Force Majeure Meets the AI Buildout: Oracle's Project Jupiter Warning Shot

Introduction

For three years, the defining image of the artificial intelligence infrastructure boom has been the construction site: turbines spooling up, substations rising out of scrubland, and a steady drumbeat of capital expenditure announcements from Microsoft, Google, Amazon, Oracle and Meta. The prevailing assumption was that the bottleneck was money, and that money was the only thing standing between the labs' ambitions and gigawatts of live compute.

That assumption took a hit this week. On 24 September, Bloomberg reported — and Reuters carried the story — that Oracle had issued a force majeure notice to Blue Owl Capital over Project Jupiter, the 1,400-acre AI campus it plans in Doña Ana County, New Mexico. Oracle shares fell roughly four percent in premarket trading. The company did not deny the report, but its public line was notably careful: "Project Jupiter remains on our planned schedule. We are fully committed to New Mexico and confident in our path forward." Alongside that, it said it was "reimagining" the campus's power plan.

A legal hedge and a reassurance in the same press release is an unusual combination. It is also, arguably, the most important cloud infrastructure story of the week — not because a single 2.4-gigawatt project slipped, but because it shows what happens when the physical world stops being a formality. The same week also delivered the counterweight: Akamai announced an $11.6 billion, seven-year commitment from Anthropic for distributed cloud infrastructure, a deal that could grow by another $9 billion. One story is about capacity that cannot arrive on schedule. The other is about a hyperscaler betting billions that capacity does not need to arrive in a single building at all.

Together they sketch the real shape of cloud infrastructure in late 2026: the grid, the permitting queue and the contractual fine print are no longer background conditions. They are the product.

Jupiter by the Numbers: 1,400 Acres, 2.45 Gigawatts, and a Loan Book Underwater

Project Jupiter is not a colocation hall. It sits near Santa Teresa and Sunland Park on the New Mexico–Texas border, and at roughly 2.45 to 2.5 gigawatts of planned capacity it would be among the largest computing campuses anywhere on earth. It is tied to Oracle Cloud Infrastructure and to OpenAI's Stargate buildout, which is what makes it a story well beyond Oracle's own balance sheet — the largest single AI infrastructure bet in the United States is now partly underwritten by a developer's debt.

The financial structure is where the strain shows. Developers have discussed as much as $165 billion in industrial revenue bond and related investment if the full vision is realized, backed by a consortium of banks that put together roughly $18 billion in loans. Those loans, according to reporting, have already been trading in the high-80s to low-90s cents on the dollar. Bonds priced at 87 cents are not mispriced by accident. Someone with money is pricing a meaningful probability that the project does not deliver on its original schedule.

The physical progress is real, and so is the political friction. Construction on the data center buildings has been underway since late 2025, and county updates in September put the campus at roughly 26 percent complete. Oracle and its partners have pitched a substantial local benefit package: hundreds of millions of dollars in tax and community payments, closed-loop cooling, no public drinking water consumed for cooling or generation, Oracle bearing its own power costs so ratepayers are not asked to fund the campus, and a 100 percent carbon-free matching commitment by 2031. For a county that has been left with very little from the fossil fuel economy that defined it, that is a serious offer.

The problem is the power. Earlier designs for the on-site microgrid leaned on gas turbines and diesel. Oracle dropped that design, and Bloom Energy fuel cells have become the centerpiece of the current plan. That pivot matters because it came late, and because fuel cells still need fuel — which brings the site back to a gas supply question that has been a persistent legal sticking point. In July 2026, the New Mexico State Land Office rejected Energy Transfer's permit application for the Green Chile Project, a 17-mile natural gas pipeline extension intended specifically to supply fuel for Jupiter. Environmental groups challenged notices and water use; the New Mexico Supreme Court lifted some of those stays in mid-September. The air-quality permit and the microgrid have become the political and legal choke point for the entire campus.

This is why the force majeure notice matters more than a routine contract provision. Force majeure is not a shutdown and it is not an exit. It is a contractual device for excusing performance when circumstances beyond a party's control intervene. In this case, the report is that Oracle is trying to defer payments if the facility misses its 2028 in-service date, rather than walking away as the main tenant. That is a very different posture from cancellation. It is a tenant who has decided the asset is still worth holding but the schedule is no longer something it will pay for on time.

The Take-or-Pay Squeeze and the Memory Pre-Purchase

It is worth sitting with what has actually changed here, because it is easy to read Project Jupiter as a story about one company's legal team being clever. The more useful reading is that risk allocation in AI infrastructure has become an explicit, priced, and litigated category.

In hyperscale contracts, the standard protections are take-or-pay structures on power, milestone payments tied to delivery, and capacity commitments that survive schedule slippage. Those protections cut in both directions. A developer who has sold a 2.45-gigawatt take-or-pay obligation to a hyperscaler is exposed if the campus cannot energise on time. A hyperscaler who has committed billions to that capacity is exposed if it cannot. Oracle's notice is an attempt to move the schedule risk from its side of the ledger to Blue Owl's — and Blue Owl's own statement that the notice does not change the project's financial commitments suggests the transfer did not fully succeed.

