On December 19, the Michigan Public Service Commission voted 3-0 to conditionally approve DTE’s 19-year, 1.4 GW, power supply agreement for a Stargate data center campus to be developed in Saline Township, Michigan. This brought to a close a tumultuous year for the relationship between artificial intelligence data centers, state governments and utilities.
The saga isn’t over yet; there are still environmental permits to be obtained, and the Michigan legislature may roll back the state’s tax incentives for data centers. But the arc of this particular Stargate data center’s story reflects nearly every aspect of this market’s development in 2025 — and the key questions that remain to be answered in 2026.
The era of gigawatt-scale data centers
A year ago, on President Trump’s first full day in office, executives from OpenAI, Oracle, and SoftBank announced a $500 billion joint infrastructure project, to finance and build the digital and energy infrastructure to develop artificial general intelligence. AGI, according to SoftBank CEO Masayoshi Son, which would one day “solve the issues that mankind would never ever have thought that we could solve.” Stargate, it was dubbed. Trump stood beaming alongside them, making clear that the federal government would play an increasingly powerful role in supporting AI in the U.S.
Already, on day one, Trump had signed an executive order rescinding President Biden’s executive order on the “Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence.” A steady drumbeat of orders and policy directives would follow, all embracing three parallel ideas: 1) the U.S. is in a race with China for global AI leadership, and so this is a strategic priority; 2) AI is in its infancy so shouldn’t be burdened by regulations; and 3) the scale and consistency of power required is best suited for fossil fuel and nuclear “baseload” power plants.
Stargate’s first data center in Texas was already under construction when the joint venture made its big White House announcement, but OpenAI made it clear there would be more — and that they would all likely be over a gigawatt each.
The concept of the gigawatt-scale data center pre-dated Trump; Biden had signed an EO just before he left office that directed federal agencies to lease land to host gigawatt-scale data centers and new clean power facilities. Throughout 2025, gigawatt-scale data centers were proposed on an unprecedented scale, and many have already secured financing and permits. EpochAi has identified five data centers that will be operating in 2026 at a gigawatt or greater, including those from Amazon, Microsoft, xAI, Meta, and OpenAI’s Stargate campus in Abilene, Texas.
For the energy market, how these and future data centers will be powered is of critical importance. The push for AI has already rewritten the forecasts for gas, geothermal, nuclear, solar, wind, fuel cells, and battery storage. And at the same time, it has forced a rethinking of energy and capacity market designs, electricity ratemaking, and the tradeoffs of building large-scale generation off-grid as developers rush to power their new facilities.
The Stargate initiative spent much of late 2024 and early 2025 evaluating sites beyond the initial campus in Abilene, Texas. And they were largely welcomed with open arms. States generally see data center development as economic development, particularly as Biden-era infrastructure spending and manufacturing support gets unwound.
In 2024, Michigan passed two important data center sales and use tax exemptions that went into effect in early 2025, while the state’s Strategic Fund and Economic Development Corporation promoted the state’s established construction and tech labor pool. And by October 2025, the basics of the deal for the Stargate campus in Saline Township, Michigan, were ready to be announced. Unlike many other planned gigawatt data centers, the one in Saline Township would rely entirely on power from the grid, which DTE Energy had agreed to supply.
In other Stargate sites, the plan is often to leverage onsite behind-the-meter gas generation, employing modular gas reciprocating engines from VoltaGrid, or GE Vernova’s stackable aeroderivative gas turbines, for example. A fact sheet that was shared from Texas Congressman Jodey Arrington’s office includes information on the use of onsite gas generation at the Doña Ana County, New Mexico, site and the Shackelford County, Texas site.
By the year’s end, there would be six Stargate developments underway in the U.S., including campuses in Texas, New Mexico, Wyoming, Wisconsin, and Michigan. Around the world, OpenAI is planning Stargate sites in Canada, Argentina, the UK, Germany, Norway, Korea, and India.
Utilities seize the opportunity
Throughout 2025, electric utilities found themselves in the unfamiliar but welcome position of having to update both load and revenue forecasts. Wall Street couldn’t have been happier. What was for years a rather boring source of stability and dividends became a hot growth sector.
