Bloomberg Sees Echoes of the 2008 Crisis in the AI Boom's Circular Deals

Look at the biggest AI deals of the past year and the same pattern keeps repeating. Nvidia announced its intention to invest up to $100 billion in OpenAI, while OpenAI buys systems from Nvidia totaling gigawatts of capacity. AMD gave OpenAI warrants to buy up to 160 million AMD shares in exchange for an order of 6 GW of chips.

Oracle, according to press reports, signed a contract with OpenAI worth roughly $300 billion, borrowing money and building data centers to fulfill it. Bloomberg devoted an episode of its Big Take podcast to this arrangement: Shawn Donnan explains to host David Gura why such a "money loop" resembles 2008. The episode announcement itself is short and gives no figures.

The numbers above come from earlier reporting on the deals.

How the loop works

The logic is simple. A chip or cloud supplier gives money to a model developer. The developer spends it buying from that same supplier. The supplier reports higher revenue, its market capitalization rises, and new capital is raised against that capitalization. Formally, each deal is legal and makes economic sense: OpenAI needs compute, and Nvidia and AMD need guaranteed demand. The problem is that a significant share of that demand is financed by the sellers themselves. Real outside payment, meaning revenue from subscriptions, API usage and enterprise contracts, is still noticeably smaller than the planned capacity commitments.

Where 2008 comes in

Bloomberg draws the parallel not from the subject of the deals but from their structure. Before the 2008 crisis, mortgage securities were sliced up and resold within the financial system, and risk ended up concentrated among a narrow circle of participants who regarded one another as reliable counterparties. Today a similar interconnectedness is emerging among a handful of companies: if one node, such as OpenAI, fails to meet its obligations, chip suppliers, cloud landlords and their creditors will all suffer.

The second part of the thesis concerns global imbalances. Building AI infrastructure pulls enormous flows of capital into the US and into a narrow set of assets. That widens distortions in trade and savings between countries, and the high share prices of tech giants give investors a sense of wealth that itself sustains spending. This is the publication's argument, not a confirmed forecast of a crisis.

How the current situation differs

There are counterarguments. The main investors still pay primarily out of operating cash flow and equity, not bank leverage. Demand for compute is real: GPU utilization is high, and AI labs' revenue is growing. But the share of debt is increasing. Separate project vehicles for data centers, large bond issues and loans secured by chips are appearing. Instruments like these are precisely what make risk less visible.

Key elements of the loop:

• Nvidia: up to $100 billion in planned investment in OpenAI, tied to deliveries of systems.
• AMD: 6 GW of chips for OpenAI and warrants for up to 160 million shares.
• Oracle: a contract of around $300 billion, financed in part with borrowed money.
• Common feature: the supplier is simultaneously an investor and the main beneficiary of the customer's purchases.

What to watch next

Investors and tech executives would do well to separate two measures: stated capacity commitments and the actual revenue of end customers. If the former grows faster than the latter, the loop is held together by expectations. Financial reports are worth reading for the share of revenue from related parties, off-balance-sheet obligations and data-center debt maturities. Companies building products on one lab's API should reasonably line up a second provider and keep contracts flexible: a failure at one node quickly reaches prices and compute availability. Finally, GPU rental prices will be a signal. If they start falling while construction continues, supply is outpacing solvent demand.