The Akamai-Anthropic deal announced on the same day illuminates what the other side of that ledger looks like when it is working. Akamai said Anthropic had committed $11.6 billion over seven years for cloud infrastructure services, requiring an estimated $5.5 billion in capital expenditure from Akamai over the next two years to support it, including the pre-purchase of memory. That is the largest contract in Akamai's 28-year history. It follows $2.8 billion in multi-year commitments the two signed earlier in 2026. Revenue is expected to begin in the second half of 2027, run between roughly $150 million and $300 million for full-year 2027, and be fully ramped by the end of 2028, after which the annualised run rate is expected to reach about $1.7 billion. The additional $9 billion is an option rather than a committed deliverable-backed increment. Akamai's stock rose about 4.9 percent, or $5.43 per share, to $115.84 following the announcement.

The detail that deserves the most attention is "including the pre-purchase of memory." That single clause reveals more about where the constraint in AI infrastructure sits today than most market commentary does. In a GPU buildout, the accelerator is the scarce item and everyone assumes the balance of the system — CPU, DRAM, networking, storage — follows. Anthropic's arrangement suggests the opposite has become true for at least one major buyer. The economics of agentic inference and long-context workloads are pushing hard on memory bandwidth and capacity. A frontier lab that can secure its accelerator allocation is not necessarily the lab that can secure the DRAM to feed it. Akamai's CFO Ed McGowan described the structure as take-or-pay, noting that because Akamai is guaranteed payment upon delivery, revenue should hold steady for the remainder of the contract.

There is also a strategic subtext. Anthropic selected Akamai's distributed cloud rather than a GPU-centric hyperscaler, for CPU workloads. Akamai's co-founder and CEO Tom Leighton framed it as extending "high-performance cloud computing from the core to the edge." McGowan said the contract has pulled in inbound interest from prospects who had not previously considered Akamai capable of operating at that scale. And Akamai noted the deal does not prevent it from working with Anthropic's competitors.

This is the structural point that has been building for two years. The set of companies that can credibly absorb frontier-lab workloads at scale is widening, but the widening is happening in the CPU, memory and distributed-colocation layers, not in the accelerator layer. An earlier wave of coverage in this space described the compute continuum and why AI was turning cloud strategy inside out. Jupiter and the Akamai deal are the same thesis arriving from opposite directions.

What Infrastructure Buyers Should Take From This

For the enterprises and platform teams actually planning around compute capacity, the practical lessons are not abstract.

First, power and permitting belong in the infrastructure equation, not the facilities annex. Jupiter is roughly a quarter built and its binding constraint is a fuel-supply permit for a pipeline that was rejected once already. Anyone modelling capacity arrival dates from signed contracts and vendor roadmaps alone is modelling the wrong variable. The grid and interconnection queue is the schedule.

Second, read the payment structure, not the headline number. A $165 billion campus and an $11.6 billion cloud contract are both marketing figures until you know who bears the risk of a miss. Jupiter's loan book trading at 87 cents is the market's own read on that risk. Akamai's take-or-pay is the same risk, allocated to the party that actually builds.

Third, expect more force majeure. A hyperscaler tenant is far more likely to invoke it and stay than to walk away, because the alternative is writing off a multi-year capacity reservation in a market where demand is still rising. Expect tenants to demand renegotiated milestone schedules, delay-linked payment terms, and power-availability conditions precedent across every large contract signed in the next two quarters.

Fourth, memory is now a first-class procurement category. The DRAM pre-purchase in the Akamai deal is not a footnote; it is evidence that the bottleneck list is longer than the accelerator.

And fifth, the counter-cyclical signal is real. While a mega-campus stalls in the desert over gas permits, a company built on a distributed footprint of smaller sites is signing the largest contract in its history. That is not a paradox — it is the market repricing what it considers low-risk capacity. For readers tracking the broader AI infrastructure buildout, the two stories together suggest that the next phase of cloud competition will be won not by who reserves the most gigawatts, but by whose gigawatts are actually energised, on time, on a schedule someone has thought hard about.

Conclusion

Oracle's force majeure notice is a legal document, not a confession. The company says the campus remains on schedule and remains committed. Both statements can be true, and both probably are — the notice exists precisely so that Oracle can keep building while deferring the cost of any slippage.

But the market's reaction was to sell the stock, the developer's debt trades at a steep discount, and the power plan is being rewritten in mid-construction. Those are not the facts of a project that is comfortably on track. They are the facts of a project that has lost the presumption of certainty, which in an industry financed on take-or-pay schedules and 2028 revenue ramp targets is a meaningful loss.

The optimistic reading of the week belongs to Akamai and Anthropic: a frontier lab has found that it can buy frontier-scale infrastructure from a company that was not yesterday's obvious candidate, and the price of that capacity can be fixed seven years out. The cautious reading belongs to everyone planning against a 2027 or 2028 compute budget, because Project Jupiter has just demonstrated that the distance between a signed megawatt and an energised one is, in 2026, a distance measured in permits and court rulings.

Both readings are true at once. That tension is not a bug in the AI infrastructure market. It is the market.

Images

A data center cold aisle lined with server racks, the kind of facility capacity eventually reaches once power and permits clear.