One of the hottest debates in the AI world over the past year or two is “circular investing.” NVIDIA announces it wants to invest in OpenAI, and OpenAI turns around and signs astronomical compute deals with NVIDIA, Oracle, and Broadcom; Amazon and Google invest in Anthropic, and Anthropic spends the money right back on their clouds. The money seems to circle among the same handful of companies, so some shout “left hand to right hand” and others shout “this is a bubble in the making.”
This piece lays the money web out flat. First, what “circular investing” actually is and what these deals look like; then why the bulls and bears are at each other’s throats; and finally a yardstick for reading it all. This is one piece in the “AI Industry Watch” series; to follow up on who is actually making money, pair it with the AI Stock Money-Flow Map, and to see whether this really counts as a bubble, pair it with Is AI a Bubble?.
What Is Circular Investing? In Plain Terms
Circular investing (circular financing) means this: one company invests in another, and the one being invested in turns around and uses the money to buy the investor’s products or services. The money loops around and lands back near where it started.
In this round of AI, the typical script goes like this: a chipmaker or cloud operator puts money into a model company (such as OpenAI or Anthropic), and once the model company has the funds, it signs big orders to turn around and buy the investor’s chips and rent its cloud. So the line between “investment” and “revenue” gets a little blurry.
The industry has a more technical term for it, “vendor financing”: the supplier puts up the money to help the customer afford its own goods. Here’s an analogy: a maker of coffee machines first lends money to a coffee shop so it can open, and the coffee shop then uses that money to buy the coffee machines from the same maker. The business is real, but an outsider will ask one more question: how much of this demand comes from genuine appetite, and how much was pushed out by the money?
Core-Data Snapshot
Below we organize the most-discussed deals into a single table. First, let’s be clear about one thing: some figures are official company announcements, while others are only media reports or market estimates, and they carry very different weight. The “Timing / Nature” column flags which is which.
| Deal | Details | Timing / Nature |
|---|---|---|
| NVIDIA → OpenAI | LOI: deploy at least 10GW of NVIDIA systems; NVIDIA plans to invest up to $100 billion in phases | 2025-09 official LOI (not finalized cash) |
| AMD → OpenAI | Deploy 6GW of AMD GPUs; grant warrants for up to 160 million AMD shares, unlocking on milestones | 2025-10 official |
| NVIDIA → CoreWeave | Equity investment of about $2 billion | Official (separate from the GPU-collateralized loan) |
| Amazon → Anthropic | $8 billion invested to date; another $5 billion in 2026, with up to $20 billion more in future; Anthropic commits to over $100 billion on AWS over ten years, up to 5GW | 2026-04 official |
| Google → Anthropic | Expanded TPU/cloud partnership, in the “tens of billions of dollars” range, up to a million TPUs, over 1GW in 2026 | 2026 official (no single dollar figure specified) |
| OpenAI → CoreWeave | Compute contracts totaling about $22.4 billion | 2025-09 CoreWeave announcement |
| OpenAI → Oracle | About $300 billion of cloud compute over roughly five years | Media (WSJ/Reuters) reports, not officially named |
| OpenAI → Broadcom | Co-develop 10GW of custom AI accelerators | 2025-10 official (amount not disclosed) |
| ASML → Mistral | Leads a €1.3 billion round, about 11% stake | 2025-09 official |
| Stargate | OpenAI/SoftBank/Oracle/MGX, declaring $500 billion of infrastructure over four years | 2025 official (figure is a stated intent) |
What This Money Web Looks Like
It’s easiest to understand if you split the main deals into three groups by “who is investing in whom.”
Group one: chipmakers investing in model companies. NVIDIA and OpenAI signed a letter of intent (LOI, not yet a finalized contract) in September 2025, pointing toward OpenAI deploying at least 10GW of NVIDIA systems, with NVIDIA planning to invest up to $100 billion in phases. AMD is also partnering with OpenAI: OpenAI plans to deploy 6GW of AMD GPUs, and AMD has granted OpenAI warrants for up to 160 million shares (a warrant is a right to buy stock at an agreed price in the future), unlocking on deployment milestones. Separately, NVIDIA has also invested about $2 billion of equity in compute-rental provider CoreWeave.
