Oracle's dominoes: Will its New Mexico data center be the trigger for AI landmines?

Oracle issued a force majeure notice to the developers of Project Jupiter in New Mexico, stating that the data center would be shut down due to political factors.

Impacted by the incident, Oracle's stock price closed down 6.9%, and the stock prices of major suppliers such as Blue Owl and Bloom Energy also declined. The market interpreted it as a "project going to fail," but this could be the first domino to fall in the AI minefield: if the data center cannot be online by 2028, can Oracle postpone rent payments?

Even more alarming is that New Mexico won't be the only obstacle for data centers; it's the first example to bring structural contradictions to the forefront. Especially with the significant political infiltration of industries before and after the election, data centers will be the most intense battleground. In particular, grid connection delays and the delivery of critical equipment can delay a server room's power supply by three years without any political bans or legislation.

I. What happened? Oracle's dominoes

Bloomberg reports that Oracle (ORCL) has formally notified the developer of Project Jupiter, Blue Owl Capital's STACK Infrastructure, of invoking the force majeure clause. Sources familiar with the matter describe the situation as follows: if Oracle and the developer jointly determine that a force majeure event related to the power supply commitment has occurred, Oracle can seek a rent deferral of up to three years. It does not intend to relinquish its major tenant status, and the rent will not be cancelled; it will continue to be paid for the entire lease term as soon as project payments commence.

In layman's terms, Oracle's "rental" contract stipulates that payment must be made no later than three years after the contract expires.

However, there's an even more crucial implicit message, as revealed in a supplementary report from the Financial Times: even if Oracle fails to complete power connection by the initial deadline, the company still has a contractual obligation to pay investors "holding costs" in lieu of rent, for a maximum period of three years; the initial deadline is set for the first half of 2027. Specifically, "the power supply risk for the data center is borne by Oracle, not the investors." This indicates that in this type of transaction structure, the investor has removed the most uncontrollable risk of US data centers (whether power will arrive on time) from their balance sheet and passed it on to the tenant. Oracle is willing to take on this risk because it has no other choice; in the competition for AI computing power in the US, no power means no capacity, and no capacity means no RPO (Recovery Points on Payment).

Let's look at how Oracle's project was delayed—first, the state right-of-way permit was rejected by the New Mexico Department of Land in March 2026, and the appeal was rejected again on July 14. Energy Transfer was forced to reroute through federal land, adding approximately 130 days to the process. Pipeline commissioning was postponed from the original August 2026 to February 1, 2027, a delay of 185 days. The delivery of the first phase of the project was postponed from the original November 2026 to the first half of 2027, a delay of approximately 210 days. This also explains why Oracle's force majeure notice pertains to "events related to the power supply commitment," rather than "construction delays"—construction delays are the result, while permit delays are the cause.

Project Jupiter, located in Dona Anna County, New Mexico, covers approximately 1,400 acres and was originally planned to house four data centers with a designed capacity of 2.45 GW, equivalent to the simultaneous electricity consumption of about 1.8 million US households, with a total investment of approximately $165 billion. It is one of five US sites under the $400 billion agreement signed last year by OpenAI, Oracle, and SoftBank, and is also one of the flagship campuses of the Stargate project, personally announced by President Trump. It was originally intended to be powered by Bloom Energy's solid oxide fuel cells (SOFCs), but fuel cells require natural gas, which necessitates a 17-mile pipeline, known as the Green Chile Project, connecting to the existing gas network. This is one of the reasons why biochemical oxygen demand (BE) plummeted last night, which will inevitably impact the domestic BE supply chain on Monday.

The domino effect has begun: a 2.45GW data center's fate hinges on a 17-mile pipeline that requires approximately 0.6 miles of right-of-way on public trust land from the State Land Administration. The reasons given for the rejection include greenhouse gas emissions, pressure on water and other resources, and the fact that "the pipeline offers virtually no benefit to the State Land Administration beneficiaries, the local community, or even the entire state of New Mexico."

How the Market Interprets It: Understanding the Wall Street Narrative

Oracle shares fell the most directly, by 6.9%.

