Oracle Counts the Ways Its AI Bet Could Blow Up in Its Face
Oracle is spending hundreds of billions building AI datacenter capacity, mostly for OpenAI. And now, buried in a regulatory filing, the company is quietly listing all the ways this could end very badly.
The filing covers Oracle Cloud Infrastructure, and it reads less like a business update and more like a stress test. Big Red has tied itself to long-term infrastructure commitments, leasing capacity through partners like Crusoe rather than building its own. That model works fine when customers pay on time and keep renewing. It becomes a problem when they don't.
Oracle doesn't name OpenAI directly, but the relationship is hard to ignore. Earlier this year, Oracle signed on to the Stargate initiative alongside OpenAI, SoftBank, and MGX, a half-trillion-dollar ambition to blanket the planet in AI infrastructure. Oracle's slice of that deal reportedly amounts to $300 billion in capacity over five years, with the company also managing OpenAI's flagship facility in Abilene, Texas. On top of that, Oracle claims roughly $155 billion in remaining performance obligations from other customers.
The numbers look impressive until you remember that OpenAI has never actually turned a profit. Its ability to pay Oracle's invoices depends entirely on its continued success raising capital from investors. Oracle acknowledged this plainly: 'Our business is, and may continue to be, exposed to risks of customer non-payment and non-performance.' That sentence is doing a lot of work.
And even paying customers might not stick around. If contracts aren't renewed, Oracle warns it may be unable to re-lease or repurpose that capacity on acceptable terms. Which means it could end up holding expensive infrastructure with nobody to bill.
Then there's the power problem. Oracle says it is already struggling to secure reliable, affordable electricity for its datacenter buildout. Demand for power is outstripping supply globally, and energy prices are volatile, particularly vulnerable to extreme weather and regional market quirks. Where customer pricing is fixed, Oracle eats the difference.
The physical act of building datacenters at this scale carries its own risks. The filing rattles through a fairly alarming list: suitable permitted land, reliable power interconnection, networking hardware, GPUs, memory components. Any of these can be delayed, restricted, or priced out of reach. Governments can impose construction moratoria. Environmental and zoning regulations can stall projects for years. Engineering and contractor problems can blow budgets. The list goes on.
Despite all of this, Oracle doesn't see a clean exit. The filing is explicit: if the company stops pouring money into AI infrastructure and products, it risks falling behind on both technology and industry standards, which would hurt its competitive position just as badly as overextending itself. Trapped between the cost of pushing forward and the cost of pulling back, it has opted to keep accelerating.
During its Q4 earnings call, Oracle said it plans to spend $70 billion on capital expenditure in fiscal year 2027, up from around $55 billion in 2026. To fund that, it wants to raise roughly $40 billion in new debt and equity next year, piled on top of the $18 billion it already borrowed back in September.
Markets are not convinced. Oracle's stock has dropped more than 40 percent over the past month.
Larry Ellison has made bold bets before. This one is in a different league.