—·
The next AI infrastructure deal is less about chips than about who finances substations, secures grid access, and absorbs the risk of 24/7 power demand.
Georgia’s regulators approved a generation buildout costing $16.3 billion in December 2025 after concluding that power demand from data centers was large enough to justify a roughly 50% increase in Georgia Power’s capacity; utility staff estimated customers could ultimately pay $50 billion to $60 billion over time once financing costs and regulated returns are included (apnews.com). That is the clearest sign yet that “AI infrastructure partnership” no longer means only cloud credits, GPU roadmaps, or model-training alliances. It increasingly means a harder bargain: who pays for gas turbines, transmission upgrades, backup capacity, and queue priority when hyperscalers want electricity on a timetable the grid was never built to meet.
That shift makes this phase of AI expansion distinct from the alliance-heavy coverage that has focused on chipmakers and frontier-model developers. The more consequential negotiation is moving downstream, into utility dockets, rate design, power purchase agreements, interconnection rules, and local permitting fights. In other words, the defining partnership in AI infrastructure is becoming a four-way contract among hyperscalers, utilities, regulators, and host communities.
The economics explain why. Lawrence Berkeley National Laboratory estimated that U.S. data centers used 176 terawatt-hours of electricity in 2023, up from 58 TWh in 2014, and projected a rise to 325–580 TWh by 2028, equivalent to 6.7% to 12% of total U.S. electricity consumption (eta.lbl.gov, newscenter.lbl.gov). EPRI separately estimated that data centers could consume 4.6% to 9.1% of U.S. electricity generation by 2030, while a later EPRI-Epoch AI analysis put total U.S. AI power capacity at about 5 GW today and potentially above 50 GW by 2030 (publicpower.org, publicpower.org). Once the load is measured in tens of gigawatts, the question is no longer whether AI needs infrastructure partners. It is which partners will underwrite the grid.
For years, hyperscalers could treat power procurement as a procurement function layered onto real estate strategy: sign renewable contracts, secure utility service, add diesel backup, and scale campuses where fiber and tax incentives already existed. AI changes that logic because the load is larger, more concentrated, and more urgent. A conventional cloud region can ramp in phases. A GPU-heavy AI campus often arrives with a much steeper demand curve and a stronger need for 24/7 firmness, high-voltage upgrades, and advanced cooling systems that increase the site’s dependence on reliable power infrastructure.
That is why the language of partnership has shifted from “sourcing clean energy” to “sharing system costs.” Utilities are now being asked to build for loads that may exceed the demand of midsize cities, but regulators are under pressure to prevent residential and small-business customers from subsidizing speculative or unusually concentrated growth. The bargain is simple to describe and hard to implement: hyperscalers want speed and certainty; utilities want long-duration revenue commitments; regulators want ratepayer protection; communities want jobs and tax base without inheriting outsized infrastructure or environmental costs.
Virginia offers one of the clearest illustrations. In November 2025, the Virginia State Corporation Commission approved a new GS-5 rate class for customers demanding 25 megawatts or more, explicitly targeting the biggest electricity users, including data centers, so they would pay rates based on their unique cost profile rather than blend into broader classes (scc.virginia.gov, scc.virginia.gov). Dominion had argued that special terms were needed to protect other customers from stranded-cost risk if large-load users scaled back after the utility had already built generation, transmission, and distribution assets to serve them (news.dominionenergy.com).
This matters because Virginia is not a fringe market. It is the core U.S. data-center market, and regulators there are effectively writing an early template for the politics of AI load growth. Once utilities begin creating bespoke rate classes, minimum-demand charges, long contract terms, and exit penalties for hyperscale users, AI infrastructure partnerships stop looking like strategic alliances and start looking like project finance.
The utility side of the bargain is becoming more explicit: if AI customers want accelerated buildouts, they must absorb more of the financial risk. Georgia has moved furthest in making that principle visible. The Georgia Public Service Commission approved a base-rate freeze through the end of 2028 in July 2025 even as opponents warned that data-center-related costs could still surface later (apnews.com). Then, in December 2025, regulators approved Georgia Power’s much larger expansion plan, with company and commission officials emphasizing that large energy users would pay more and that the revenue could place “downward pressure” on residential bills after the freeze ends (apnews.com).