Blue Owl fell by about 3.6%, and Bloom Energy fell by 3.1%. The real pressure signals appeared in the credit market, not the stock market: Reuters reported on September 18 that the approximately $18 billion construction loan for Project Jupiter (provided by a syndicate of about 20 banks) was already trading on the secondary market at face value of 89 to 91 cents; on the same day, Bloomberg reported that metrics measuring Oracle's credit risk hit record highs. In other words, six days before Oracle issued its force majeure notice, the lending market had already provided a precise assessment.

Why is Oracle experiencing more pain than other companies during the same delay?

KeyBanc analyst Jackson Ader hit the nail on the head: contract payments follow a fixed schedule, while revenue depends on project progress, creating a cash flow gap between the two; Oracle will feel this mismatch more acutely than hyperscale cloud vendors with more abundant cash reserves.

Looking at its balance sheet, Oracle has $122.3 billion in long-term debt, approximately $236 billion in total debt including current liabilities, $261 billion in data center lease commitments, and a $95 billion cap on its FY2027 capital expenditure guidance. The lease commitments alone exceed its total debt, indicating that Oracle has shifted a significant amount of capital-intensive activities to off-balance-sheet lease structures. Oracle accounts for approximately $117 billion in the Bloomberg U.S. Investment Grade Corporate Bond Index, making it the largest single issuer in the U.S. non-financial sector.

Next, let's look at Oracle's cash flow situation, which is the market's biggest concern regarding the sustainability of AI.

Oracle's latest fiscal quarter saw a free cash flow of approximately -$5.4 billion, and the company plans to raise approximately $40 billion this fiscal year through a combination of debt and equity. The conversion pace of remaining performance obligations (RPOs) is also a key variable: only about 12% of intended RPOs are expected to convert within 12 months, and 34% within three years. This means that two-thirds of Oracle's $664 billion in revenue will only be realized several years later, but the pressure of capital expenditures and interest rate constraints will be felt immediately. If a flagship site is delayed, the loss will not only be in revenue but also in the on-balance-sheet costs already priced into anticipated financing arrangements and market confidence.

The biggest problem and concern with Oracle's AI infrastructure narrative isn't demand, but rather the discount imposed by the "delivery certainty" of US infrastructure capabilities. Evidence on the demand side (RPO growth, OCI growth, 97.9% GPU utilization, GPU resale premium) all point to industry-wide tension rather than oversupply; however, Oracle's capital structure dictates that if delivery is delayed, the losses will immediately manifest as interest and lease liabilities, while revenue will only materialize years later. Giants like Microsoft, Amazon, and Google can absorb delays with operating cash flow, but Oracle cannot.

This is also the real reason why the market has been worried that it will become the first domino in the AI world.

II. Why is it important? The US attitude is changing.

In 2026, the attitude of US state politics towards data centers underwent a fundamental shift, which can be summarized in one sentence: subsidies are still being issued, but every new law passed adds conditions to those subsidies. Simply put, the barriers to entry are getting higher and higher.

At least 14 states do not disclose the cost of subsidies, making it impossible for outsiders to assess the fiscal consequences. In the 2026 legislative session, 25 states proposed bills to cut or eliminate subsidies, more than double the number from the previous year; however, only 8 states actually passed such legislation, and only Maine completely eliminated all subsidies. Texas and Virginia each suffered annual tax revenue losses exceeding $1 billion, while 10 other states experienced annual losses exceeding $100 million.

Categorizing state-level measures reveals three typical paths. The first is cutting funding: Maine completely eliminated data center tax subsidies; Arizona passed an $18.3 billion budget bill suspending new data center tax credits for three years (effective July 1, 2026); Illinois Governor Pritzker suspended new data center tax incentive agreements starting July 1; Ohio Governor DeWine instructed the Tax Credit Administration to suspend accepting new applications; and Utah passed legislation stipulating that no new incentives will be provided to "high-load data centers" starting May 6, 2027. The second is shifting costs: Oklahoma requires data centers with peak demand exceeding 75 megawatts to fully bear utility service costs; Tennessee requires data center owners to bear all costs of associated utility infrastructure; Florida's SB 484 prohibits utilities from shifting data center costs to residential and small business customers; and Maryland's Utility RELIEF Act restructured grid cost-sharing rules. The third approach is to decentralize decision-making power: the laws of Florida and South Dakota explicitly retain the authority of local governments in zoning, permitting, and land use, allowing them to set stricter standards or even veto projects.

