the master switch on upstream duration — queue years, transformer/substation lead times, new interconnection capacity predict upstream profit duration better than any model benchmark. Two: bet along the bottleneck — not just generation, but the overlooked transmission, transformers and switchgear, high-voltage substations, and the siting rights that hold interconnection capacity; these are priced with the most lag. Three: watch for “release” signals (nuclear restart, SMR commercialization, storage biting) — they’re the early warning to shorten upstream duration.

The AI upstream name you’re watching — is its scarcity the physical kind held up by power, or the capacity kind capital will catch up to? Comments open.

Sources (June 2026): data-center power doubling to ~945 TWh by 2030, AI portion quadrupling, US ~half of growth (IEA Energy and AI 2025, A-grade); interconnection queues 4–7 yrs, AI infra avg >7 yrs to operation, substation lead 3–5 yrs, 2026 entrants energizing 2030–2033 (grid-operator data incl. PJM, as reported, B+).

— From Chapter 12 of a book in progress, working title The Deflation Sandwich

#AI #Energy #DeflationSandwich