PJM at $329/MW-day — How the US Power Grid Becomes the AI Bottleneck Capex Can't Solve
An 11.4× increase in capacity-clearing price in two years is not a forecast. It is a cleared market price reflecting physical scarcity that capex cannot resolve inside the thesis window.
Published May 22, 2026 · CrossVol Research
The Number That Changes Everything
The PJM Interconnection capacity auction for the 2026/2027 delivery year cleared at $329.17 per megawatt-day. The 2024/2025 delivery year cleared at $28.92 per megawatt-day. An 11.4× increase in two years.
This is not a model output. It is the clearing price of a regulated auction where every load-serving entity in PJM territory must procure capacity for the delivery year. The auction reveals two facts simultaneously:
- Marginal generation capacity in PJM territory is structurally short relative to forecast load.
- The market is willing to pay roughly an order of magnitude more for it than it was 24 months ago.
Goldman Sachs' April 2026 note "US Data Center Power Demand Projected to Double by 2027" treats power capacity as a soft constraint that capex can solve. The PJM auction is the empirical disproof. You cannot capex your way to capacity that doesn't physically exist on the timeframe the AI infrastructure buildout requires.
The 300 GW Gap — Decomposed
Multiple independent estimates (EPRI, Lawrence Berkeley National Lab, Gartner via Brookings, internal hyperscaler planning leaks) converge on a US incremental power demand of 200–350 GW above current baseline by 2030, dominated by datacenter load and electrification. We use 300 GW as the central case.
To put 300 GW in perspective: that is roughly equivalent to the entire installed capacity of Germany, plus France's nuclear fleet, plus the UK's total grid combined. The US is being asked to build that incremental capacity by 2030 — five years.
Decomposing the gap by source-type capacity to be added:
| Source | Realistic 2026–2030 Add (GW) | Structural Constraint |
|---|---|---|
| Natural Gas Combined Cycle | ~25 | Siemens / Mitsubishi / GE turbine order book 4-year lead time; pipeline permitting; interconnection queue |
| Solar (utility-scale) | ~45 (firm equivalent ~12) | Land + transmission + interconnection queue (5–7 year cycle); capacity factor 22–27% |
| Wind | ~15 (firm equivalent ~5) | Onshore siting + offshore permitting paralyzed; capacity factor 30–40% |
| Battery Storage (effective firm) | ~10 firm-equivalent | 4-hour storage at scale; recycles renewable not net new |
| Nuclear (existing restart + life-extension) | ~3–5 (Three Mile Island, Palisades, etc.) | Re-licensing cycle ~3 years; PPA negotiation |
| Small Modular Reactors (SMR) | ~0–1 by 2030 (first commercial deployment 2029 at earliest) | NRC certification, vendor delivery, fuel supply chain |
| Large new nuclear (AP1000) | ~0 (no new project clears commissioning by 2030) | Vogtle-style ~8–10 year build cycle |
Summing the realistic firm-equivalent capacity additions: approximately 60–70 GW by 2030. Against a 300 GW demand gap. The arithmetic is uncomfortable.
This is not a forecast about a policy or political environment. It is a forecast about physical infrastructure delivery. Turbine order books are full through 2029. Solar interconnection queues in PJM, MISO and ERCOT are 5–7 years deep. Nuclear is a multi-decade asset class incompatible with the AI thesis window.
Why Executive Orders Don't Close This Gap
A frequent counter-argument is that the Trump administration's deregulatory approach — fast-tracked permits, federal-land siting, suspension of environmental review — closes the gap. We disagree with the diagnosis. The binding constraints are not regulatory:
- Turbine manufacturing capacity. Three vendors (Siemens, Mitsubishi, GE) manufacture utility-scale gas turbines globally. Their order books are full to 2029. An executive order does not add turbine manufacturing lines.
- Transmission interconnection physics. Connecting a new generation source to the grid requires line studies, transformer ordering, substation construction. Transformer lead times are now 3–4 years globally. Executive orders do not produce transformers faster.
- Skilled labor. Linemen, transformer technicians, gas turbine commissioners — the trade unions are at full capacity. Executive orders do not train workers faster than apprenticeship programs.
- Capital allocation. Utilities are regulated entities with cost-of-capital constraints. They cannot deploy capex at 3× their historic rate without rate-base adjustments that go through state PUC processes. Executive orders cannot bypass state utility commissions.
The constraint is physical and capital, not regulatory. The thesis-relevant time horizon (Q1 2027) is approximately 8 months from this writing. Nothing the federal government does in that window changes the realized supply of US electricity.
The China Counter-Position
While the US power-grid story is constraint, the China power-grid story is overcapacity. China added ~370 GW of net new generation capacity in 2025 alone — roughly equivalent to the entire US peak demand. The mix is heavily weighted toward solar (160 GW), wind (75 GW), thermal (80 GW), nuclear (5 GW), and hydro/biomass (50 GW).
