Oracle's SEC Filing and the BIS Report Just Made the AI Bubble a Lot Harder to Dismiss

Two fairly significant voices have joined the growing chorus of AI sceptics, and neither of them is a disgruntled blogger. The Bank for International Settlements, which functions essentially as the central bank for the world's central banks, published a report warning that the AI bubble has uncomfortable similarities to historical financial manias: British railway speculation in the 1800s, canal investment frenzies, the dot-com collapse. The common thread in all of those was that capital flooded in far faster than the industry could ever realistically absorb or return. The BIS thinks AI might be next.

Around the same time, Oracle filed documents with the SEC that read less like routine boilerplate and more like a company genuinely reckoning with how exposed it is. Oracle's stock has shed over 40 percent in the past month. That's not a dip. That's investors quietly reconsidering their life choices.

Oracle's position in this is worth understanding, because it's genuinely unusual. The company signed up as one of the founding partners of Stargate, the OpenAI-led datacenter initiative alongside SoftBank and MGX. Oracle is on the hook for financing a chunk of what is nominally a $500 billion project, with its share reportedly somewhere around $300 billion. To fund that, Oracle has reportedly needed to borrow around $25 billion a year. The slight problem is that the customer at the end of all this, OpenAI, currently cannot pay its own bills and relies entirely on external capital to stay afloat.

So the business model, simplified, is: Oracle leases datacenter space it doesn't own, then sub-leases it to a company with no reliable revenue, in the hope that AI becomes profitable enough for everyone to get paid. It does bear a passing resemblance to WeWork, now you mention it.

The SEC filing catalogued the risks in unusual detail. Lease commitments that might go unpaid. The difficulty of accurately forecasting demand when demand is largely theoretical. Permitting obstacles. Growing local opposition to datacenter construction. Power availability problems. The reality that if OpenAI can't meet its obligations, there is no obvious queue of alternative customers ready to absorb hundreds of billions of dollars worth of compute capacity. xAI, for context, has reportedly been sub-leasing spare hardware because it has more infrastructure than it knows what to do with.

For the hyperscalers with diversified revenue streams, a partial AI collapse is survivable. Meta can point its GPUs at advertising algorithms. Microsoft, Google, and Amazon have cloud and enterprise businesses that will carry them through a correction. Oracle doesn't have that cushion. Its exposure to the AI build-out is proportionally enormous, which is probably why the market is pricing in some fairly grim scenarios.

The broader concern flagged in the BIS report goes beyond the hyperscalers themselves. Datacenter construction firms, infrastructure suppliers, component manufacturers, power grid operators, all of these sit downstream of capital commitments that rest on the assumption that AI demand will materialise at scale and at the right price. If it doesn't, the fallout travels well beyond Silicon Valley balance sheets.

There are also structural problems building on the demand side. Enterprise customers are increasingly uncomfortable with the opacity of frontier AI providers, the unpredictability of pricing, the lack of control over their own data, and the difficulty of switching providers. Palantir's Alex Karp, never one to understate anything, made this point publicly in early July, arguing that businesses fundamentally want transparency, portability, and cost predictability, none of which the current frontier AI market offers particularly well. He was obviously pitching Palantir at the same time, but the underlying observation isn't wrong.

Meanwhile, the software built around these models is increasingly optimised for using them less. Developers are building systems to reduce token consumption, route simpler queries to cheaper models, and avoid frontier systems for anything that doesn't strictly require them. The logic being promoted is that frontier models are only necessary for perhaps a quarter of real-world queries. If that framing takes hold, the revenue projections underpinning hundreds of billions of dollars of infrastructure spending start to look optimistic.

The squeeze is coming. At some point these companies need to demonstrate that AI can be run profitably without indefinite venture capital subsidy. If they raise prices to cover costs, they risk pushing customers toward open-source alternatives or Chinese models that are already competitive enough to make the comparison uncomfortable. If they hold prices down, they continue burning cash at a rate that can't go on forever.

The bubble probably won't disappear entirely when it corrects. Railways survived the railway mania. The internet survived the dot-com crash. But a lot of the money currently sloshing around will evaporate, and the companies left holding the most speculative commitments will feel it hardest. Oracle, right now, looks like a fairly strong candidate for that list.