Each line is a deal, and its arrow points the way the money goes. Follow the arrows and the money often comes back to where it started.
Bigger bubbles are bigger companies by market value or latest valuation, ranked roughly rather than drawn to scale. Many headline figures are letters of intent, not contracts, and are drawn dashed; each deal's status is shown in its detail. The $600B of Nvidia–OpenAI financing reported in July 2026 became a signed $105B guarantee in August: headline and contract can differ by half a trillion dollars.
The AI circular economy, mapped. Nvidia funds the labs, the labs buy compute, the clouds buy Nvidia's chips. Every deal above is sourced.
Anthropic’s amended S-1, filed on 2 November, adds up its purchase commitments and leases at about $895.4B, substantially all of it for compute, against $17.0B of revenue in the third quarter. About $447.5B falls due before the end of 2030; $213.8B can be cancelled or depends on conditions. Its suppliers are funding part of the bill. Broadcom, Google and Amazon have up to $77B of note facilities open to it, and on 26 October four Wall Street firms committed $63B to a vehicle that buys Broadcom chips and leases them to Anthropic, with Broadcom supporting their resale value. On 4 November TeraWulf said Google will back the debt on Anthropic’s $19B lease in Kentucky. Google is Anthropic’s investor and supplier, and holds warrants over about 14% of TeraWulf. Its backstops on other companies’ data centres reached $75.7B in September, counting those it has promised.
The market has started to charge for it. OpenAI cut its round from about $30B to about $20B, at $1.25T. Oracle capped new funding at $20B on 28 October, filed to raise all of it five days later and fell 8.8%. S&P cut Meta’s outlook to negative. Marvell agreed to buy Lumentum for $115B in stock, and by 4 November that stock had lost 36% of the value in the headline. At midday on 5 November, after a Democratic sweep of Congress raised fears for the data-centre build-out, the Nasdaq was down about 4% and more than 10% below its October record, and CoreWeave touched $40.58, close to where it listed in March 2025.
"It's not circular because we put a little bit of money in, and a lot of money comes back."
Jensen Huang, Nvidia CEO, at the Goldman Sachs Communacopia + Technology conference, 10 September 2026, with $99B of Nvidia's money in its customers
"Equity and debt investments in the firms depreciate in value precisely when future revenue streams tied to them are impaired."
International Monetary Fund, Global Financial Stability Report, 13 October 2026, on the “wrong-way risk” in circular AI financing
"We respectfully disagree with S&P's assertion about Meta's AI monetization capabilities."
Meta spokesperson to CNBC, 30 October 2026, after S&P cut Meta’s credit outlook to negative
Historical scores are applied retrospectively by hand, using the same six indicators. They are judgements about the record, not measurements — each carries its basis, listed below. The telecom marker is the closest structural analogue to what this map shows: Lucent, Nortel and Motorola lent their own customers the money to buy their equipment, booked the sales as revenue, then wrote the loans off.
What this is. A measure of how closely the current picture matches the characteristics common to historical asset bubbles. It is not a forecast, it says nothing about timing, and it is not investment advice. Half the weight is computed from the tracked deals on this page; half is hand-set and sourced.
What it is not. These tracked deals were selected because they are circular — that is the point of the map. The circularity figure describes this deal set, not the market as a whole, and the denominator is always shown. An index where every needle pointed the same way would be a press release; market behaviour currently points the other way, and says so.
Newest first, by the date each deal began, each with its source. The tags say how real a deal is: