Oracle Invoked Force Majeure on Its $165 Billion AI Data Center Because New Mexico Can't Deliver the Gas on Time
Oracle sent a force majeure notice to Blue Owl Capital this week over "Project Jupiter," a 2.45-gigawatt, $165 billion AI data center campus in New Mexico tied to the OpenAI and SoftBank Stargate buildout. The headline sounds like Oracle backing out. It isn't. It's Oracle telling its own investors, in a legal filing, that the physical infrastructure behind the AI boom (pipelines, permits, turbines) is moving slower than the compute demand it's supposed to feed, and asking to be excused from penalties while that gap gets sorted out.
What force majeure actually means here
Force majeure is a contract clause that excuses a party from its obligations when something outside its control gets in the way. Oracle isn't invoking it to exit Project Jupiter as the main tenant. It's invoking it to protect itself from penalties if construction misses the 2028 completion target. Blue Owl, the developer on the other side of the notice, was careful to say the filing "does not change the financial commitments to this multi-year project." Read those two statements together and you get the real story: Oracle wants legal cover for a delay it didn't cause, while Blue Owl wants to reassure its own investors that the money is still coming. Both can be true at once, and the gap between them is exactly where the actual risk lives.
The three things actually holding this up
None of the holdups are about chips or capital. They're about the boring infrastructure underneath AI data centers that almost never makes it into a keynote:
An Energy Transfer gas pipeline meant to feed the site is now running nearly six months behind, with a new delivery target of February 1, 2027, after regulators repeatedly rejected the original route and forced a reroute. The 2.45-gigawatt campus also depends on Bloom Energy fuel cells for power, and the air-quality permit those fuel cells need is still pending, with state environment officials facing a November 23 deadline to decide. Oracle's stock dropped roughly 3% on the initial CNBC report of the force majeure filing, which tells you the market read this as more than routine paperwork.
Why this matters if you rent GPU capacity from anyone
If you're building a business on top of "cheap AI cloud" capacity, whether that's a GPU rental marketplace, a wrapper around someone else's inference API, or your own fine-tuned model hosted on rented hardware, you are downstream of exactly this kind of bottleneck. A vendor's marketing page can promise unlimited scale while the actual data center backing that promise is waiting on a pipeline permit that a state regulator hasn't signed off on yet. Oracle is one of the most sophisticated infrastructure operators on earth, backed by OpenAI and SoftBank's money, and it still can't force a gas pipeline to move faster than the regulatory process allows. A smaller GPU cloud provider has even less leverage over the same physical constraints.
The honest counter-take
I don't want to overstate this into "the AI buildout is collapsing." It isn't. Force majeure notices on multi-year infrastructure projects are a normal, almost boring part of how these deals get structured, precisely because everyone involved knows pipelines and permits slip. Oracle and Blue Owl are both signaling, in their own ways, that the underlying commitment hasn't changed. The risk here isn't that Project Jupiter collapses. It's narrower and more useful than that: promised capacity and delivered capacity are not the same thing, and the gap between them is measured in months, sometimes years, even for the best-capitalized players in the industry.
What I'd actually do
Before you build a paying customer's SLA around any GPU cloud provider's uptime promise, ask which specific data center your workload is provisioned against, and whether that facility is live today or still under construction. If your vendor can't answer that clearly, assume you're exposed to exactly the kind of delay Oracle just disclosed about its own flagship project. Diversify your inference across at least two providers on different physical infrastructure if your business genuinely depends on availability, the same way you'd avoid putting all your production traffic on a single database region.
Author
Lukas
@lukcombinator