The differences between these three paths are crucial. Cutting costs affects the project's economics, which can be negotiated; shifting costs affects the electricity price structure and alters the contract terms between utility companies and data centers; while decentralizing decision-making power directly entrusts the project's fate to the lowest level of government, the one most easily influenced by public opinion. Project Jupiter encountered the third path.

In 2026, at least 14 states attempted to pass some form of data center construction moratorium, but none of the legislation was ultimately implemented. Maine and Vermont went the furthest: both state legislatures passed bills, only to have them vetoed by their governors (Maine Governor Janet Mills expressed support in principle for the moratorium but opposed the bill's lack of exemptions for specific local projects; Vermont Governor Phil Scott believed the state's existing regulations were sufficient), both attempts to overturn the veto failed. New York was the closest exception: a bill imposing a one-year moratorium on new projects exceeding 20 megawatts passed the Senate 44-16 and the House 102-39, with Governor Hochul yet to comment; however, she subsequently established a one-year moratorium for specific large-scale new facilities through an executive order, making New York the first state in the nation to implement a state-level moratorium. The legislative path to moratoriums failed, but the executive path has been successfully implemented.

This also implies a framework for extrapolating the "midterm election effect." If Democrats gain a majority in more state legislatures and gubernatorial seats after November 3, then the proposals that failed in the 2026 legislative session (versions still under discussion in committees in states like Michigan, North Carolina, South Carolina, and Minnesota) will return in a more mature form; conversely, if Republicans retain their state-level majority, the 2026 landscape will likely continue—more cost-shifting provisions and fewer hard suspension orders. New Mexico, where Project Jupiter is located, falls into a more delicate position: State Land Commissioner Stephanie Garcia Richard is running as the running mate for Democratic gubernatorial candidate Deb Haaland in the November election, and one of her key actions during her term was twice rejecting pipeline access rights for the project.

A symmetrical fact that is easily overlooked is that, concurrent with the "tightening," states are still fiercely competing for projects. States with abundant electricity, such as Indiana, Louisiana, Texas, and Ohio, continue to increase their investment in infrastructure, and developers' behavior in 2026 has shifted from "choosing land" to "choosing electricity." This means that the net effect of tightening state policies is not a reduction in the total number of projects, but rather a redistribution of project locations—a redistribution process that itself incurs a 1-2 year time lag. Turning our attention back to New Mexico: it lies in the middle ground on the US map, "neither significantly tightening nor significantly increasing." What truly stalled Project Jupiter was not state legislation, but the land approval power held by an elected commissioner. This is the most universal aspect of this case—it doesn't require any new laws; it only requires an approver to write "no" on a single application.

Electricity: The sole veto power, and it's raising prices.

If policy is a soft constraint, then electricity is a hard constraint. The controversies surrounding AI data centers in the US ultimately converge on the same number: the electricity bill. And when this number begins to affect the monthly expenses of ordinary households, the controversy shifts from an environmental issue to an election issue.

PJM Interconnection, the largest regional grid operator in the United States, serves 67 million people across 13 states and Washington, D.C., in what is also the world's most densely populated data center region. Its capacity auction (BRA) clearing price provides the most direct price signal in this resource competition: $28.92/MW·day for delivery year 2024-25, jumping to $269.92 in 2025-26 (a record increase of approximately 9.3 times), $329.17 in 2026-27, $333.44 in 2027-28, and $325.00 in 2028-29. This represents an increase of approximately 11.5 times in three years, and for three consecutive years, the price was sold at the "price cap"—the 2028-29 cap was triggered by a settlement in a lawsuit filed in Pennsylvania; without the cap, the clearing price would have been $389/MW·day. This is crucial: the cap was held up by political pressure, not by supply easing.

The U.S. top energy regulatory agency has intervened. In June 2026, the Federal Energy Regulatory Commission (FERC) unanimously voted to instruct six regional grid operators to ensure that AI data centers and other large loads can be connected to the transmission system "in a timely and orderly manner." The FERC also required them to explain within 30 days how they would guarantee power generation capacity and submit grid connection rule adjustments within 60 days. These adjustments included co-location arrangements, off-meter generation, and new transmission service categories for large loads with flexible adjustment capabilities. The FERC also stipulated that data centers must bear all grid upgrade costs required for their connection. These administrative requirements have led to a surge in demand for power – for example, small modular reactors (SMRs) in the U.S. take 8-12 years to develop, large transformers take 24-40 months, and gas turbine orders are already booked until 2030.

