Power Grid Gets $967M DOE Boost as AI Data Center Demand Surges

A Familiar Program Gets a Bigger Spotlight

On Tuesday, Oct. 20, the U.S. Department of Energy announced about $967 million in new grid funding. The department said the money would modernize infrastructure, head off outages and lower what Americans pay for electricity.29 Energy Secretary Chris Wright called lower costs and prosperity top priorities for President Trump. He described the grants as sending money directly to states, territories and tribes to improve reliability and reduce bills.25

Much of the coverage presented the money as the administration's answer to surging power demand from AI data centers and the risk of blackouts.21 The program's details tell a narrower story. The money goes out through the Grid Resilience State and Tribal Formula Grants program, which was built to protect infrastructure from storms, wildfires and other disasters.21 It covers fiscal years 2025 and 2026, the last two years of a five-year program created by the 2021 bipartisan infrastructure law. Recipients have until Feb. 17, 2027, to apply.2124

So this is not a new White House initiative built around AI. It is the final installment of a Biden-era formula program, rebranded around affordability. The guidance is also broad: diesel, propane and natural gas backup generators qualify, and so does battery storage.2122

Even the outlet that broke the story with the most favorable framing noted a gap. DOE did not say how much consumers would save or when.21 The same report pointed out that resilience spending can soften the impact of outages without fixing the separate problem of building enough generation to meet demand.21 That distinction matters more than anything else in this story.

The Demand Problem the Money Doesn't Touch

The pressure on the grid is real, and it is concentrated in a few places. PJM Interconnection serves about 67 million people across 13 states and Washington, D.C. It expects roughly 32 gigawatts of new peak demand by 2030, with data centers responsible for about 94% of that growth.21 In September, Pennsylvania regulators warned of a high-demand scenario with little new generation. In that case, parts of PJM could face more than 13 days a year when supply falls short of demand by 2030.22

Demand records are already falling. PJM hit an hourly peak of 162.6 GW on July 2, 2026, and ERCOT in Texas reached 91.1 GW on July 22.39 Texas officials say ERCOT is weighing more than 474 GW of large-load requests, about 90% of them from data centers. That is more than five times the grid's record peak.39 In Virginia, Dominion Energy reported about 53.8 GW of contracted data center capacity as of July.39

Forecasts for how big the AI load will get vary widely, and that variation is a story in itself. Goldman Sachs projects U.S. data center demand rising from 31 GW in 2025 to 41 GW in 2026 and 66 GW in 2027.34 Morgan Stanley's figures for data center IT load run 17.96 GW in 2026, 35.46 GW in 2027 and 52.31 GW in 2028. The bank puts total new U.S. data center power needs over that period at 97 GW, against about 19 GW of available grid capacity.36 Gartner expects AI-optimized servers to use 175 TWh this year, close to double their 2025 consumption.32 One analysis says utility load forecasts filed with FERC run about 40% above credible estimates of what will actually be built.34

The numbers differ, but they all point the same way. Analysts disagree on how large the gap is, not on whether it exists. Some of the projected demand will also never show up. A Relae dataset counted 46 projects in 20 states, worth more than $170 billion, that were blocked, withdrawn or stalled between January 2024 and mid-May 2026. The pace has picked up sharply.34 Gartner expects 40% of AI data centers to face power limits by 2027, and new grid connections in major markets now take 24 to 36 months to approve.31

Bills, Blame and a Partisan Split

On affordability, coverage splits along familiar lines. Conservative outlets tied the grant announcement to rising costs and pointed to strain from data centers in places like Maryland.21 The Natural Resources Defense Council puts more of the blame on the administration. It says average household rates have climbed 15.1% since January 2025, from 15.9 to 18.3 cents per kilowatt-hour. The group cites increases of 25% in Ohio and a third in Iowa.27 NRDC lists data centers, transmission buildout, volatile gas prices and grid hardening as causes. It also argues that federal moves against wind, solar and storage, along with tariffs on grid equipment, have made the shortage worse.27

The NRDC numbers come from an advocacy group, but the trend they describe matches other reporting. PJM's independent market monitor links about 12 GW of added data center demand in 2026/27 to a doubling of capacity costs in the region.32 Policy is moving too. Federal regulators rejected parts of a PJM plan that would have let it force some large customers to cut consumption in emergencies. They approved a registry of major power users instead.21

