Utility demand-response events run in 2026
73
Record, versus 49 events in 2025, per Oracle's Nov. 11, 2026 announcement
Peak demand reduction delivered
1 GW
Year to date 2026, summed across Opower utility programs
PJM 2028/2029 capacity price
$325/MW-day
Cleared at the FERC-approved cap, Base Residual Auction held Jul 14, 2026
PJM capacity shortfall
6,831 MW
Cleared capacity versus reliability requirement, 2028/2029 delivery year
PJM backstop auction ceiling
$555/MW-day
Maximum willingness to pay for new supply, amended proposal Oct 29, 2026
Oracle quarterly capital expenditures
$28.5B
Q1 FY2027, quarter ended Aug 31, 2026, versus $8.5 billion a year earlier
Oracle has just shown that a gigawatt of peak demand can be pulled out of 4.3 million households with software and a text message — no new plant, no interconnection queue, no fuel. The catch sits on the other side of the meter: PJM's auction to close a 6,831 MW capacity gap admits only new supply, so the region is preparing to pay up to $555 a megawatt-day for megawatts it could reduce for cents.
The record, decomposed
Oracle's Gigawatt Is 73 Draws on One Base, Not One Dispatchable Block
The headline number is a stack of small wins. Oracle reports 73 events in 2026 at an average peak saving of 16,592 kW per event — about the peak demand of 4,150 homes — and calls the total 1 gigawatt. Multiply73 events by that average and the sum is closer to 1.2 GW of event peaks, slightly above the figure Oracle published, which tells you the gigawatt is an accumulation of separate hours rather than a block of power you could call on at 5 p.m. on a July afternoon.
| Opower program metric | 2025 | 2026 | Change |
|---|---|---|---|
| Utility demand-response events | 49 | 73 | +49% |
| Customers engaged | 3.3 million | 4.3 million | +30% |
| Average peak savings per event | not disclosed | 16,592 kW | not disclosed |
| Peak reduction reported | not disclosed | 1 GW year to date | record |
Depth, not breadth, is the binding constraint. The average event moved the power of about 4,150 homes, a slice roughly 0.1% of the 4.3 million customers Oracle says the programs engaged, so the next gigawatt costs more customer attention rather than more customers. Puget Sound Energy's Flex Events program has delivered up to 16.6 MW per event across three successive seasons for a $0.50/kWh bill credit; CPS Energy's behavioral program has pulled more than 350,000 customers into its virtual power plant and reached nearly 19 MW across 12 August events. Both are retail utility programs rather than PJM capacity resources, so the megawatt is worth avoided purchases rather than capacity revenue.
The price of a megawatt
There Are Three Prices for the Same Megawatt, and One of Them Is Almost Free
| Route to a megawatt | Price tag | Who pays | Evidence |
|---|---|---|---|
| Existing supply in PJM's capacity market | $325/MW-day, or $118,625 per MW-year | PJM load: ratepayers plus large loads | 2028/2029 Base Residual Auction |
| New supply in PJM's backstop auction | up to $555/MW-day, contracts up to 15 years | Large Loads, per PJM's cost allocation | Amended Reliability Backstop Procurement |
| Behavioral demand response | $0.50/kWh bill credit, about $500 per MW-hour curtailed | Utility shareholders, spread over retail customers | Puget Sound Energy Flex Events |
| Build it yourself | $28.5B of capex in one quarter; $13.3B of power commitments | Oracle's shareholders and lenders | Oracle Q1 FY2027 10-Q and FY2026 10-K |
Normalize for the hours that matter and the gap is stark. Firm capacity at $325/MW-day costs $118,625 per megawatt-year: about $13.50 per megawatt-hour spread across all 8,760 hours, or roughly $1,186 per megawatt-hour if you count only the 100 highest hours of the year. A $0.50/kWh bill credit costs about $500 per megawatt-hour of load actually curtailed. The credit is not the whole cost — Oracle does not say what utilities pay for Opower — and it only exists on event days, but even at ten times that estimate behaviorally flexed load would cost hundreds of times less per peak hour than capacity payments alone.
PJM capacity prices have cleared at the cap in three straight auctions
PJM's RTO capacity clearing price by delivery year, in dollars per megawatt-day. The 2028/2029 price eased 2.5% from $333.44 while cleared capacity still fell 6,831 MW short of the reliability requirement.