As we reported in August last year, analysts were calling this the “golden age of utilities,” and quarterly earnings were marked by upward revisions in customer pipelines and capital expenditure plans. The investor-owned utility trade group Edison Electric Institute forecast capex of $1.1 trillion between 2025 and 2029, and identified 488 GW of planned or proposed generation capacity through 2029. The uncertainty around these figures, however, led to fears of widespread rate increases on the backs of what remain largely unproven business models in the AI era.
Data center operators, in the best case, represent a class of industrial customers that are well capitalized and eager negotiators as they look to build quickly, at gigawatt scale. Forecasts range widely, but there is relative consensus of around 100 GW of data center demand in the U.S. between 2025 and 2035, representing 2.35% annual load growth, far outpacing the last decade’s 0.6% annual figure.
Utilities have a strong hand to play. Utilities and grid operators can be slow to get projects interconnected, but where capacity exists — as was the case with DTE — they remain the preferred partner for data centers. Being grid connected means reliability, resilience, and operational familiarity. (The alternative is for operators to build their own generation and storage onsite and behind the meter, which means adding further levels of redundancy and power management and significantly increasing the complexity of a project.)
In Michigan, DTE Energy was the model of a large utility’s fortune in 2025. In Q3, its five-year capex plan was raised by $6.5 billion, or 22%, which it ascribed to data centers, and included nearly $2 billion of incremental energy storage. The 1.4 GW Stargate deal would increase its load by an astonishing 25%, and help it continue to exceed its earnings targets, with further upside from over 3 GW of additional data center load in its pipeline.
Meanwhile, it planned to update its IRP to include more renewables, storage and gas investment, all in service of a wave of data center deals.
The contract specifics
As the year went on, DTE sought fast-track approval for the Stargate data center electricity contracts. The utility claimed that Green Chile Ventures, the subsidiary of Oracle taking the contract, had an aggressive timeline that needed to be met, and pushed for approval at the December 5 meeting of the state’s PSC. But for residents in the area there was a catch: The ‘ex parte’ motion meant DTE wouldn’t have to go through a hearing, where the public or opposing groups could submit expert testimony and challenge DTE’s assumptions.
DTE, like many other utilities negotiating power contracts with data center operators, had to balance their economic interest in this deal with two critical factors: 1) counter-party risk (i.e. whether the Stargate consortium was as credit-worthy as a Google or Meta), and 2) ratepayer risk (i.e. whether the added cost of serving Stargate would be socialized across its customer base).
To address these concerns, the approved contracts for the Saline Township data center included elements that are now becoming common across the U.S.:
- Extended contract terms: The DTE-Stargate contract is for 19 years, compared to what is typically a five-year contract for large energy customers.
- Take-or-pay terms: These require the customer to pay 80% minimum billing demand; between 50% to 60% is more typical.
- Curtailment: This is often lumped in with discussions of flexibility, but in this case it refers to the requirement that in the event of an energy emergency “the data center’s load is reduced or interrupted before interrupting service to other DTE Electric customers.” How the customer manages that curtailment obligation is up to them; they could choose to just rely on traditional backup power with diesel gensets, or move to more advanced flexible capacity solutions based on onsite batteries, generation, or even power management solutions that operate at the compute level.
- Termination payment: Utilities have begun requiring up to 10 years’ worth of minimum billing demand if the data center stops operating before the contract ends.
The Michigan PSC also included approval of a 1.4 GW energy storage facility that DTE would develop, own, and operate and that Green Chile Ventures would finance. This also comes with its own collateral requirements and an early termination payment obligation.
Finally, MSPC added requirements for DTE to update their IRP to account for the size of Stargate’s load, and update its analysis of the impact of the data center on its capacity requirements and energy waste reduction target.
This type of negotiation with utility commissions is happening across the country. The day after the Michigan PSC approved DTE’s request, the Georgia PSC voted to approve Georgia Power’s own request to add 9.8 GW of new power generation assets to their network, largely to power data centers. They included an interesting condition that Georgia Power would “financially backstop” these new projects, which include 3.6 GW of combined cycle natural gas generation, 3 GW of battery energy storage systems, and over 2.8 GW of PPAs. The backstop is meant to protect against the data center operators terminating a contract early and shifting the costs on to ratepayers. Additionally, Georgia Power agreed the incremental revenue from large load customers of at least $556 million annually will benefit customers in its next rate case, in 2028.