Georgia’s own Public Service Commission data center fact sheet underscores how quickly the scale changed. It notes that an estimate of data-center demand had reached 6,600 MW by 2023, forcing regulators to confront whether existing tariff structures were adequate for a class of customers whose size and load shape differ radically from ordinary commercial users (psc.ga.gov). Georgia Power’s 2025 Integrated Resource Plan was framed around the need to meet the energy requirements of a fast-growing state economy, but the underlying policy issue was more specific: can a regulated utility build ahead of massive AI-related load without socializing too much risk onto captive customers (georgiapower.com)?
Texas shows the problem in an even starker way because the queue itself has become a signal of speculative and real demand colliding. ERCOT’s May 2025 operational overview said the grid operator was tracking about 156,000 MW of large loads seeking interconnection as of June 2, 2025 (ercot.com). By October 2025, ERCOT said it was tracking roughly 205 GW of large load interconnection requests (ercot.com). A National Association of State Energy Officials presentation citing Texas regulatory proceedings said nearly 69% of 189 GW of large-load requests were from data centers (naseo.org).
Those numbers do not mean all requested loads will materialize. But they do reveal why utilities and grid operators are changing their posture. The grid is being asked to plan for AI expansion under conditions of uncertainty, where the cost of underbuilding is lost investment and the cost of overbuilding can be passed through regulated bills. Partnership, in this context, increasingly means getting hyperscalers to sign terms robust enough that utilities can finance the build with political cover.
The first case is Amazon Web Services and Talen Energy in Pennsylvania. In November 2024, FERC rejected a proposal tied to the Susquehanna nuclear plant that would have allowed an Amazon data center to expand the amount of co-located power drawn behind the meter from 300 MW to 480 MW, questioning whether PJM’s tariff provided fair and transparent treatment for similar customers (ferc.gov, powermag.com). In February 2025, FERC launched a broader review of co-location arrangements for large loads, and in December 2025 it directed PJM to create clearer rules for such deals (ferc.gov, ferc.gov). The important point is not simply that a deal was challenged. It is that direct power access for AI campuses has become a regulatory category of its own.
The second case is Amazon’s broader Pennsylvania expansion. In June 2025, Amazon said it would spend $20 billion on two data-center complexes in the state, including one adjacent to Susquehanna and another at Keystone Trade Center, where power would come through the grid rather than a direct plant tie-up (apnews.com). That split is revealing. Hyperscalers are not choosing between grid service and dedicated power partnerships; they are pursuing both, depending on what regulators, grid operators, and local infrastructure can support.
The third case is Microsoft and Constellation at Three Mile Island. In September 2024, Constellation announced plans to restart Unit 1 at Three Mile Island, targeting 2028 operations under a 20-year agreement with Microsoft to buy the plant’s output for data-center power matching (cnbc.com, apnews.com). AP later reported that the U.S. Department of Energy offered a $1 billion loan in November 2025 to help finance the restart (apnews.com). This is not the old renewable-PPA model. It is an AI-era structure in which a hyperscaler’s long-term demand helps justify the revival of a large generation asset, backed by federal financing.
The fourth case is Meta and Constellation in Illinois. In June 2025, the companies signed a 20-year power purchase agreement for the Clinton Clean Energy Center, with Constellation saying the deal covered 1,121 MW of emissions-free nuclear energy beginning in 2027 (constellationenergy.com, cnbc.com). AP reported that the agreement would help preserve 1,100 local jobs, support $13.5 million in annual tax revenue, and add 30 MW of expanded clean-energy output at the site (apnews.com). Again, the pattern is clear: AI infrastructure partnerships are being used not merely to source electrons, but to finance asset life extension, licensing confidence, and local political legitimacy.
A fifth, more recent case is Google and Xcel Energy in Minnesota. In February 2026, Google announced a data center in Pine Island, Minnesota, and said the agreement with Xcel would bring 1,400 MW of wind, 200 MW of solar, and 300 MW of iron-air battery storage from Form Energy onto Xcel’s grid; Xcel summarized the package as 1,900 MW of new clean energy tied to the project (blog.google, newsroom.xcelenergy.com). The significance lies in the structure: the partnership is not just a bilateral energy claim. It is a utility-integrated package combining data-center siting, grid additions, and storage deployment.
