On September 24, Oracle issued a force majeure notice related to Project Jupiter, a 1,400-acre data center campus under development in New Mexico by Stack Infrastructure, a subsidiary of Blue Owl. Blue Owl has provided an $18 billion debt facility to support the project, which is part of the Stargate initiative involving OpenAI and SoftBank, and expected to come online in 2028. The project has faced delays related to natural gas pipeline deployment, water access, and environmental permitting.
Under the lease structure, Oracle cannot terminate its commitment. However, citing delays in delivering power to the site, Oracle will continue paying the lower development-stage rent rather than transitioning to higher operational payments. Oracle is also responsible for servicing the project’s debt obligations.
I believe this marks the first time a high-profile AI infrastructure participant has invoked the rarely used force majeure provision to delay payment obligations. The move highlights the critical importance of power availability and permitting timelines. More importantly, it exposes structural risks embedded within the rapidly expanding AI infrastructure sector.
The Hidden Risk in Plain Sight
Project Jupiter is important not because it appears headed toward failure, but because it provides one of the first public examples of how execution delays can ripple through the increasingly complex financing structures underpinning AI infrastructure.
In its Fall 2026 In The Gaps newsletter, Ares Management highlighted force majeure by an offtaker as one of the key risks facing AI infrastructure investments. Remarkably, the report was published only days before Oracle issued its notice. Ares identified Oracle as one of eight core counterparties, alongside Meta, Microsoft, Google, Amazon, Nvidia, OpenAI, and Anthropic, that collectively support more than $573 billion in AI infrastructure debt, leases, guarantees, and other obligations that form what Ares described as an “interconnected web” of financial commitments.
One of the primary risks identified by Ares was the growing execution bottleneck surrounding power procurement and grid interconnection [see here for power risk assessment]. In many regions, timelines have stretched beyond four years. At the same time, resistance to large-scale data center development continues to increase. New York recently became the first state to impose a temporary pause on environmental permitting for data center projects exceeding 50 MW (Executive Order 62). Opposition to data center expansion has also gained political visibility in Texas, where concerns over power consumption and infrastructure strain have entered the public debate, affecting the gubernatorial race.
The Infrastructure Trap
Ares described the theoretical mechanics through which AI infrastructure financing structures could come under pressure from non-financial execution risks. Oracle’s notice now provides a real-world example of those mechanics in action, even though Oracle and Blue Owl have both stated that Project Jupiter remains on schedule and that all core financial commitments remain intact.
The important issue is that the financing structure can obscure where the real risk resides. Debt investors often believe they are underwriting investment-grade corporate credit backed by a hyperscaler’s balance sheet. In reality, investors also assume the risks associated with utility delays, permitting obstacles, construction challenges, and contractual provisions that allow an offtaker to seek temporary relief from its obligations when qualifying events occur [see here]. The underlying risk is not solely credit risk. It is execution risk.
The Great AI Accounting Debate
Following publication of the Ares report, Michael Burry expanded the discussion in a Substack post outlining his bearish thesis on AI infrastructure. Using Ares’ $573 billion financing matrix alongside SEC disclosures, Burry argued that markets significantly underestimate both future financial obligations and the depreciation risk associated with AI infrastructure investments.
According to Burry’s analysis, Microsoft, Meta, Google, Amazon, and Oracle collectively carry roughly $3 trillion of future commitments, many of which remain embedded in contractual disclosures rather than appearing as conventional debt obligations. He estimates that Google alone carries approximately $900 billion of commitments, while Meta carries roughly $700 billion, a figure that could exceed $1 trillion when broader commitments are included.
Whether or not these estimates prove accurate, the broader observation deserves attention: much of the financial exposure associated with AI buildouts sits outside traditional balance-sheet metrics.
In short, lenders often commit billions to AI infrastructure projects because investment-grade companies such as Oracle, Google, and Microsoft serve as tenants or contractual backstops. But when physical bottlenecks emerge, those same companies may be able to invoke contractual provisions that reduce or defer certain obligations.
Compounding the Risk: GPU Depreciation
This perceived safety is further weakened by another challenge: the rapid depreciation of compute hardware. Data center buildings may have useful lives measured in decades, but the GPUs inside them do not. As new chip generations enter the market, existing hardware can lose 50% to 60% of its rental value within two years. Nvidia has been on a rapid product cycle, moving from Hopper (2023) to Blackwell (2025), Rubin (2026), and eventually Feynman.
