How durable is the supply-constrained data center regime?

2026-07-30 · four research passes: energy first-principles, demand durability, component supply chains, siting opposition · companions: deep dive · distributions · not licensed financial advice

Verdict: the supply-constrained regime is durable through 2028 with high confidence, and plausibly through 2029-2030. Every independent constraint stack outlasts the investment window: gas turbines are sold out through 2029, transformers run 2.5-4 year lead times with order books into 2028-29, Samsung says the memory shortage worsens into 2028, and demand keeps compounding (tokens +4-7x/yr, hyperscaler lease commitments +63% YoY). The glut scenario requires capex to decelerate while capacity lands - but the physics says capacity can't land fast before 2029, and the commitments say capex isn't decelerating. Siting opposition is a 5-10% haircut, not a constraint. The one genuine break: enterprise agent ROI patience running out in the next 12-18 months - a demand-side event no supply analysis can rule out.

1 · The watts model (energy stream)

Year	US DC demand (avg GW)	New FIRM supply deliverable (GW/yr)	Verdict
2026	38-41	~8-10 (only 6.3 GW of new gas nationally; 86 GW additions are mostly solar/storage)	Balanced nationally, already short in PJM
2027	45-66	~10-14 (TMI restart, pre-2023 turbine orders)	Gap widens - turbines ordered today deliver 2028-30
2028	60-80	~15-20	Still short; PJM projects 15 GW shortfall by 2030
2029-30	70-95	~25-35 best case	Earliest broad easing 2029; full catch-up 2030-32, ERCOT clears first, PJM last

GE Vernova turbine backlog: ~116 GW, sold out through 2029; manufacturing ramps only 20→30 GW/yr (global, ~20% DC-earmarked).
Behind-the-meter gas: ~101 GW announced, ~2 GW actually online - same turbine bottleneck.
Nuclear: restarts + uprates = low single-digit GW through 2028; SMRs are 2030s.
Prices confirm scarcity: PJM capacity cleared $329/MW-day (+22% YoY, ~10x 2023 levels); EIA scenario has ERCOT +79% by 2027.

2 · Demand durability (the Jevons evidence)

Price -2 orders of magnitude, volume +2-3 orders: Google tokens/month 9.7T (2024) → 480T (2025) → 3.2 quadrillion (May 2026); OpenRouter independently +4-5x YoY. Cheaper inference has expanded total spend every single time.
Constraints are mechanistic: Anthropic throttled peak hours (Mar '26) and only raised limits after new capacity; MSFT CFO: "demand exceeds supply" into FY27; OpenAI pulled GPUs off Sora.
The router/efficiency risk has been live for 2 years and lost every round: routing + distillation cut per-task cost 40-85% - and total tokens grew anyway. Agents are why: 10-100x tokens per task vs chat; coding went ~11% → >50% of platform tokens in a year; Claude Code took Anthropic $9B → $47B run-rate in 5 months.
Subscription-subsidy cuts convert demand, don't kill it: labs are already un-bundling (~$14k cost-to-serve on $200 plans was "accidental" pricing); usage migrates to metered billing - hidden demand becomes disclosed revenue.
The break scenario: enterprise ROI patience - Uber burned its 2026 AI coding budget in 4 months ("ROI link is not there yet"); MSFT canceled Claude Code in one division. If that becomes a wave without new agent categories offsetting, demand growth breaks. Watch renewal rates + any hyperscaler CUTTING (not re-mixing) FY28 capex guidance.

3 · Component constraints - when does each un-tighten?

Constraint	Severity now	Un-tighten estimate	Note
Memory (HBM/DRAM)	Extreme - sold out through '26, prices +80-110% QoQ	2028 earliest, contested	Samsung (7/30/26): worsens into 2028. SK Hynix full ramp 2030. Glut requires capex decel FIRST.
Transformers/switchgear	Severe - 2.5-4 yr leads, books full into '28-29	2027-2028	~$2B of Siemens/Eaton/Hitachi US plants have dated online schedules
Packaging (CoWoS)	Severe - NVDA alone absorbs ~all 2026 capacity	2027-2028 (first to ease)	ABF substrate is the deeper bind to late '27
Construction labor	340-500k worker shortfall	No dated resolution	The quiet long-duration constraint

The most capex-deflationary event available: a genuine DRAM/HBM rollover - historically prices fall by multiples, fast, once it turns. That would let hyperscaler capex guides FLATTEN while delivered compute keeps growing - the benign path for GOOGL's margin story (memory ≈ 30-48% of capex by '27).

4 · Siting opposition - quantified

Numerator inflation: the "$130B blocked in Q1 2026" counts delays (~72% historically) at full-decade buildout valuations. Formal government rejections/restrictions: 15 actions in 10 states since 2025, vs ~1,086 facilities planned/under construction - under 2%.
Denominator: $850B+ lease commitments, 331 GW pipeline. Blocked share by any consistent measure: ~3-7%.
Geography: 38 states still offer DC incentives (TX/OH/IN/VA leading); NY's moratorium is a 1-yr pause on 20MW+; Maine's was vetoed. Blocked projects re-site (documented pattern). 23-24 states adopted large-load tariffs - accommodation, not rejection.
Haircut: 5-10% of announced 2026-28 capacity, concentrated in a few jurisdictions. Not the binding constraint - power is.

5 · What this means for the thesis and the trade

Claim	Anchors on regime?	Verdict
310/250 Dec-28 spread breakeven ($287)	Barely	Survives the orderly end of the regime (glut → dead money = spread wins). Killed only by the disorderly end (funding-cycle break at expiry).
Base case $405-475 (backlog converts at margin)	Heavily	Regime holds through window per all four streams. Real risk = delivery slippage: only ~5 GW under construction vs 16 GW scheduled '26; Goldman discounts announced pipelines 40-50%. But slippage defers revenue AND depreciation - a timing risk (~1-2% of revenue/yr), not a loss.
Bull case $500+ (TPU externalization, margin holds)	Heavily + needs Google to out-deliver	Google owns its chip supply chain (TPU/Broadcom) but faces the same grid physics as everyone; ERCOT-heavy siting (TX $40B) is the right side of the map
Crash scenario ($185-235)	Inversely	Requires the demand break (enterprise ROI wave) or funding-cycle crack - NOT supply physics. Supply constraints actually protect margins.

The regime-watch dashboard - six tripwires, in signal order: (1) any hyperscaler cutting FY28 capex guidance (Q4'26/Q1'27 calls); (2) enterprise agentic-tool renewal rates / more Uber-style budget kills; (3) DRAM contract prices rolling over (TrendForce quarterlies); (4) GE Vernova turbine backlog growth decelerating; (5) PJM capacity auction clearing lower; (6) Anthropic's next funding round (the backlog's counterparty). Tripwires 1-2 are demand breaks (bearish GOOGL); 3-5 are supply easing (mixed: margin relief + glut risk); 6 is the concentration risk.

Synthesized 2026-07-30 from four sourced research passes (energy watts model, demand elasticity, component supply chains, siting quantification) - full citations in session research outputs, key figures flagged for confidence inline in those passes. Demand GW projections diverge 2-3x across Goldman/McKinsey/LBNL/EPRI; the model uses ranges. Not licensed financial advice.