As these cases accumulate, regulatory institutions are being pushed into a new role. They are no longer just refereeing rates after infrastructure is built. They are being asked to define the terms of entry for AI load itself. FERC’s intervention in PJM is the clearest federal example. The commission said in February 2025 that it was reviewing whether PJM’s tariff needed rules for the co-location of large loads such as AI-enabled data centers, specifically to preserve reliability and ensure fair cost treatment (ferc.gov). By December 2025, FERC had ordered PJM to establish transparent rules for servicing AI-driven data centers co-located with generation (ferc.gov).
PJM’s own forecasting shows why the issue escalated. PJM said its 2025 long-term forecast and subsequent evaluations pointed to up to roughly 30 GW of projected electricity-demand growth from data centers between 2025 and 2030 (insidelines.pjm.com). Its updated 2026 long-term load outlook said the current forecasted annual growth rate was more than ten times the 0.3% ten-year average annual growth rate projected in the 2021 forecast, a sign of how dramatically large-load expectations have shifted regional planning assumptions (insidelines.pjm.com).
That regulatory pivot changes the commercial logic for hyperscalers. Securing land and fiber is no longer enough if substations, transmission capacity, or generation rights are trapped in contested queues. The real scarcity is increasingly “firm deliverability with political permission.” That is why partnerships are extending into rate settlements, utility-specific tariffs, and behind-the-meter proposals. Grid access is becoming a negotiated asset class.
The final change is political. Host communities are no longer mere beneficiaries of construction jobs and tax abatements; they are part of the financing equation because local acceptance affects timelines, cost recovery, and permitting risk. Meta’s Clinton agreement was sold not only as a clean-energy arrangement but as a deal preserving employment and tax revenue in central Illinois (apnews.com). Google’s Minnesota announcement similarly foregrounded not just compute capacity but a grid buildout with wind, solar, and long-duration storage additions that can be presented as broader system value rather than private benefit alone (blog.google, newsroom.xcelenergy.com).
That messaging is not cosmetic. It reflects the reality that communities and regulators increasingly want proof that hyperscaler projects deliver system benefits commensurate with the infrastructure they require. If the public is being asked to tolerate new transmission lines, power plants, water use, or land-use change, then the old promise of abstract digital growth is not enough. Partnerships must show tangible local returns and credible insulation for ordinary ratepayers.
The likely result is that AI data-center expansion will bifurcate. Some projects will proceed in regions where utilities can still socialize part of the system build under conventional growth assumptions. But the most power-intensive campuses will increasingly need dedicated structures: special tariffs, minimum-bill commitments, utility-owned transmission expansions, nuclear or gas-backed PPAs, long-duration storage packages, and explicit cost-sharing terms. In practical terms, the next generation of AI infrastructure partnerships will look less like Silicon Valley alliance press releases and more like hybrid utility-regulatory financing compacts.
By 2027 to 2028, the center of gravity in AI infrastructure is likely to move decisively toward power-backed siting strategies rather than pure compute-led expansion. That timeline is already visible in the start dates of major arrangements: Meta’s Clinton agreement begins in 2027 (constellationenergy.com), while Microsoft’s Three Mile Island-linked supply target is 2028 (cnbc.com). If Berkeley Lab’s 2028 upper-bound projection of 580 TWh for U.S. data-center electricity use proves directionally correct, the industry will not be able to rely on ad hoc procurement and queue gaming alone (eta.lbl.gov).
The policy recommendation is straightforward: state regulators and grid operators should require large AI data-center customers above a clear demand threshold—such as Virginia’s 25 MW standard—to sign long-duration service agreements with minimum-demand charges, transparent exit provisions, and project-specific cost allocation for major transmission and distribution upgrades (scc.virginia.gov). FERC should also push regional grid operators to standardize co-location and large-load interconnection rules so that access to power cannot depend on bespoke negotiations alone (ferc.gov).
For investors and operators, the implication is equally clear. The premium asset in AI infrastructure over the next three years may not be the newest accelerator or even the lowest-cost land parcel. It may be the project that can prove a bankable answer to three questions at once: who supplies the power, who pays for the grid, and who bears the downside if demand forecasts disappoint. The firms that solve those questions first will shape where AI actually gets built.
Electricity-market rules shape AI data-center timelines by determining who pays for grid upgrades and reliability, and when.
The DOE’s PORTS Technology Campus model turns hyperscaler–chipmaker deals into sovereign, power-contracting arrangements with grid risk and site governance embedded.
Planned US data centers face power delays tied to grid hardware lead times and interconnection limits, forcing hyperscalers and utilities into new PPA and reliability fights.