POWER — CONSTRAINED
Power is a binding constraint on the AI buildout — in many regions, the tightest one
72/100 conviction
Last reviewed 17 July 2026 · our own structural assessment, not a price target
The thesis
Through 2027, electrical infrastructure is becoming a binding constraint for many AI data-centre projects — grid interconnection, generation and high-voltage transformers — not just compute. Increasingly, hyperscalers route AROUND the constraint (on-site gas, imported transformers, prepaying for queue position) rather than being stopped outright, so it shows up more in cost and detour than outright cancellation — but the gap between sub-18-month build cycles and multi-year equipment and interconnection lead times remains real.
The case for
The constraint is now showing up in cleared market prices, not just commentary. On Tuesday 14 July 2026 PJM Interconnection — the grid operator for 13 states and Washington DC, about 67 million people, running what it calls the world's largest competitive wholesale electricity market — published its 2028/29 Base Residual Auction. The auction does not buy electricity; it buys commitments to be available. It procured 138,317.8 MW UCAP of accredited capacity at $325/MW-day, the FERC-approved cap, for the entire footprint, and still finished 6,831.3 MW UCAP short of its own reliability requirement — against 6,516 MW short in the prior auction. Those are the first two occasions in PJM's history that the entire RTO has fallen short. The decisive detail is PJM's own appendix simulation ('BRA Clearing Simulation Without Cap and Floor', 2028/2029 BRA Report, pp.17-18): re-clearing the same seller offers with the price collar removed lifts the RTO clearing price to $554.72/MW-day (ComEd $776.69) and the cleared MW multiplied by clearing price by $13.3B to $29.7B — while procuring just 76.8 MW UCAP more capacity and moving the total reserve margin from 14.7% to 14.8%. PJM cautions it does not know how sellers would have bid without the collar. On equipment, the same week: generator step-up transformer lead times surpassed 160 weeks by Q1 2026, up from a 143-week average in 2024, and high-voltage circuit breakers went from 77 weeks (2023) to 125 weeks (H2 2025); Wood Mackenzie forecasts transformer prices rising 4-10% over the next year and expects breakers and switchgear to face larger deficits than transformers (Reuters/Wood Mackenzie, 9 Jul 2026).
The case against
Capital is paying up and building through, not stopping. Meta said on 13 July 2026 it is expanding Richland Parish, Louisiana to '5 GW in compute capacity' representing 'an investment of more than $50 billion in the Richland Parish region' — up from the 'more than $10 billion in Louisiana' it announced on 4 December 2024 (both are undefined promotional regional aggregates, not audited capex). Microsoft's CFO Amy Hood told investors on 29 April 2026 she feels 'quite good about our ability to work through the physical limitations', emphasising CPU, GPU and storage deployment rather than again singling out power — though Microsoft still expects to be capacity-constrained through 2026. SemiAnalysis (18 Jun 2026) argues the 'roughly half of 2026 US data-centre capacity cancelled' headlines are overstated: its year-end-2026 North American hyperscaler self-build forecast moved only ~1% over six months and its colocation forecast <5%. And the PJM shortfall is not a simple data-centre story — cleared supply actually ROSE 3,733 MW UCAP (which PJM says came primarily from higher Accredited UCAP factors and fewer unoffered megawatts — not 3.7GW of new plant); the reliability requirement rose faster, alongside accreditation changes, only 317.1 MW UCAP of new generation plus 207.6 MW of uprates, delayed transmission in-service dates, a compressed ~23-month auction schedule, and, by PJM's own admission, the price cap itself, which 'can reduce the amount of investment, and therefore supply, in the PJM region.'
What would change our mind
PJM's September Backstop Procurement — if FERC approves it — procuring most of the 6,831 MW shortfall without a further price escalation; generator step-up transformer and high-voltage circuit-breaker lead times contracting from their Q1-2026 readings (>160 weeks and 125 weeks); or hyperscaler AI capex plans cooling rather than being reiterated or raised.
Where the exposure sits
The theme plays out in grid and high-voltage power-equipment supply chains, on-site/behind-the-meter generation, and electric utilities rather than in the chip layer. Which reliever ultimately dominates — grid build-out, imported equipment, or on-site gas — remains unsettled.
Show our work — the evidence
The key evidence behind the score, dated and sourced — supporting and disconfirming. This is how the conviction is built, and what we re-check as it moves.
