The AI bubble: everyone is lending money to each other

The circle of AI money

Here's the simplest version of what's happening. Microsoft writes OpenAI a check for $13 billion. OpenAI takes that money and spends a huge chunk of it on Microsoft Azure — Microsoft's own cloud platform. Microsoft then reports this as cloud revenue growth. Wall Street cheers. Microsoft's stock goes up. Microsoft uses its higher valuation to invest more in AI.

That's not a supply chain. That's a revolving door with a cash register.

And it's not just Microsoft. The pattern repeats across the entire AI ecosystem.

Nvidia: the arms dealer who funds both sides

Nvidia is worth over $5 trillion — the most valuable company in the world as of mid-2026. That valuation is built on one thing: selling GPUs to AI companies. H100s, B200s, GB200s — the gold rush hardware that every AI lab, cloud provider, and enterprise is racing to buy.

But here's the twist. Nvidia doesn't just sell picks and shovels. It also invests in the miners. Nvidia Ventures has backed CoreWeave, Recursion, Cohere, Inflection, and dozens more AI startups. Those companies then spend their funding on — you guessed it — Nvidia GPUs.

The question nobody wants to ask: how much of that $115B in GPU revenue is funded by venture capital and corporate investment — rather than actual end-user AI demand? If the customers buying GPUs are burning through investor money, Nvidia's "revenue" is really just a downstream echo of speculative capital.

The Anthropic triangle

Anthropic — the company behind Claude — is funded by both Amazon ($8B) and Google ($3B+). Anthropic runs its infrastructure on both AWS and Google Cloud. So Amazon and Google are effectively funding a company that pays them back in cloud bills.

It gets better. Amazon offers Anthropic's models through AWS Bedrock. Google offers them through Vertex AI. So the investment isn't just about equity — it's about locking in a marquee AI brand as a feature of their own cloud platform. The investment is the customer acquisition cost.

Stargate: the $500B bet

In January 2025, SoftBank, Oracle, and OpenAI announced "Stargate" — a $500 billion AI infrastructure project. The initial $100B commitment was the largest single AI infrastructure deal ever announced.

But look at who's involved. SoftBank — the company behind the WeWork debacle and the Vision Fund's spectacular losses — is the lead investor. Oracle provides cloud infrastructure. OpenAI provides the AI models. The money flows from SoftBank → to building data centers → which run on Nvidia chips → which host OpenAI models → which Microsoft also licenses.

Everyone touches the money. Everyone books revenue. And the ultimate customer — the business that actually pays for an AI product that solves a real problem? Still mostly theoretical at $500B scale.

AMD: the underdog's gamble

AMD sees the gold rush and wants in. They've invested heavily in their MI300X GPU line to compete with Nvidia. They've acquired Xilinx for $49B to strengthen their AI hardware portfolio. They're courting every cloud provider, every AI lab, every enterprise that's tired of Nvidia's pricing power.

The play makes sense — if AI demand is real and growing. But if a significant chunk of that demand is funded by circular investment rather than end-user revenue, AMD is building a $49B bridge to a market that might be smaller than everyone thinks.

September 2026 update: the loop got bigger

Six months after this article was first published, the circle hasn't broken — it's widened. Three developments from the last two weeks of August 2026, each from a primary source:

Nvidia now says it's a quarter of its own demand. On its August 26 earnings call (revenue: $96.2 billion for the quarter), CEO Jensen Huang said: "We've invested nearly $50 billion in the Frontier AI labs. This was a meaningful commitment, but it represented a small fraction of our expected free cash flow." CFO Colette Kress went further: the portion of demand supported by Nvidia's own balance sheet "is expected to contribute toward roughly a quarter of our business next year." She also put OpenAI's existing and planned commitments at approximately 12 gigawatts of Nvidia compute through 2030, and said six infrastructure capital partners "will raise over $500 billion of third-party capital." Read that again: the chipmaker is financing a quarter of next year's chip sales.

Anthropic's compute deals route back to Nvidia. On August 31, Anthropic signed a reported $35 billion cloud agreement with Lambda, a cloud provider Nvidia has invested in, for a Texas data center developed by Hut 8 — with Nvidia itself holding the lease on the site. It followed a separate, reported $45 billion capacity deal the week before. So Nvidia is the chip supplier, an equity backer of the cloud provider, and the anchor tenant on the building — while Amazon and Google, Anthropic's largest investors, also sell Anthropic's models through their own clouds.

The exit is now on the calendar. On August 19, OpenAI CFO Sarah Friar told employees the company "will be a public company in 2027," or sooner if "our business continues to inflect." Her framing: "The IPO is not a finish line, it is a milestone, another fundraise. We raised $122 billion in March, and that gives us flexibility." She added that Anthropic is also confidentially on file and there is "a chance they pull the cover off that confidential file in the coming weeks and become public in September." An IPO doesn't resolve the question this article asks — whether end customers pay enough to justify the buildout. It transfers that question to public shareholders.

Is this actually a bubble?

The honest answer: it depends on what happens next. Circular capital flows aren't inherently fraudulent. They're how ecosystems bootstrap. Amazon lost money for years while building the infrastructure that now prints cash. The question is whether AI revenue eventually comes from real customers solving real problems — or whether it stays a loop of tech companies paying each other.

For every dollar spent building AI, only about 9 cents comes back as actual product revenue. The rest is funded by investment, stock dilution, and corporate balance sheets betting on future demand. That gap has to close — or the music stops.

The dot-com parallel

In 1999, dot-com companies bought ads on each other's websites. Company A's revenue was Company B's marketing spend, and vice versa. Revenue went up for everyone. Then it didn't.

The AI version is more sophisticated. Instead of banner ads, it's cloud credits, equity investments, and GPU purchases. But the underlying pattern is the same: companies counting each other's spending as revenue. The loop works until someone asks where the outside money comes from.

The internet bubble ended, and the internet turned out to be real. The overvalued companies died, but the infrastructure survived and became the backbone of the modern economy. AI might follow the same path — real technology, fake valuations, painful correction, eventual transformation.

None of this means AI isn't transformative. It means the financial layer on top of AI — the valuations, the investment rounds, the revenue claims — might be running ahead of reality. The technology is impressive. The accounting is creative. And the money? It's going in circles.