When an offtaker invokes a contractual provision that lowers or suspends payment obligations during development delays, lenders remain exposed to assets that continue depreciating regardless of project progress. If restructuring or default occurs, hardware liquidation values may recover only a fraction of originally financed amounts.
Adding to the debate, hyperscalers have extended useful-life assumptions for AI infrastructure assets from roughly 2-3 years to 5 or 6 years. These adjustments reduce depreciation expense and improve reported earnings. For instance, Meta disclosed that extending the useful life of certain servers to 5.5 years reduced annual depreciation expense by approximately $2.9 billion. Based on similar assumptions, Burry estimates that Oracle’s earnings could be overstated by 26.9% and Meta’s by 20.8% by 2028.
The issue has become sufficiently prominent that Nvidia has pushed back on claims of rapid obsolescence. In a September 27 investor presentation titled “NVIDIA AI Infrastructure Retains Value Beyond Accelerated Depreciation Schedules,” Nvidia argued that 4-6 years is a reasonable depreciation window, noting that A100 systems introduced in 2020 remain widely deployed. Jensen Huang further highlighted rising H100 rental rates and pointed to reports that CoreWeave has already booked portions of its A100 fleet through 2029. [For Burry’s latest response to Nvidia, see here.]
The Vendor Financing Flywheel
Vendor financing introduces another dimension to the AI infrastructure ecosystem. The implicit assumption behind many of these arrangements is that Nvidia and other hardware providers can improve the financeability of AI infrastructure projects by supporting revenues, lease structures, or residual values. Nvidia’s commitments to CoreWeave, residual-value support arrangements involving OpenAI, and Broadcom’s financing structures associated with Anthropic represent different variations of the same concept.
By supporting cash flows, vendors help transform otherwise difficult-to-finance infrastructure investments into bankable assets. The result is a potentially self-reinforcing cycle. Hardware vendors support demand and infrastructure expansion. That support helps AI labs and neoclouds attract external financing. The new financing funds additional hardware purchases, which in turn generates further demand for vendor products. The structure bears similarities to financing models used during earlier waves of infrastructure investment, particularly in telecommunications and Internet infrastructure during the late 1990s.
Concluding Thoughts
There is still remarkably little visibility into the ultimate trajectory of the AI infrastructure market. The limits, capabilities, and long-term return on investment of large language models remain uncertain, as do their broader impacts on productivity, employment, and economic growth. At the same time, the competitive landscape continues to evolve rapidly. Open-weight models are challenging the economics of frontier labs, while custom silicon developed by Google, Amazon, and others is increasingly competing with Nvidia-based architectures.
At the center of this dynamic are OpenAI and Anthropic, the critical lynchpins to watch. While hyperscalers provide the financial backstop and private credit provides the funding, OpenAI and Anthropic provide the underlying demand narrative that justifies the multi-hundred-billion-dollar infrastructure buildout. If either frontier lab experiences demand softness, monetization challenges, or margin compression, the revenue required to support this vast web of leases, guarantees, and hardware financing comes into question. In fact, Oracle’s force majeure notice may ultimately prove less important for its immediate financial impact than for what it revealed: in the AI infrastructure boom, some of the greatest risks are not technological or financial, but physical, regulatory, and contractual.
Yet the market’s willingness to fund these commitments is driven by more than commercial expectations alone. AI has increasingly become a geopolitical imperative [see here]. First, it is viewed as central to maintaining technological leadership over China. Second, it is seen by many policymakers as a catalyst for productivity growth at a time when the U.S. faces more than $40 trillion in national debt and mounting fiscal pressures. The stakes, therefore, could hardly be higher. For technology companies, investors, and policymakers alike, AI has become a race they cannot afford to lose. If we are in an AI bubble, it is precisely this must-win mentality that will keep inflating it, extending investment well beyond what fundamentals alone might justify and significantly amplifying systemic risks when the cycle eventually turns.
For additional insights into AI infrastructure investments, download my recent Xona Insight Note: The Diverging Investment Strategies Behind AI Infrastructure.