“Circular investing” is one of the hottest debates in AI over the past two years: chipmakers and cloud giants invest in OpenAI and Anthropic, and the model companies spend the money back on their investors’ chips and clouds. Some call it “left hand to right hand”; others call it a sign of a bubble.
This piece lays out the money web, sums up both sides, and gives you a yardstick for reading it. To see who is actually making money, pair it with the AI Stock Money-Flow Map; to see whether this counts as a bubble, pair it with Is AI a Bubble?.
What Is Circular Investing? In Plain Terms
Circular investing (circular financing) means one company invests in another, and the company receiving the money uses it to buy the investor’s products or services. The money loops back near where it started. The industry also calls it “vendor financing.”
An analogy: a coffee machine maker puts up money to help a coffee shop open, and the shop uses that money to buy coffee machines from the same maker. The business is real, but outsiders will ask: how much of this demand was pushed out by the money?
Core-Data Snapshot
Official announcements and media reports carry very different weight; the “Timing / Nature” column marks which is which.
| Deal | Details | Timing / Nature |
|---|---|---|
| NVIDIA → OpenAI | $30 billion investment; OpenAI commits to 3GW of inference and 2GW of training compute on Vera Rubin systems | 2026-02 official announcement |
| NVIDIA × OpenAI | Conditional guarantee for the Ohio data center, up to $105 billion; on default, NVIDIA covers an agreed shortfall left after subleasing or selling the capacity | 2026-08 official (SEC filing) |
| AMD × OpenAI | Purchase of 6GW of AMD GPUs; OpenAI can earn warrants for up to 160 million AMD shares on milestones | 2025-10 official |
| Amazon → OpenAI | Commits to invest $50 billion (according to Bloomberg, $35 billion of it depends on an IPO or reaching AGI) | 2026-02 official |
| OpenAI → Oracle | About $300 billion of cloud compute over roughly five years | Media (WSJ/Reuters) reports, not officially confirmed |
| OpenAI → CoreWeave | Compute contracts totaling about $22.4 billion | 2025-09 official |
| OpenAI → Broadcom | Co-develop 10GW of custom AI accelerators | 2025-10 official (amount not disclosed) |
| NVIDIA → CoreWeave | Additional $2 billion equity investment | 2026-01 official |
| Amazon → Anthropic | $13 billion invested, plus up to $20 billion more tied to milestones; Anthropic commits to spend over $100 billion on AWS over ten years | 2026-04 official |
| Google → Anthropic | Up to $40 billion ($10 billion up front, $30 billion tied to performance); Google Cloud provides 5GW from 2027 | 2026-04, according to multiple media reports, confirmed by both sides |
| NVIDIA, Microsoft → Anthropic | Up to $10 billion and $5 billion respectively; Anthropic commits to buy $30 billion of Azure compute | 2025-11 official |
What This Money Web Looks Like
The deals in the table fall into three groups: NVIDIA taking a stake in OpenAI, while AMD gives OpenAI warrants; cloud giants investing in model companies and then collecting cloud orders (Amazon, Google, Microsoft); and model companies placing orders with compute providers like Oracle and CoreWeave.
For companies that rent out GPUs, read neocloud compute rental; for the physical hardware supply chain, read The AI Hardware Supply Chain, End to End.
Why Some People Worry
When a company both invests in a customer and collects money from that customer, outsiders find it hard to tell how much of the demand its own investment created. This is called “round-tripping.” If one model company can’t raise its next round, the whole chain could loosen. NVIDIA guaranteeing OpenAI’s lease goes a step further: the supplier not only puts in money but also takes on the risk that the customer can’t pay rent.
The common analogy is the dot-com bubble of the late 1990s: Lucent and Nortel lent money to telecom startups to buy their own equipment, many startups failed, and the equipment makers crashed with them. In September 2026, the Managing Director of the International Monetary Fund (IMF) also warned that if profits or investment plans disappoint, leverage, circular financing, and global linkages could turn that disappointment into a broad correction.
Why Others Say It’s Fine
The other side says this is normal in a capital-intensive industry: strategic investment plus advance purchasing. NVIDIA CEO Jensen Huang has publicly pushed back on the “circular” framing, saying the investments are about securing capacity and helping customers grow the market.
Airlines pre-order planes and power plants sign long-term power purchase agreements; AI infrastructure locks in capacity with equity plus long contracts, which is similar. And the investors are mostly cash-rich big companies, not the debt-laden startups of that era.
How to Read This Web
This piece only describes the industry’s money structure. It does not list beneficiary stocks or rank individual stocks, and it is not investment advice. Check these four points:
- Equity investment or tied purchasing. NVIDIA’s stake in OpenAI and how much NVIDIA compute OpenAI commits to use are two different things.
- Letter of intent or signed contract. NVIDIA’s September 2025 letter of intent to invest “up to $100 billion” in OpenAI was never finalized; in 2026 it became a $30 billion equity stake.
- Officially confirmed or media report. The roughly $300 billion figure for OpenAI’s contract with Oracle comes from media reports.
- Does revenue keep up with commitments. Annualized revenue (ARR) projects current revenue over a full year. According to the FT, OpenAI was at about $50 billion at the end of September 2026 (by its own accounting); Anthropic officially reported above $47 billion in May 2026, and according to Bloomberg, above $65 billion at the end of July. The two count revenue resold through cloud partners differently, so they can’t be compared directly. The other thing to watch is whether revenue can support the compute commitments.
Penchan’s summary
Whether circular investing is normal pre-purchasing or inflated demand, both sides have a case.
Rather than rushing to call a bubble, it’s more useful to sort each deal into “equity vs. purchasing,” “intent vs. contract,” and “official vs. media,” then check whether revenue keeps up with the commitments.