This led to a rather rare combination in American politics in 2026: candidates from both parties were vying to attack the same industry, which had almost no voter base speaking up for it.

According to UMass Amherst data, the percentages opposing the construction of local AI data centers are as follows: liberals 84%, 2024 Harris voters 81%, Democrats 76%, national average 65%, Republicans and Trump voters 52% each, conservatives 51%, and hardline MAGA supporters 42%. In other words, even among the hardline MAGA group, who are most supportive of Trump and closest to his pro-AI stance, opponents are still 1.7 times more numerous than supporters (25%); and among Republicans, opponents are approximately 2.9 times more numerous than supporters (18%). Raymond La Raja, head of polling at UMass Amherst, stated, "We can't find a single group of voters who truly want these data centers." This statement is the starting point for understanding the political impact of the midterm elections.

Specifically, in electoral politics, the landscape in 2026 presents three characteristics. First, Republican candidates are beginning to distance themselves from the president's pro-data center stance. Ohio Senator Jon Husted, when asked about the president's statement that "boycotting data centers will impoverish communities," publicly responded, "That's not my position." The Republican Senate campaign described the data center issue as a "burden" hanging over Husted's campaign. The House of Representatives passed a bill by an overwhelming majority of 210 Republicans and 206 Democrats to protect consumers from the energy costs of data centers, while Husted's Senate version failed to pass in September. Second, the two parties have formed a rare point of convergence on the specific issue of "cost shifting." The House passed the Ratepayer Protection Act by a vote of 417 to 3, transferring the increased electricity costs driven by data centers from consumers to the technology companies operating these facilities. Third, the president himself has become the focus of controversy rather than the solution to this issue. Trump claimed that communities opposing AI data centers "will eventually fall behind and become impoverished," which contradicts the views of most Americans, further complicating the Republican Party's situation in swing districts.

For data center projects, the "election result" is less important than the "election process." This is because, regardless of who wins, the election process itself alters developers' incentives: to mitigate election risks, most developers will choose to postpone submitting applications for sensitive projects, increase community concessions, and add more liability waivers to contracts in the second half of 2026—Oracle's force majeure notice occurred just six weeks before the election. These contractual changes will not automatically revert after the election.

III. What's next? Will the force majeure continue to escalate?

Oracle's Domino Effect: Ranking of Six Risk Factors

Breaking down the risks of Oracle's AI data centers into six independently assessable factors, it's unsurprising that power supply and licensing rank first. However, counterparty concentration, ranking before local politics, is often underestimated in market discussions. This is due to Oracle's contractual structure with OpenAI: OpenAI is the largest single customer in Oracle's $664 billion RPO. S&P cited "extreme concentration risk" as the primary reason for its downgrade. Oracle's FY2026 financial report shows that it has already reduced operating costs through a plan to lay off 162,000 employees (expected to affect approximately 30,000 jobs, with a total restructuring cost of approximately $2.8 billion in 2026)—treating operating expenses as adjustable items and AI infrastructure capital commitments as fixed items. This order itself illustrates where the pressure will be released first. In contrast, equipment and timelines, and state policy diffusion rank lower because these two have alternative paths: equipment can be paid at a premium to jump the queue, and state policies can be implemented in different states. The truly irreplaceable paths are "electricity" and "customers."

This ranking also provides an actionable tracking clue: of the six factors, any deterioration in the first four will directly impact CDS rather than stock prices. This is because their common consequence is "increased uncertainty regarding payment timing," and the credit market is far more sensitive to payment timing than the stock market.

Data center risks for other US giants: the same storm, different boats

"Oracle's risks" and "industry risks" must be priced separately. The same force majeure notice, if issued by Microsoft, would elicit a much smaller market reaction—not because of business issues, but because of its capital structure and power strategy.