With a goal of weeding out speculators and helping speed interconnection, large-load tariffs proliferated across the country as 2025 came to a close. Over 60 have been proposed, and at least 36 already adopted by utilities, according to Enverus Intelligence Research. Much like in Michigan, the principles of these tariffs focus on defining cost allocation, ratepayer protections, customer eligibility criteria, and grid impacts.
Financing looked easy, until it didn’t
AI data centers are wildly expensive to build: Estimates range from $35 billion to $60 billion per gigawatt, largely down to the costs of high-density GPU racks, advanced cooling systems, power infrastructure upgrades and redundancy, and the complexity of networking and interconnection.
In 2024, the industry shrugged off concerns of an infrastructure bubble by claiming the largest tech companies had enough free cashflow to finance these data centers straight from their balance sheets. In 2025, however, that narrative was constantly challenged.
First, the largest tech players looked to corporate bonds. Alphabet, Meta, Oracle, Amazon and neo-cloud hyperscalers like CoreWeave and Crusoe issued over $120 billion in corporate debt in 2025, nearly five times their recent annual averages. Next, they looked to participate in infrastructure partnerships, such as the one between Microsoft, Blackrock, and Global Infrastructure Partners and industry partners called the AI Infrastructure Partnership, where they could keep these dollars off their corporate balance sheets.
More creative financing followed. Hyperscalers pursued data center development via the use of special purpose vehicles, where the sponsoring AI company cedes control, again to keep the project debt off their balance sheet. They have also embraced asset-backed securitization, where the collateralized assets are increasingly the Nvidia GPUs themselves.
Nvidia’s role in providing the financing to its customers, often referred to as circular financing, was for some a hallmark of a bubble: extraordinary sums of money changing hands among buyers and sellers in a complex web of transactions that often mask the underlying risks that they all share.
Oracle, in the months leading up to finalizing the power deal with DTE, had it the worst of all the major tech players. The corporation has never been as well-capitalized as Google, Meta, Amazon, or Microsoft, and its AI infrastructure move led it to make larger and larger commitments to its partners — more often with debt rather than its own cash. Oracle’s stock price soared on the September news of a record $300 billion cloud contract with OpenAI, only to come falling back down after its third quarter earnings revealed its margins on renting chips disappointed investors who expected this aggressive move to AI would also increase profits.
For the Stargate project in Saline Township, Oracle and OpenAI partnered with the newly formed Related Digital to develop the project. The real estate development company, Related Companies, launched its data center development arm in March, claiming a$45 billion pipeline of 5 GW of data centers across the U.S. and Canada; these included a 1 GW campus in Chicago, Illinois, a 1.2 GW data center in Missouri, and others in Wyoming and Canada.
Saline Township wasn’t mentioned at launch, but in October — after negotiations about rezoning the site — Related announced it would lead development of the Stargate campus in Michigan. But raising the $7 billion to $10 billion needed was not simple, and it’s not yet complete.
Related and Oracle initially turned to Blue Owl Capital, one of the largest alternative asset managers in the U.S. Blue Owl was a major financial partner for the Stargate site in Abilene, Texas, as well as for a number of other hyperscale data centers. Just days before the announcement of the Michigan PSC’s approval of DTE’s power supply contracts, however, the Financial Times reported Blue Owl would not be leading the financing for the project, with sources claiming Blue Owl couldn’t agree on leasing and debt terms with Oracle.
Blackstone was rumored to be the next stop for Related and Oracle, but a final deal remains unannounced.
Still, no one likes data centers
In a sprawling 98-page note to investors to start 2026, Jefferies analysts said the “data center honeymoon is over” — and in fact they have “become the villain.” It pointed to how the AI infrastructure buildout has collided with concerns about energy affordability, which will likely arise as a key wedge issue in the year’s state and federal elections.
“What is striking is how swiftly hyperscalers have pivoted from being the biggest buyers in the market for power (with clear potential to bid utilities, IPPs, and renewable companies against each to the lowest levels) to being price takers,” Jefferies notes. “This pivot over what amounts to a [roughly three-year] period has been remarkable.”
Concern in some communities has morphed into outright opposition. In Saline Township, where locals have called the data center project “uniquely evil,” residents are concerned about air and water pollution, land use, the diversion of resources, and general mistrust of large tech companies.
