Why power belongs on ComputeTape
Power is one of the main constraints on AI compute supply. GPUs only become usable cloud capacity when data centers have electricity, cooling, interconnection, and backup power.
Power Watch
Power Watch: learn the grid constraints behind AI compute
Track the grid and power constraints shaping AI compute capacity.
AI data centers do not just need GPUs — they need available power, interconnection approvals, substations, transmission, and cooling capacity. Power Watch tracks the regional signals that may limit cloud capacity, delay data-center buildouts, or increase compute costs.
70 / 100Power Score
Power constraint is concentrated in ERCOT, PJM and Southeast, where data-center load growth is outpacing grid planning.
ElevatedStatus
Trend: Tightening
Regions
Regional Heat Map
Power-constraint scores (0–100; higher means more constrained) for AI data-center and GPU cloud capacity, as of Jun 21, 2026. Scored from public grid-operator and utility data using the Regional power constraint scoring method.
| Region | Status | Score | Trend | Confidence | Main Constraint | AI Compute Impact | Source |
|---|---|---|---|---|---|---|---|
| PJM | Severe | 85 / 100 | Tightening | High | Capacity costs at record: 2026/2027 cleared at the price cap and the 2027/2028 auction hit a new ~$16.4B record while falling short of the 20% reserve margin. Data centers drive nearly all of a forecast 32 GW peak-load rise to 2030. | Large AI campuses face record capacity costs and multi-year energization timelines; PJM accelerated a backstop reliability auction and FERC large-load interconnection reform is pending but unresolved. | PJM Inside Lines |
| ERCOT | Severe | 90 / 100 | Tightening | High | Large-load queue above 233 GW (about 70%+ data centers) against roughly 85 GW peak; in early June 2026 ERCOT moved to a 'Batch Zero' process that triages large loads against what the transmission system can support today. | Fastest-growing buildout region, but Batch Zero deliverability screening and tight reserve margins mean projects without grid readiness face rejection or long delays. | ERCOT — Large Load Integration |
| Southeast | Elevated | 65 / 100 | Tightening | Medium | Utilities planning 10+ GW of new capacity for data centers (Georgia Power ~10 GW; Southern 50 GW pipeline; Dominion 47 GW contracted), though independent analysts call the forecasts aggressive. | Heavy utility commitments and new gas-plus-storage, but contested load forecasts and rate/tariff decisions add cost and approval risk. | Georgia PSC — Data Center Fact Sheet |
| MISO | Moderate | 55 / 100 | Easing | High | 2026/27 capacity prices eased (summer $424/MW-day, down from a $666 record); adequacy met across all zones but flagged at risk as large loads accelerate. | More near-term headroom than PJM or ERCOT, but interconnection timelines lengthen as large-load requests grow. | Utility Dive |
| CAISO | Moderate | 50 / 100 | Stable | High | Modest data-center forecast (+1.8 GW by 2030); summer 2026 assessed well-positioned with ~8 GW of new resources, but large-load cost allocation and 'pulsating' loads remain unresolved. | Lower AI-campus pressure than PJM or ERCOT; high prices and interconnection-policy uncertainty are the watch items. | California Energy Commission |
| SPP | Watch | 45 / 100 | Tightening | Medium | Peak load projected to nearly double (56 to 105 GW over 10 years); 26+ GW of >100 MW load requests (9 GW data centers), served by new but curtailable large-load tariffs (HILL/CHILL). | Current headroom and a fast 90-day interconnection path, but curtailable terms and future growth temper long-term firmness. | SPP — High Impact Large Load Integration |
| Pacific Northwest | Moderate | 50 / 100 | Stable | Medium | Data-center load could reach ~4 GW by 2030; BPA and the NW Power Council find the central scenario adequate but the high scenario short, with peak-driven needs by 2031. | Attractive low-cost hydro power, but transmission availability and hydro variability cap how fast new load can energize. | BPA — 2026 Loads & Resources Study (White Book) |
| National | Elevated | 70 / 100 | Tightening | Medium | AI data-center load growth is outpacing grid planning in the highest-demand regions (PJM, ERCOT, Southeast). | Power remains a binding medium-term constraint on GPU cloud capacity, concentrated in the Eastern Interconnection. | Load-weighted composite |
Deals
Power Deal Tracker
Large power agreements, generation deals, and utility arrangements tied to AI data-center capacity.