The administration's bigger grid bet is upgrading the lines that already exist. In September, DOE committed $1.9 billion to 31 "Speed to Power" projects in 26 states. Including the recipients' own contributions, the total is $5.25 billion, which the department says will add more than 23 GW of capacity.28 Like the newer grants, that announcement does not promise any specific bill reduction. Rates are still set by state utility commissions.28 Smaller awards follow the same approach, such as a $4.1 million grant to Entergy Mississippi for dynamic line rating on 375 miles of transmission.25 This is cheap capacity, and probably the smartest federal spending in the mix. But it cannot add generation that doesn't exist.

Nuclear Becomes the Hyperscaler Workaround

While federal money goes into wires and backup generators, tech companies are paying directly for firm power. Google signed a 20-year agreement with Constellation for 890 MW of new nuclear output from uprates at 11 reactors across six PJM sites in Illinois, Pennsylvania and New Jersey. The deal is backed by more than $4.3 billion in Constellation investment.7 The first uprate is expected in 2028, with the full capacity in place by about 2032.3 A separate 15-year deal covers another 2,700 MW across PJM.7 Constellation called the agreement a direct response to PJM's "Bring Your Own Power" proposal, consistent with the White House's ratepayer-protection pledge.7

That came a week after Amazon's 20-year, 690 MW agreement tied to Maryland's Calvert Cliffs plant, which includes about 190 MW of new capacity.67 Oracle, building the $15 billion Lighthouse campus in Wisconsin with OpenAI and Vantage, plans to buy 125 to 250 MW from the Point Beach plant.10 DOE is also offering a $4.2 billion loan for Vistra uprates that would supply Meta.6 In Kansas, Deep Fission signed a 25-year energy purchase agreement for a commercial project at its DOE pilot site.25

The trend is clear: uprates and restarts, not new reactors, because new plants take a decade or more to permit and build.9 Existing sites already have grid connections, cooling systems and operating licenses.8 Still, nuclear does not make the cost question go away. Wisconsin's Point Beach shows why. Its cost per megawatt-hour has risen from $45.94 in 2016 to $75.51 in 2026, and the plant's contract drives about 20% of a proposed $176 million We Energies rate increase.410 Oracle says its purchase could save other customers $300 million. A consumer advocate said that without the full filing, it's impossible to tell whether customers are getting a big win or only small steps forward.10

Cooling Is Quietly Reshaping the Load

The least-noticed part of the story is inside the data hall. Air cooling stops making economic sense at around 40 to 50 kW per rack, and current accelerators draw 700 to 1,200 watts each.13 A full GB200 NVL72 rack uses roughly 120 to 140 kW.32 Morgan Stanley estimates a Vera Rubin rack at about 234 kW and a Rubin Ultra rack at around 600 kW.36 IDC expects 90% of large performance-intensive deployments to use direct liquid cooling by 2027.13

Nvidia says its Rubin platform's 45°C liquid loops can cut cooling-tower water use from about 2.6 million gallons per megawatt per year to nearly zero in favorable climates.16 Microsoft describes its newest AI sites as using closed loops that consume almost no water.20 Industry studies credit full liquid cooling with roughly 10% lower total facility energy use compared with air.15 There is a trade-off, though. Dry cooling can raise electricity use, and waterless designs can push water use and emissions back to the power plant.1320 Better cooling cuts overhead, but it does not offset rack power that rises with every chip generation.

The Reading

Read together, the coverage shows two separate grid strategies. The federal one is spread thin and focused on resilience: about $1 billion in end-of-program formula grants, plus line upgrades that free up capacity. The private one is concentrated and focused on capacity: hyperscalers paying billions for nuclear uprates and dedicated supply. The second will do far more to determine whether data centers get powered without pushing costs onto households.

The cheapest relief may be flexibility. A TU Wien–led study found that the U.S. data center sites it modeled could stay within existing grid capacity by cutting consumption for no more than 35 hours a year.37 The Google–Constellation deal includes load-shaping and demand-response provisions.7 If regulators make that kind of flexibility standard, the grid problem gets easier to manage. Until then, the $967 million is mostly a way to keep the lights on during storms, and not much help with bills.