More importantly, China's national grid (State Grid Corporation of China) operates with structural over-build incentives: PRC industrial policy treats electricity as a strategic input with policy-priced supply. The result is that industrial electricity prices in China are approximately 60% of US industrial pricing, with no grid-constraint markup.
For AI inference workloads — which are continuously power-consuming, geographically flexible, and primarily sensitive to electricity cost — the China grid is a structural advantage of a magnitude the US cannot match within the thesis window. This is what makes the Huawei Atlas / Ascend platform a credible substitute at the inference layer: not just chip cost parity, but total operating-cost parity inclusive of power input.
Cross-Asset Implications
The grid bottleneck transmits into several discrete asset classes:
1. Datacenter REITs — Asymmetric Long
Datacenter REITs (Equinix, Digital Realty, GDS) with existing power-supplied capacity are positioned as a scarce-asset class. The thesis is bearish on AI-infrastructure equipment (NVDA, MU, VRT) because the asset providers compress; it is structurally bullish on the small number of physical real estate owners that already have powered capacity contracted.
The trade is not a clean expression of the thesis — datacenter REIT valuations are heavily levered to interest rates, and a recession-tilt re-rating would compress them despite the operational story. But on a long-term horizon, the scarcity premium is real.
2. Independent Power Producers — Asymmetric Long
Vistra, NRG Energy, Constellation Energy own existing generation assets that benefit from the capacity-price spike directly. Constellation in particular is the operator of restartable nuclear capacity (Three Mile Island restart contracted to Microsoft). The cleaner expression is buying the IPPs directly versus hedging through the utility sector.
However, IPP equity has already moved meaningfully (Vistra is +200% from 2024 lows, Constellation +250%). The expression is more in long-dated call options than spot equity.
3. Natural Gas Producers — Indirect Long
If utility-scale gas is the marginal source of incremental firm capacity, natural gas demand from the power sector grows structurally. EQT, Range Resources, Antero Resources are positioned as long-dated calls on the Henry Hub forward curve. Long-dated WTI / Henry Hub spread divergence is also a viable expression.
4. AI Infrastructure Equipment — Structural Short (via the Pillar)
The same logic that makes datacenter capacity scarce makes the AI capex deployment slower. If hyperscalers cannot get powered capacity to install GPUs into, GPU demand pulls into 2028–2029 rather than landing in 2026–2027. This removes a forcing function from the NVDA / MU / AVGO bull case in the thesis window. The grid bottleneck is mechanically consistent with the pillar's vol re-rating call on the equipment cohort.
Daily Monitoring — Grid-Specific Signals
Six data series we track:
- PJM capacity auction results — annual, but the most direct read on structural scarcity.
- MISO and ERCOT capacity prices — proxy series for cross-RTO comparison.
- Turbine vendor order book disclosures — Siemens Energy quarterly investor updates explicitly disclose AI-driven order intake.
- FERC interconnection queue reports — quarterly reports detailing the project pipeline depth.
- Henry Hub forward curve, 2027–2029 strip — leading indicator of forward demand assumptions.
- Datacenter REIT funds-from-operations growth and powered-capacity additions — quarterly earnings.
Conclusion
The PJM capacity auction at $329/MW-day is not a forecast. It is a clearing price — the market's actual price for marginal capacity given physical scarcity. The 300 GW US demand gap by 2030 is decomposable into supply additions of approximately 60–70 GW realistic by 2030. The arithmetic is uncomfortable, and the gap is not closeable by executive order because the binding constraints are physical infrastructure, capital allocation, and skilled labor — not regulation.
For the AI thesis, the grid bottleneck operates as a negative forcing function on capex deployment within the window. Hyperscalers cannot install GPUs faster than they can get powered datacenter capacity, and the capacity is structurally short. This is Vector 3 of the pillar thesis, and it is the most under-modeled of the five vectors because the sell-side treats power as a soft constraint that capex solves.
For the cross-asset expression, the right side of the trade is generation, the wrong side is equipment. Datacenter REITs, IPPs, and natural gas producers are structurally long. AI infrastructure equipment is structurally short via the pillar trade construction.
For the full thesis: The China AI Disruption Thesis. For the financial-architecture mechanism: The Hyperscaler Bond Wall. For the cost-side compression on AI revenue: DeepSeek's $0.27 Token Economics.
Disclaimer: This document is for informational purposes only and does not constitute investment advice, an offer, or a solicitation. CrossVol Research does not make trade recommendations. The opinions expressed are those of the authors at the time of writing and may change without notice. CrossVol Research and its principals may hold positions, directly or indirectly, in entities mentioned herein. Past performance is not indicative of future returns. Communication promotionnelle non-MIFID dans l'Union européenne.