- Context$554.72 is not an uncapped price — it is the default VRR curve ceiling. From PJM's published 2028/29 parameters: max(1.15 x Gross CONE $223,800/MW-yr ICAP - 0.75 x Net E&AS $129,886.71, 0.2 x CONE) = $159,954.97/MW-yr ICAP; /365 = $438.23/MW-day ICAP; /0.79 accreditation factor = $554.73/MW-day UCAP. PJM printed $554.72. Removing the collar reverts to the standard ceiling — it does not discover a market price. Both the actual and simulated auctions cleared AT a cap.
- ContextThe price collar covers FOUR auctions (2026/27, 2027/28, 2028/29, 2029/30). Three are complete; the last closes 15 December 2026. Per OATT Attachment DD 5.10(a)(i) the tariff stops referencing the collar from the 2030/31 Delivery Year — there is no auto-renewal, and extending it requires a new FPA 205 filing and FERC acceptance. FERC called the collar 'time-limited'.
- ContextPJM does not attribute the shortfall to data centres. Cleared supply ROSE 3,733 MW UCAP (primarily from higher Accredited UCAP factors and fewer unoffered megawatts, not new plant); the reliability requirement rose faster. PJM lists accreditation changes, only 317.1 MW UCAP of new generation plus 207.6 MW of uprates, delayed transmission in-service dates, a compressed ~23-month auction schedule, and its own price cap, which 'can reduce the amount of investment, and therefore supply, in the PJM region.'
- ContextMeta is expanding Richland Parish, Louisiana to '5 GW in compute capacity', representing 'an investment of more than $50 billion in the Richland Parish region' — up from 'more than $10 billion in Louisiana' announced 4 Dec 2024. Capital is paying up and building through the constraint.
- AgainstSemiAnalysis's year-end-2026 North American hyperscaler self-build forecast moved only ~1% over six months and its colocation forecast <5% — arguing the dramatic "cancellation" figures count speculative announcement-stage projects as committed capacity.
- Against~$1.8–2B of new North American transformer capacity is being added (Hitachi Energy, Siemens Energy), but most comes online 2027–28 — after the thesis window.
- ContextPJM's own appendix simulation ('BRA Clearing Simulation Without Cap and Floor', pp.17-18): re-clearing the same offers without the price collar raises the RTO price to $554.72/MW-day (ComEd $776.69) and cleared value by $13.3B to $29.7B, while procuring only 76.8 MW UCAP more capacity (138,394.6 vs 138,317.8) and moving total reserve margin 14.7% -> 14.8%. PJM caveats: 'PJM does not know how offers may have changed if the temporary cap and floor was not implemented.'
- ContextReal new physical supply in the 2028/29 BRA was 317.1 MW of new generation plus 207.6 MW of uprates — 524.7 MW, ~0.4% of the market, against a 6,831.3 MW shortfall. Coal cleared 2,941 MW lower. Demand response — the fastest, cheapest resource — went BACKWARDS by 277 MW into the highest-priced capacity environment in PJM's history.
- ContextPJM's 2028/29 Base Residual Auction (14 Jul 2026) procured 138,317.8 MW UCAP at $325/MW-day — the FERC-approved cap — and finished 6,831.3 MW UCAP short of the reliability requirement (prior auction: 6,516 MW short); the first two RTO-wide shortfalls in PJM history. PJM plans to seek FERC approval for a special Backstop Procurement in September.
- ContextGenerator step-up transformer lead times surpassed 160 weeks by Q1 2026, up from a 143-week average in 2024; high-voltage circuit breakers rose from 77 weeks (2023) to 125 weeks (H2 2025). Wood Mackenzie forecasts transformer prices rising ~4-10% over the next year and expects breakers and switchgear to face larger deficits than transformers. Data centres were just under 2% of the electrical-equipment market in 2020; WoodMac sees up to 40% under accelerated scenarios.
- ContextThe 2028/29 cap was $325.00/MW-day UCAP. PJM's Net CONE for 2028/29 was $325.69/MW-day UCAP — the ceiling sits $0.69 below the level PJM's own market-design paper says prices must reach 'at or above ... for an extended period in order to attract entry of new supply'. PJM also reports the collar excluded over 800 MW of cost-based offers above the prior $333.44 cap.
Tags mark how each fact is sourced: Company disclosure (the company said it) · Official data (a government or agency figure) · Analyst estimate (a third party's number) · Reported (via the press).