The AI bubble is no longer a concern confined to skeptical analysts and short-sellers. The Bank for International Settlements, the institution owned by the world’s central banks, has published a report drawing direct parallels between today’s AI infrastructure buildout and history’s great financial manias, from the British railway mania of the 1800s to the dot-com bubble. And hyperscale tech companies, once the most vocal proponents of unlimited AI spending, are beginning to show cracks.
Oracle’s stock has fallen more than 40% in the past month. The company filed an unusually detailed SEC disclosure outlining specific financial risks tied to its AI bets, a rare departure from boilerplate language that analysts say signals genuine alarm inside the company’s leadership.
The Oracle Problem
At the center of the concern is Oracle’s role in the Stargate project, the US$500 billion (approximately £397 billion) AI data center buildout spearheaded with SoftBank and MGX. Oracle may need to borrow approximately US$25 billion (approximately £20 billion) per year to fund its share, and its business model depends heavily on a single customer: OpenAI.
“OpenAI cannot pay its own bills,” said Tobias Mann of The Register, who analysed Oracle’s SEC filing. “It is like leasing property to someone who may or may not be able to pay, but probably can’t.”
Oracle’s filing discloses that there is no guarantee OpenAI will renew or pay its leases, that up to US$155 billion (approximately £123 billion) of committed capacity could become stranded if OpenAI defaults, and that permitting delays, power sourcing issues, and growing community moratoriums against AI data centers all threaten the Stargate timeline.
Hyperscaler Capex: A Trap of Their Own Making
The Big Four hyperscalers have committed staggering sums to AI infrastructure in 2026 alone:
| Company | 2026 AI Buildout Capex |
|———|———————-|
| Amazon | >US$200 billion |
| Microsoft | US$190 billion |
| Google | US$180 billion |
| Meta | US$140 billion |
The combined US$710 billion (approximately £564 billion) dwarfs any previous technology infrastructure cycle. The BIS report specifically warns that the AI investment boom has attracted “a lot more capital than the resulting industry could actually produce”, a pattern the institution says has historically preceded severe economic corrections.
The hyperscalers face a structural catch-22: if they do not spend on AI infrastructure, investors penalise them. If they spend and the promised AI revenue fails to materialise, investors will penalise them too. Amazon, Microsoft, and Google have diversified cloud revenue streams that can absorb the risk. Meta can repurpose its GPU fleet for advertising. Oracle, with its near-total exposure to OpenAI, cannot.
Enterprise Pushback and the Search for Alternatives
Enterprise customers are voting with their wallets. Rising token prices from OpenAI and Anthropic are driving companies to explore open-source models, private cloud deployments, and on-premises AI infrastructure. Palantir CEO Alex Karp criticised “the very gate-kept world of frontier AI,” saying customers are demanding transparency and predictable pricing that the frontier labs have not delivered.
Apple’s private cloud computing architecture, and Google’s adoption of a similar approach, signals a broader industry shift toward data-local AI. Developers are actively optimising models to use fewer tokens, a sign that current pricing levels are unsustainable for real-world deployment.
Even xAI, Elon Musk’s AI venture, has overbuilt its infrastructure and is now leasing excess capacity to other companies, suggesting that even the most well-funded AI labs cannot fully utilise the hardware they have purchased.
Meanwhile, Chinese AI models are gaining traction as lower-cost alternatives, though the geopolitical uncertainty created by the Trump administration’s short-lived export controls on Anthropic’s Mythos models has added risk for enterprises considering them.
Not Every Company Will Survive
The BIS report draws a distinction between past bubbles and the current one: while the British railway mania and the dot-com crash destroyed companies, they also built lasting infrastructure, rail networks and internet backbones, that underpinned future growth. The same may prove true for AI data centers. But the transition period, the BIS warns, could take down parts of the global economy with it.
Sources: Even banks and hyperscalers are now sounding the alarm about the AI bubble (The Register, July 6); Bank for International Settlements report (July 2026); Oracle SEC filing (June 2026)