Related Digital’s first proposal to rezone 575 acres for development was rejected by the planning commission in August 2025. Related followed with a lawsuit, claiming the decision qualified as exclusionary zoning. By October, the township had settled and allowed the project to move forward. The PSC decision approving DTE’s energy contracts faced opposition as well, and though the contracts have now been conditionally approved, there are still more local permitting hurdles for Related to clear.
Data center resistance is increasing across the US, and remarkably bipartisan. This is arguably a healthy response to what can feel like opaque and unbalanced processes to develop large-scale infrastructure, and it can be a truly animating issue for voters.
This organized opposition can pose real delays — almost equal to those caused by projects languishing in interconnection queues, or else waiting on generation, transformers, or switchgear. In early December a coalition of mostly environmental groups called on Congress to impose a moratorium on the development of new data centers in the U.S. And similar groups are making an anti-data center push in Europe and Latin America as well.
What comes next
As we enter 2026, the AI infrastructure market appears to be rationalizing. Regulators have moved surprisingly quickly to explore and adopt new tariffs and rules around interconnection to put some downward pressure on speculation. In the discourse around a potential bubble, investors and lenders are acting (perhaps?) more cautiously, and have increasing power to set terms. And the AI labs continue to release more powerful and capable updates to their models, causing the market to expand beyond chatbots, to AI agents and physical AI platforms like autonomous vehicles and robotics.
For Oracle and the Stargate consortium, though, the outlook is less clear. OpenAI has the largest market share among the AI labs, but Oracle continues to struggle to find its footing in a space dominated by much larger rivals. Where it made its name as a successful enterprise software company with little capex required, it is now attempting to rival the largest hyperscalers in capex — without their free cashflow and solid balance sheets.
It hasn’t been easy for SoftBank either. As the founding financial partner of Stargate, it barely made its year-end cash commitment to the group. Of the $40 billion it had committed to Stargate in early 2025, a $22.5 billion tranche was due in December, but the money wasn’t there. It quickly sold major stakes in Nvidia and T-Mobile, and Bloomberg reported that it drew from a margin loan secured by shares of the chip architecture company Arm Holdings to complete the deal. For what could have been one of the most important investments in SoftBank’s history, it had the whiff of desperation.
For the AI infrastructure market, 2026 comes with more questions than obvious answers.
- Are the companies behind the Stargate project actually the best combination to compete with the likes of Google, Meta, and Amazon — not to mention China? OpenAI may be in the lead for Chatbot adoption, but that lead feels tenuous as rivals like Google leverage their platforms to spread usage. As a relative newcomer to cloud infrastructure, Oracle has more to reckon with than just a poor debt-to-equity ratio, and is exposed to the risks of the “circular” financing it’s in with Nvidia and others. Meanwhile, SoftBank’s struggles to meet its obligations to the group in 2025 may be even harder in 2026.
- Will the projections that nearly a third of all new data center development will deploy behind-the meter gas generation come to fruition? Are data center operators actually prepared to become power plant operators as well? The Stargate campus in Michigan was fortunate that DTE Energy offered 100% grid-connected power and battery storage, but many other Stargate projects are planning for onsite BTM generation. 2026 will likely be a year when these plans meet the realities of implementation.
- Will the gigawatt era of data centers result in a redesign of power markets in the U.S.? It has already led to important new rulemaking processes to address interconnection bottlenecks.
- What form will flexibility take as a result of the industry’s speed-to-power imperative? Ideally, we’re entering a world where large-scale data centers emerge as flexible grid assets, not just gigantic load centers, but the mechanics of flexibility are complex and the AI companies have a strong financial incentive to operate 24/7.
- Will community resistance lead to better development outcomes for residents, ratepayers, and data center operators — or just forestall the inevitable proliferation of the only infrastructure being financed and built in this country?
- And finally, can this market weather a bubble forming and bursting without socializing too much of the damage? Surely pipelines will be rationalized and scaled back, load forecasts trimmed, and projects canceled in 2026. But will this lead to a collapse in financing?
A version of this story was published in the AI-Energy Nexus newsletter on January 7, 2026. Subscribe to get pieces like this — plus expert analysis, original reporting, and curated resources — in your inbox every Wednesday.