Comparing the seven major AI infrastructure players across two dimensions—"financial and leverage risk" and "power and licensing risk"—clearly reveals three distinct situations. The first category comprises companies with "high risks on both ends"—Oracle (92/88) and Stargate (85/62), led by OpenAI. Oracle simultaneously bears $236 billion in debt and the licensing risk of a 17-mile pipeline, making it the only company to rank high in both dimensions. OpenAI itself has no balance sheet pressure (it doesn't hold assets), but as a computing power buyer, it bears the fulfillment risk of all sites—if any site is delayed, the "computing power gap" directly impacts its model iteration schedule. The second category consists of companies with "low financial risk and moderate licensing risk"—Microsoft (30/65), Amazon (28/55), Google (25/52), and Meta (45/58). The combined capital expenditure guidance for these four companies in 2026 is approximately $720 billion to $745 billion (Amazon approximately $220 billion, Alphabet $195 billion to $205 billion, Meta $130 billion to $145 billion, and Microsoft approximately $175 billion), far exceeding Oracle's scale. However, their delays can be absorbed by operating cash flow rather than new debt. They face common constraints related to electricity and licensing, with Microsoft having the highest electricity risk score (65) because it publicly acknowledges the "order backlog but inability to deliver" capacity constraints on Azure—precisely the reason why it has a large number of AI servers in inventory that are idle due to power supply delays. The third category is "structurally unique"—xAI (78/45). Its financial risk score is high (high leverage, reliance on a single founder), but its power and licensing risks are significantly lower than its peers because its strategy is to rapidly deploy gas turbines and mobile units for self-generation, and to replace new construction with warehouse renovations. Colossus' first 100,000-card cluster went live in approximately 122 days. This comparison provides an important conclusion: players whose core strategy is self-generation and rapid deployment have significantly less power risk exposure than players whose core strategy is grid connection—this is precisely why Oracle chose Bloom Energy SOFC, and why this event surprised the market: it chose a path that bypassed the grid, but was blocked by ground pipelines.

Core judgment

Judgment 1: This event is characterized as a "risk reallocation," not a "project termination." The actual content of the force majeure notice is only one point—if the project cannot go live by 2028, Oracle will seek a rent deferral of up to three years; it does not constitute a default event for the $18 billion financing, Oracle has not withdrawn from its major tenant status, Blue Owl has confirmed that its financial commitments remain unchanged, and the project's construction is still progressing. Reading it as "Stargate crashed" is an overinterpretation of the legal document.

Judgment Two: The key point is that "the risk of power supply failure is borne by Oracle, not the investors." This statement means that in this type of transaction structure, the most uncontrollable variable has shifted from the investors to the tenants. It also signals that when contracts begin to include clauses for liability relief in case of "no power," it indicates that the probability of this event occurring has been priced as "not ruled out" by professional institutions. Oracle is willing to take on this risk because in a market where there is no capacity without power, the cost of not signing is greater than signing the wrong contract.

Judgment Three: New Mexico is not the only obstacle; it is the first example to bring structural contradictions to the forefront. The full spectrum of obstacles includes: a 17-mile pipeline requiring 0.6 miles of state trust land for right-of-way; a 2,300-person town voting 44-0 to ban a data center; a large transformer requiring a 24-40 month wait; a capacity electricity bill increasing 11.5 times in three years; a petition against it garnering 500,000 signatures in Tennessee; and organized opposition activities in 49 states across the country. A common characteristic of these obstacles is that none require the passage of any new law. The suspension order legislation attempted by 14 states in 2026 failed to materialize, while $198 billion in projects were still blocked or delayed in the first half of that year—this demonstrates that the real binding force lies not at the legislative level, but at the executive and local levels.

Judgment 4: The midterm elections are less important than the "election process." The election results determine what tools (legislative or executive) will be used to continue tightening, while the election process has permanently altered developers' behavior: delaying the submission of sensitive projects, increasing community concessions, and adding more disclaimers to contracts. Oracle's notification six weeks before the election is no coincidence. These contractual changes will not regress after the election.

Judgment 5: By 2026, the bottleneck for AI infrastructure will have shifted from chips to electricity and licensing, and this is an irreversible shift. Among the seven uncontrollable factors, the fastest-transmitting are grid connection delays (95) and critical equipment delivery (88), while the slowest is chip supply (30). This ranking means that investing money in computing power is no longer enough to buy time for computing power to go online—GE Vernova has a backlog of 100 gigawatts, Siemens Energy has a backlog of approximately 136 billion euros, and the production capacity of the three leading gas turbine manufacturers is scheduled until 2030, while there is no other company besides IBM that can provide expedited production capacity for large transformers.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.