Datacenter Power Constraints in 2026: The Real Bottleneck on AI

The constraint on AI infrastructure in 2026 is not GPUs. Nvidia is shipping. The hyperscalers have allocations. CoreWeave, Lambda, and the GPU-cloud cohort have capacity. The constraint is power — interconnection queues at the major US transmission operators are measured in years, large-load tariffs are being renegotiated under pressure, and the 2030 demand projections from the EIA and DOE assume a doubling of US data center load that has no clean answer for where the electrons come from.

This is the operational picture for anyone planning data center capacity through 2026 and 2027 — what is actually constrained, where the workarounds are, and how the math has changed.

ERCOT Texas and the curtailment conversation#

ERCOT — the Electric Reliability Council of Texas, the grid operator for most of the state — has been the most exposed US grid to the AI data center buildout. Texas combines cheap land, a permissive permitting regime, and the strongest renewables deployment in the country, which made it the default destination for data center growth from 2022 onward. By 2025 ERCOT was projecting 90+ GW of new large-load interconnection requests through 2030, against a 2024 system peak of roughly 86 GW.

The 2025 ERCOT board adopted curtailment provisions for large flexible loads — data centers above a threshold can be required to curtail consumption during peak hours in exchange for faster interconnection. The trade-off is meaningful. AI training workloads can be curtailed in principle (checkpoint and pause); AI inference workloads less so. The hyperscalers have negotiated mixed-use agreements that protect inference and accept curtailment on training.

The Texas PUC’s 2025 large-load tariff proceedings continued through 2026 with the central question being who pays for the transmission upgrades — incumbent ratepayers or new data center loads. The political answer is converging on the new loads paying more, which raises the effective cost of Texas data center capacity and partially closes the gap with other states.

AEP, Dominion, and the PJM interconnection queue#

PJM — the Mid-Atlantic and Midwest transmission operator — has the most-cited interconnection queue problem. As of mid-2025 the PJM queue had hundreds of GW of generation pending interconnection studies, with average study-to-energization timelines of 4 to 6 years. The reform package PJM adopted in 2023 reduced the queue somewhat but did not solve the underlying problem: the speed of new generation interconnection is materially slower than the speed of new load interconnection.

The PJM operating territory includes Northern Virginia (Dominion Energy) and the AEP service territory in Ohio, Indiana, and West Virginia. Northern Virginia is the largest US data center cluster — roughly 25 percent of US data center capacity by some measures — and is the region where the constraint has been most visible. Dominion’s 2024 and 2025 large-load tariff filings, the 2025 Virginia SCC proceedings on data center cost allocation, and the political conversation about whether residential ratepayers are subsidizing data center expansion have all been operational issues for capacity planners.

AEP has been on the supply side of the same conversation. The AEP-Indiana Michigan Power agreements with major data center customers include cost-sharing for transmission upgrades; the AEP Ohio rate cases have argued for similar structures.

The Three Mile Island restart and the Constellation pivot#

The Microsoft-Constellation Energy agreement announced September 2024 — to restart Three Mile Island Unit 1 and dedicate its output to Microsoft data centers under a 20-year PPA — was the single most-cited transaction in the AI-power conversation. The unit had been shut down in 2019 for economic reasons; the restart targets 2028 with regulatory approval from the NRC required.

The structure became a template. Constellation Energy has positioned itself as the nuclear-PPA-for-AI counterparty of choice; subsequent agreements with Meta and Amazon followed similar shapes. The premium price the hyperscalers are paying for 24/7 carbon-free power — materially above grid-average — reflects both the sustainability commitments and the practical reality that firm power is what data centers need.

The broader nuclear pivot includes Vistra’s announcement of plans for additional uprates at its Comanche Peak units, NRG’s small-modular-reactor commitments, and the operational restart conversations around Palisades in Michigan (Holtec, with Department of Energy loan support) and Duane Arnold in Iowa. The economics of restart depend heavily on PPA pricing — at 2024 grid-average prices these would not pencil; at 2025-2026 AI-data-center PPA prices they do.

