Jim Rickards spotted a pattern in the AI industry that nobody else is talking about. The biggest companies building AI keep investing in each other, and the money that changes hands tends to come back around.
A chipmaker takes a stake in an AI startup. That startup buys the chipmaker’s chips with the money it just received. A cloud provider funds a model builder’s expansion, and the model builder pays the cloud provider to run its systems. OpenAI takes funding from a chip company that then gets paid for the hardware it ships. Each deal looks like growth on its own. Rickards says the picture gets murkier when you trace the dollars through the full loop.
His question is structural. How much of the revenue driving AI valuations comes from genuine outside customers, and how much comes from the same handful of companies passing money back and forth?
Rickards has been reading financial architecture as a vulnerability map for most of his career. In 2009, the Pentagon brought him into a top-secret Applied Physics Laboratory outside Washington to run the Defense Department’s first-ever financial war game. About sixty experts from the military, intelligence, and academic communities were split into teams representing the United States, Russia, China, East Asia, and a catch-all group. The only weapons allowed were financial instruments — currencies, stocks, bonds, derivatives. By the end of two days, the Chinese team had emerged in the strongest position, partly because the American and Russian teams had spent so much energy attacking each other that they weakened their own standing. Rickards wrote about the exercise in his 2011 book Currency Wars and returned to the Pentagon in 2015 for a smaller session on financial conflict in the South China Sea. The throughline of both exercises: the financial system is infrastructure, and infrastructure has failure points. When he looks at AI companies passing money in circles and calling it revenue, he is reading the architecture the same way he reads a payments network or a currency regime — looking for where the load-bearing wall is thin.
The circular spending thesis
Rickards laid out the pattern in a July 14, 2026 presentation for the AI Black Paper campaign, released through GlobeNewswire. The core claim is that AI’s biggest players have built a network of cross-investments where capital flows between the same names repeatedly. Nvidia invests in an AI lab that turns around and buys Nvidia GPUs, Microsoft backs an AI startup that pays Microsoft for Azure cloud credits, and OpenAI takes funding from a chip company that then gets paid for the hardware it ships.
On its own, each arrangement resembles ordinary venture investing, and companies back partners all the time. Rickards acknowledges this. His concern is scale and concentration. When the same dollars rotate between a small group of companies, reported revenue can look larger and more durable than the underlying demand actually is.
There is a certain elegance to the arrangement. The same dollar creates growth on two balance sheets at once — the investor records a stake, and the investee records revenue from the investor’s own product. Wall Street calls this “strategic partnership.” The rest of us might call it a mirror.
The distinction matters because stock prices depend on it. If a meaningful share of AI revenue comes from companies funding their own customers, the true strength of the market is harder to judge than the numbers suggest. Rickards says that gap is where investors get caught off guard.
Why the loop amplifies the debt problem
The circular spending thesis connects directly to Rickards’ broader AI debt framework, laid out in the Rickards AI debt warning. AI companies issued $236 billion in debt during the first five months of 2026, up from $200 billion for all of 2025. The borrowing pace is accelerating. To put the scale in context: the five largest hyperscalers — Amazon, Alphabet, Microsoft, Meta, and Oracle — issued about $122 billion in corporate bonds in 2025, more than four times their average annual borrowing of roughly $28 billion between 2020 and 2024, according to OECD and Bank of America data. Goldman Sachs estimates that AI-related borrowing now accounts for roughly 30% of net investment-grade bond issuance. The OECD projects $4.1 trillion in hyperscaler capital expenditure between 2026 and 2030, with nearly four-fifths of the external funding expected to come from debt markets. The $236 billion is the opening tranche of a financing wave that will run for years.
If revenue from outside customers is real and growing, the debt funds expansion that pays for itself. If a chunk of that revenue is circular, the debt funds spending that feeds back into itself without generating independent returns. The debt still needs to be repaid from actual profits. Rickards’ line from the presentation cuts to the core: technology changes quickly, balance sheets do not.
The combination is what makes the circular pattern worth watching. Heavy borrowing funds spending that looks like demand. When the loop slows or breaks, the debt remains and the revenue recedes.
The July 29 test
Rickards keeps returning to one date. Around July 29, the largest AI companies report quarterly earnings. Those reports contain the data that resolves the circular spending question.
Buried in the filings are customer concentration disclosures, revenue breakdowns by segment, and capital expenditure commentary. An investor can see how much revenue comes from a small number of customers. They can check whether those customers are also investment partners. They can compare CapEx guidance against free cash flow generation. The line items that matter to someone who has read through a cycle ending before: the related-party transaction footnote, usually buried forty pages into a 10-Q; the customer-concentration table, where companies name any buyer accounting for 10% or more of revenue; and the gap between capital expenditure guidance and free cash flow. When a company says it will spend $40 billion on infrastructure and generate $8 billion in free cash flow, the $32 billion gap has to come from somewhere. In 2000, Cisco’s filings showed vendor financing receivables growing faster than product revenue. The revenue looked like growth. The receivables were the real story. The same pattern — revenue rising while the quality of that revenue quietly deteriorates — is what an operator watches for in these filings. The numbers tell the truth. They bury it in the footnotes and wait for you to find it.
If the reports show broad-based customer growth and revenue diversifying beyond the hyperscaler circle, the circular spending concern recedes. If customer concentration is high and a significant portion of revenue traces back to companies that are also investors or investment targets, the loop Rickards describes becomes visible in the numbers.
The historical parallel
Rickards draws a line to the dotcom era, a parallel the AI bubble debate tracks across the analysts weighing in on both sides. In 2000, Cisco warned it could not see demand ahead. The Nasdaq’s unwind accelerated from there. The parallel Rickards draws is about revenue quality. During the dotcom bubble, a significant share of telecom equipment revenue came from companies buying gear to build infrastructure for other companies that were also buying gear. The spending fed itself until it stopped. Cisco’s vendor financing operation extended credit to telecom companies that used the credit to buy Cisco routers. By 2000, vendor financing and lease deals accounted for roughly 10% of Cisco’s $20 billion in annual revenue. When the telecoms collapsed, Cisco wrote off nearly $900 million in bad loans. The equipment was real and the demand was circular, so revenue looked real until you followed the money in circles.
The AI pattern has structural similarities. The same companies are funding each other’s spending while reporting that spending as revenue. The 2000 comparison is an argument that the same accounting dynamics are in play, and that the same kind of earnings report can expose them.
What changes if the loop is real
If the circular spending thesis holds, several things follow. AI revenue growth rates overstate genuine customer demand. The debt raised against that revenue carries more risk than the credit ratings suggest. And the stocks priced for sustained high growth face a repricing when the loop slows, because the revenue base is smaller than reported figures implied.
None of this means AI fails. Rickards says the technology is real. Whether the financial architecture around it is as solid as the market believes is the open question. July 29 is the first date where the answer starts to come into focus. The earnings reports will show who is actually paying, and how much of the money comes from outside the circle. Rickards applies the same financial-architecture reading across his campaigns. More campaign breakdowns in the Promo Watch.