Morning Analysis — 2026-07-19

Dek

The AI power gap just got a number — 100 gigawatts — and the capital isn't waiting in the interconnection line anymore.

The Big Shift

Bank of America now projects that AI data-center demand will outrun planned US utility capacity by more than 100 gigawatts through 2030, pushing builders toward on-site gas and batteries rather than the public grid (¹). That's roughly a hundred large power plants of demand with no home. Why it matters: the bottleneck for AI has officially moved from chips to electrons, and the market response is to build power *beside* the load instead of queuing for grid hookups that take years. What it signals: a durable bid under anyone who owns firm, dispatchable generation — nuclear operators, gas turbines, storage — and a widening premium for "behind-the-meter" power.

Analysis

Power is the binding constraint, and existing nuclear is the cleanest way to fill it. The NRC is weighing a subsequent license renewal for Constellation's Ginna plant (²), and a Seeking Alpha piece frames Constellation bluntly: "the inventory is smaller than the gap" (³). Translation: there are only so many always-on, carbon-free megawatts to sign away, and every hyperscaler wants them. This is the core thesis in one line — scarce firm power gets repriced upward, and the owners capture it.

The grid can't keep up, so capital routes around it. PJM, the mid-Atlantic grid operator, hit its price cap *again* at its latest capacity auction (⁴) — a signal that buyers will pay the maximum allowed for a promise of future power. Meanwhile the political version of "just build more" is stalling: Trump has spent billions but added few new gigawatts to the grid, having blocked tens of GW of wind and solar (⁵). What this implies: the grid is not the answer on the AI timeline, which validates the on-site gas-plus-storage pivot BofA describes and rewards developers who bring their own generation.

Demand keeps proving itself on the compute side. TSMC — the company that actually fabricates the AI chips — says the supercycle "just got stronger" (⁶), and the pick-and-shovel trade (optical interconnects, chip suppliers, hyperscaler capex) is still running hot across the watchlist. This matters because it removes the easy bear case: the demand isn't softening, so the power squeeze isn't a passing spike — it's structural. Every incremental chip shipped is another claim on electricity that doesn't exist yet.

The materials layer is quietly tightening under the same demand. The IEA now calls the copper supply outlook "worsened considerably," partly from Middle East disruption and a Beijing export move (⁷), and rare earths are picking up AI as a genuinely new demand source on top of defense and clean energy (⁸). Copper is the wiring of every data center and every new generator; rare earths are in the magnets. The implication: even if you solve generation, the physical build-out has its own input scarcities — and China/Greenland supply politics (⁹) can throttle it.

Geopolitics sits underneath all of it as a fat tail. Iran is stepping up attacks on shipping in the Strait of Hormuz, with a maritime-risk CEO calling it a "worst case scenario" (¹⁰), and Iraq/Chevron are reportedly planning a Syria route to bypass the strait entirely (¹¹). Why it matters for the thesis: a Hormuz shock spikes oil and, more importantly for us, the price of the gas that data centers are now leaning on for on-site power — the exact hedge BofA says developers are reaching for could get more expensive to run.

What Would Prove Us Wrong

• Grid catches up faster than feared. If interconnection queues clear or PJM's next auction clears *below* cap, the "build beside the grid" premium compresses — hitting the firm-power scarcity thesis (nuclear/gas owners like Constellation) directly.
• AI capex actually rolls over. A TSMC guidance cut, a hyperscaler trimming its 2026–27 data-center capex, or canceled projects (note the already-canceled Stack site in Hesse, Germany) would undercut the demand leg the whole power squeeze rests on.
• Materials fears ease. Copper prices falling back or a China/Greenland rare-earth thaw would signal the input-scarcity thesis is overstated — watch the LME copper curve and any concrete rare-earth supply deals over the next few weeks.

Thesis Impact

No thesis-moving signal.

Inflection Radar

[emergent] Grid Modernization Scale | Utility-scale procurement for 1,650 GWh renewables and 465 MW storage signals accelerating grid stress/decarbonization mandates. | Touches: NEW | ¹²

[emergent] Traditional Industry AI Adoption | A spirits company securing contracts for AI data center construction and network infrastructure highlights the rapid, non-tech sector adoption of compute power. | Touches: NEW | ¹³

QA & Caveats

No issues found.

Sources

1. AI data center growth could force US utilities to rethink generation plans, BofA says utilitydive.com
2. Constellation Energy Generation, LLC; R.E. Ginna Nuclear Power Plant; Subsequent License Renewal Application federalregister.gov
3. seekingalpha.com seekingalpha.com
4. PJM Once Again Hits Its Price Cap at Latest Auction heatmap.news
5. Trump has spent billions on energy. An electricity boom has yet to materialize. Trump vowed to unleash energy to meet su reddit.com
6. seekingalpha.com seekingalpha.com
7. Acid test: IEA warns copper supply outlook has “worsened considerably” northernminer.com
8. AI emerges as new driver of rare earth demand, Sprott says mining.com
9. ETM battles Greenland over rare earth ‘expropriation’ northernminer.com
10. Oil tankers face 'worst case scenario' in Hormuz as Iran steps up attacks on ships, maritime risk CEO says reddit.com
11. Iraq and Chevron Are Planning a Route Through Syria to Evade Strait of Hormuz heatmap.news
12. Hawaiian Electric seeks 1,650 GWh renewables, 465 MW storage procurement - Solarbytes news.google.com
13. Agencia Comercial Spirits Ltd Announces Agreements for Indonesia AI Data Center Construction, Network Infrastructure Pro news.google.com