The AWS-Talen 960 MW deal and the behind-the-meter pattern#

AWS’s March 2024 acquisition of the Cumulus data center campus from Talen Energy — co-located with the Susquehanna nuclear plant in Pennsylvania, with up to 960 MW of behind-the-meter power — was the deal that established behind-the-meter co-location as the workaround for grid-interconnection delays. The structure: build the data center on the same site as the generator, take power directly behind the meter, avoid the transmission queue entirely.

FERC’s November 2024 order rejecting the proposed ISA amendment for the AWS-Talen arrangement complicated the model. FERC’s concern was that behind-the-meter data center loads at grid-connected generators created cost-shifting questions for the broader system. The follow-on regulatory uncertainty pushed several similar deals into pause through 2025.

By mid-2026 the FERC posture had clarified somewhat. Behind-the-meter arrangements at nuclear sites with co-located dedicated generation are largely workable; arrangements at gas-fired sites with grid backup are more complicated. The hyperscalers have continued to negotiate behind-the-meter deals on the workable side of the line.

Sumitomo and the gas-turbine waitlist#

The most operationally surprising constraint of 2025 was the gas-turbine waitlist. GE Vernova, Siemens Energy, and Mitsubishi Power — the three major heavy-duty gas-turbine OEMs — have order books extending into 2029 and 2030 for the H-class and J-class machines used in major combined-cycle plants. The Sumitomo Heavy Industries waitlist became a quoted reference point in 2025 utility filings.

The implication is that even if a developer can clear interconnection and permitting on a gas-fired generator, the actual physical equipment is on a multi-year wait. The 2025 surge in announced gas-fired data center generation projects ran straight into the OEM constraint. Aero-derivative turbines (smaller, faster delivery) became more attractive but at higher operating costs and shorter expected lifetimes.

The 2030 forecast that everyone references#

The EIA’s 2025 Annual Energy Outlook projected US electricity demand growth from data centers at roughly 6 to 8 percent of total US load by 2030 — up from approximately 4 percent in 2024. The Lawrence Berkeley National Laboratory data center report, the DOE Loan Programs Office projections, and the major utility 5-year integrated resource plans all converge on similar magnitudes. The doubling, in absolute terms, of US data center load between 2024 and 2030 is the operational planning assumption.

The geographic distribution of that growth is uneven. Northern Virginia, Texas, Oregon, Iowa, and the upper Midwest are the largest existing clusters. The new builds skew to Oklahoma, the Mississippi Delta states, and the South Atlantic — places with available land, friendly permitting, and the ability to negotiate large-load tariffs. The international map is shifting in parallel; Mumbai, Hyderabad, Singapore, Dublin, and Frankfurt are the international counterparts.

What capacity planning looks like in 2026#

For organizations planning major AI compute commitments, the practical sequence has reordered. The 2022 question was “can we get the GPUs.” The 2026 question is “can we get the GPUs and where can we put them and when does the power come on.” The lead time for greenfield data center capacity from site selection to operational has stretched from 24 to 30 months in 2022 to 36 to 48 months in 2026, with the long pole almost always grid interconnection.

The workarounds enterprises and hyperscalers are using:

Co-location with existing nuclear or large hydro, at premium PPA pricing.

Behind-the-meter arrangements at sites with dedicated generation.

Phased build-out with initial capacity at constrained interconnection levels and additional capacity contingent on transmission upgrades.

Demand-side flexibility — training workloads designed to curtail, deferred-batch inference, geographic shifting of non-latency-sensitive workloads.

International expansion to jurisdictions with more grid capacity (the Saudi, UAE, and Singapore programs are explicitly competing on this).

Where pdpspectra fits#

We help enterprises plan AI infrastructure that respects the actual power and capacity constraints — the workload-shifting architecture, the multi-region inference patterns, and the realistic capacity planning that the 2026 grid environment requires. Our enterprise practice does this work.

Related reading: Snowflake vs Databricks vs BigQuery 2026, sovereign AI initiatives 2026, and tech IPO market 2025-2026.

---

Power is the binding constraint on AI infrastructure in 2026. Talk to our team about a capacity plan that accounts for the realities of grid interconnection, behind-the-meter arrangements, and workload-shifting flexibility.