Unit: $ per MW-day (UCAP)
2025/2026
269.9
2026/2027
329.2
2027/2028
333.4
2028/2029
325
The settlement problem
PJM's Fix for the Shortfall Excludes the Cheapest Megawatt
This is not a forecasting problem; it is a settlement problem. PJM's 2028/2029 Base Residual Auction cleared 138,318 MW of unforced capacity and came 6,831 MW short of its reliability requirement, with the reserve margin at 14.7%. The price cleared at the FERC-approved cap of $325/MW-day — down 2.5% from the $333.44 paid for 2027/2028 — and demand response accounted for 5% of the cleared supply mix. Total cleared capacity rose only 3,733 MW while PJM's peak load forecast for the new delivery year came in about 2,000 MW above the prior auction's forecast.
- The backstop auction pays up to $555 per megawatt-day on a megawatt-weighted average, for contracts of up to 15 years.
- Only new supply qualifies: existing retail demand response cannot bid into the auction built to fix the gap.
- New resources must reach commercial operation no later than June 1, 2032, pushing the supply decision past the current AI build cycle.
- The bid window opens March 1, 2027, closes March 28, and results are posted May 10, 2027, with FERC's suspension of the original order running to Feb. 28, 2027.
- PJM allocates the cost of new capacity to the Large Loads that drove the shortfall, so the load that caused the gap funds the megawatts that fix it.
None of that is neutral for the software layer. Oracle's gigawatt comes from retail customers who already have meters, thermostats and a bill — the cheapest source of peak flexibility available, and the one resource that cannot enter the auction designed to solve the shortage. In the near term that is good for utilities, because it lowers what they must buy. In the long term it caps what generators can charge for being available, since a megawatt that can be switched off by text message is not the same commodity as a megawatt that takes five years to build.
Supply chain
Software Goes In, Turbines Come Out: Who Actually Gets Paid
| Layer | Entity | Evidence of linkage | Where the money lands |
|---|---|---|---|
| Software dispatch (upstream of utilities) | Oracle | 73 events, 1 GW of peak reduction, 4.3 million customers engaged | Utility operating budgets, not capacity revenue |
| Behind-the-meter storage (upstream) | Vertiv | EnergyCore Grid BESS running through CPower's virtual power plant platform; 1 MW Ohio microgrid earning demand response revenue | Equipment sales plus aggregation fees |
| On-site generation (upstream) | Bloom Energy | Up to 2.8 GW committed with Oracle, 1.2 GW deploying, a system delivered in 55 days | Fuel-cell sales that bypass the grid queue |
| Central equipment (upstream of the auction) | GE Vernova | Gas backlog grew from 100 GW to 116 GW in the June 2026 quarter, at least 125 GW targeted by year-end | The 2029-2031 slot book the backstop auction feeds |
| Contracted supply (downstream) | Constellation Energy and Vistra | 890 MW of uprated nuclear under a 20-year PPA with more than $4.3B invested; 20-year PPAs supporting three PJM nuclear plants | 20-year contracted revenue at a scarcity price |
| Retail utilities (downstream) | Puget Sound Energy and CPS Energy | 16.6 MW per event across three seasons; 350,000 customers in a virtual power plant and nearly 19 MW across 12 August events | Avoided peak purchases, if regulators allow |
Read it as a routing map. The gigawatt Oracle produced is priced as a customer service; the megawatts PJM needs are priced at $325 to $555 a megawatt-day, and that price is set by the equipment queue. GE Vernova's gas backlog has gone from 100 GW to 116 GW in a single quarter, while Constellation is spending more than $4.3 billion to uprate 11 nuclear units by 890 MW. Behind the meter, the money moves to whoever delivers power without waiting for the grid: fuel cells that ship in 55 days, and storage paired with an aggregator that knows how to sell its output into the capacity market.
The load is now a grid resource
Washington Already Treats AI Load as Dispatchable Power
On July 2, 2026, PJM forecast a peak of 166,241 MW, above the 165,563 MW record set in 2006, and activated pre-emergency demand response into the evening peak. The grid operator also obtained a Department of Energy emergency order allowing transmission owners to curtail data centers and other large loads that have backup generation, as a last resort before voltage reductions or load shedding. That is the federal government formally treating AI load as a controllable resource, with the cost of flexibility borne by the data center rather than by a ratepayer.
Oracle Is Buying Power Two Ways: Capital, and Software
Cash used for capital expenditures rose from $8.5 billion in the quarter ended Aug. 31, 2025 to $28.5 billion in the quarter ended Aug. 31, 2026.