| Buyer / Operator | Power Type | Region | Size | Status | Expected Timing | Impact | Confidence |
|---|---|---|---|---|---|---|---|
| Hyperscaler / AI cloud | Nuclear / PPA | Not disclosed | Not disclosed | Announced | Not disclosed | Long-term capacity support | Medium |
| Data-center operator | Grid interconnection | PJM | Not disclosed | Pending | Not disclosed | Energization risk | Medium |
| AI infrastructure provider | Gas / behind-the-meter | ERCOT | Not disclosed | Proposed | Not disclosed | Potential capacity unlock | Medium |
| Utility / data-center customer | Tariff / rate case | Southeast | Not disclosed | Watch | Not disclosed | Cost and approval signal | Medium |
Queue
Data-Center Power Queue
Major AI data-center projects and their power-access risk.
| Project / Area | Operator | Region | Estimated MW | Utility / Grid | Status | Power Risk | Timeline | Confidence |
|---|---|---|---|---|---|---|---|---|
| Northern Virginia cluster | Multiple | PJM | Not disclosed | Dominion / PJM | Active | High | Ongoing | Medium |
| Texas AI campus pipeline | Multiple | ERCOT | Not disclosed | ERCOT utilities | Expanding | Medium / High | Ongoing | Medium |
| Southeast data-center growth | Multiple | Southeast | Not disclosed | Regional utilities | Rising | Medium | Ongoing | Medium |
| Pacific Northwest AI sites | Multiple | PNW | Not disclosed | Regional utilities | Watch | Low / Medium | Ongoing | Medium |
Context
Why Power Matters
GPU availability alone does not translate into usable AI cloud capacity. Large AI clusters require interconnection, transmission capacity, cooling, backup power, and affordable energy. Power constraints can delay data-center timelines, raise operating costs, and limit where new AI compute capacity can come online.
Signals
Top Power Signals
Jun 30, 2026
PJM Capacity Auction for 2028/2029 Delivery Year Opens
Signal: PJM Capacity Auction for 2028/2029 Delivery Year Opens. Impact: Bears on compute cost and GPU availability for AI data centers through power, grid, or interconnection constraints. Affected market area: Capacity timing.
Jun 29, 2026
PJM opposes waiver for $2B gas-fired plant in fast-track interconnection review
Signal: Gas turbine backlogs prompted Advanced Power Services to seek changes to its project in PJM's fast-track Reliability Resource Initiative interconnection review, but the grid… Impact: Bears on compute cost and GPU availability for AI data centers through power, grid, or interconnection constraints. Affected market area: Availability and Capacity timing.
Jun 9, 2026
FirstEnergy asks FERC to require data centers to pay for transmission interconnection costs
Signal: FirstEnergy's proposal adopts a cost allocation practice from the gas pipeline sector. It comes ahead of the Federal Energy Regulatory Commission's expected large load… Impact: Bears on compute cost and GPU availability for AI data centers through power, grid, or interconnection constraints. Affected market area: Cost and Capacity timing.
Jun 30, 2026
Nebius and Bloom Energy partner to power AI infrastructure build-out May 20, 2026
Signal: Nebius and Bloom Energy partner to power AI infrastructure build-out May 20, 2026. Impact: Bears on compute cost and GPU availability for AI data centers through power, grid, or interconnection constraints. Affected market area: Capacity timing.
May 27, 2026
25MW AI data center to be built in McMinnville, Tennessee
Signal: Project to be completed in Q1 2028 Impact: Bears on compute cost and GPU availability for AI data centers through power, grid, or interconnection constraints. Affected market area: Availability and capacity timing.