Unit: $ billions
Q1 FY2026
8.5
Q1 FY2027
28.5
Oracle's own answer to the same constraint is capital, not curtailment. Capital expenditures ran $28.5 billion in the September 2026 quarter against $8.5 billion a year earlier, and the company closed the quarter with $37.1 billion of cash and securities against $664 billion of remaining performance obligations. Its balance sheet also carries $13.3 billion of unconditional purchase obligations, primarily for data center power arrangements, and it has committed to up to 2.8 GW of Bloom Energy fuel cells with 1.2 GW already deploying. For an AI campus that cannot shed a quarter of its load for an hour without missing customer commitments, that is the correct answer — but it is an answer to a different question than the one utilities face.
Horizons
What Moves First, and What Decides the Rest
- Feb. 28, 2027: FERC's five-month suspension of the original backstop order expires and the amended terms take effect.
- March 1, 2027: the backstop bid window opens, closes March 28, and results land May 10 — the moment to test whether new supply stays the only eligible resource.
- July 2027: the next Base Residual Auction clears for 2029/2030, with a peak load forecast that rose more than 5,400 MW for 2026/2027 and about 2,000 MW again for 2028/2029.
- Through 2027: GE Vernova's gas output is scheduled to step from 20 GW in the third quarter of 2026 to 24 GW in 2028 and 30 GW planned for 2030 — the physical capacity the backstop auction bids against.
- 2028: the first uprated unit from Constellation's 890 MW nuclear program is scheduled to be delivered, alongside a 2,700 MW supply agreement.
- One to three years: if data-center curtailment becomes standard in interconnection agreements and power contracts, dispatchable AI load turns into a priced product and the software behind it earns revenue outside a utility operating budget.
- One to three years: if new-supply-only eligibility survives, the marginal megawatt keeps coming from turbines, uprates and fuel cells, and behavioral software stays a cost-avoider rather than a capacity earner.
- One to three years: the downside case for the software layer is behavioral decay — repeated events wear out customer response, and no platform manufactures participation households have lost.
- One to three years: the risk to the new-build case is the mirror image — if load growth slows, the gap is filled by demand response that was never going to be paid for, and turbine slots booked to 2031 look stranded.
Investable takeaways
- Opower's 1 GW across 73 events proves a utility software franchise exactly when grid capacity is scarce; renewals follow the 6,831 MW PJM shortfall, not the software demo.
- No utilities line exists in Oracle's reporting: Opower sits inside $17.2 billion of cloud and software revenue for the September 2026 quarter.
- Near term the stock trades on capex, not Opower: $28.5 billion spent in one quarter against $664 billion of remaining performance obligations.
- 890 MW of uprated nuclear under a 20-year hyperscaler PPA plus a 2,700 MW supply agreement converts cap-clearing scarcity into 20-year contracted revenue.
- A capacity market clearing at $325/MW-day and an auction willing to pay $555/MW-day support pricing for existing PJM nuclear.
- Longer term, every software-flexed megawatt is a megawatt of capacity not built, a structural cap on scarcity value.
- 20-year power purchase agreements supporting three PJM nuclear plants lock the reliability gap into contracted revenue through the AI build cycle.
- New-supply-only eligibility in the backstop auction favors Vistra's existing nuclear fleet over aggregators that cannot bid new supply.
- Horizon: near-term capacity revenue follows the $325/MW-day cap through 2028/2029; longer term the uprates answer data-center load growth.
- Vertiv's EnergyCore Grid BESS running through CPower's virtual power plant platform turns installed backup power into priced grid capacity, starting with a 1 MW Ohio microgrid.
- PJM's July 2, 2026 emergency order allowing curtailment of data centers with backup generation makes every UPS and storage megawatt a grid asset.
- Horizon: near-term sales come from build-outs; grid-services revenue is a 2027-and-beyond option.
- Gas equipment backlog grew from 100 GW to 116 GW in the June 2026 quarter, targeting at least 125 GW by year-end, with new-supply-only rules feeding 2029-2031 slots.
- Near term the ramp from 20 GW to 24 GW of annual output by 2028 is the physical constraint on how fast PJM can close the gap.
- Bear case: cleared capacity rose just 3,733 MW in the 2028/2029 auction while demand response took 5% of the mix, so software-flexed load caps turbine demand after 2030.
- Oracle's commitment of up to 2.8 GW of fuel cells, with 1.2 GW already deploying, prices the AI build's impatience rather than the grid's willingness.
- A fully operational system delivered to Oracle in 55 days, ahead of a 90-day plan, is the product: speed to power, not price per megawatt.
- Longer term, a $0.50/kWh bill credit caps what utilities will pay for behind-the-meter generation.