Group two: cloud giants investing in model companies, and the model companies buying the cloud back. Amazon’s cumulative investment in Anthropic has reached $8 billion, and in April 2026 it announced another $5 billion, with up to $20 billion more in future; Anthropic, in turn, commits to spending over $100 billion on AWS over the next ten years, using up to 5GW of compute. Google has also expanded its partnership with Anthropic, in the “tens of billions of dollars” range, with up to a million TPUs on the table. Microsoft and OpenAI updated their partnership framework in April 2026, with Microsoft remaining OpenAI’s primary cloud partner.
Group three: the model companies’ enormous compute orders. OpenAI has the biggest appetite for compute: its contracts with CoreWeave total about $22.4 billion; media report that it signed a roughly $300 billion, five-year cloud deal with Oracle (this figure comes from Wall Street Journal and Reuters reporting, not an officially named announcement from either side); and it is also working with Broadcom to design its own chips, planning a 10GW deployment. On top of that, ASML, the sole supplier of lithography (chip-imaging) machines, crossed over to lead a €1.3 billion round, about 11% stake, in the European model company Mistral.
Connect all of this and you get a web of “chipmakers, equipment makers, clouds, and model companies investing in each other and buying from each other.” Add the Stargate infrastructure plan, formed by OpenAI, SoftBank, Oracle, and MGX and declaring $500 billion over four years, and it’s no wonder outsiders find it dizzying.
Why Some People Worry
The worriers fix on one core problem: when a company is both “investing in a customer” and “collecting money from that customer,” an outsider finds it very hard to tell whether the demand is real or was pushed out by the company’s own investment.
They call this structure “round-tripping”: the investor’s money flows to the model company, which uses that money to turn around and place orders, making the investor’s revenue and backlog look great. If any one link breaks, say a model company can’t raise its next round and can’t keep burning cash, the whole chain could loosen at once.
The most common analogy is the dot-com bubble of the late 1990s. Back then, telecom equipment makers like Lucent and Nortel lent money to newly formed telecom startups to buy their own gear; demand looked booming, and after a swath of those startups went under, the equipment makers crashed with them. Some investment-bank reports have also listed “the rising circularity of AI deals” as one reason bubble concerns have resurfaced.
Why Others Say It’s Fine
The other side argues that this is perfectly normal in a capital-intensive industry, a mix of strategic investment and advance purchasing.
NVIDIA CEO Jensen Huang has publicly pushed back on the negative “circular” framing, his point being roughly: the return on an investment and the revenue from selling a product are two different things; these investments are mainly about securing capacity, aligning technology roadmaps, and supporting promising customers so the market grows bigger. Supporters also offer examples: airlines pre-order planes, power plants sign long-term power-purchase agreements, and AI infrastructure locks in capacity with equity investment plus long contracts, which is similar in nature. Besides, the cloud giants’ customer base is broad, so even if some AI startups have taken on too much money, enterprise and consumer demand still has a chance to fill that capacity over the long run.
This camp’s main point is that the contracts are mostly multi-year, “up to” commitments backed by real compute demand, not figures conjured out of thin air.
How to Read This Web
This piece doesn’t predict whether the bubble will pop, and it offers no buy or sell advice. But it can give you a few observation points to check against yourself:
- Separate “equity investment” from “tied procurement.” NVIDIA’s equity investment in OpenAI (still a letter of intent for now) and how many GPUs OpenAI commits to buy are two things of different natures; lumping them together makes it easy to scare yourself or get carried away.
- Check whether a figure is “officially confirmed” or a “media estimate.” NVIDIA’s investment in OpenAI reads as “a letter of intent, plans to invest up to,” not cash that has been booked; that $300 billion with Oracle is a media report, not an officially named contract. They carry very different weight.
- Check whether it’s a letter of intent or a formal contract. A letter of intent (LOI) signals direction; it doesn’t equal a final amount in black and white.
- Check whether the model companies’ revenue can support the commitments. OpenAI’s and Anthropic’s annualized revenue (media estimates put each at roughly $25 billion and $30 billion, and neither is an auditor-verified figure) is growing fast, but compared with their compute commitments, which run into the hundreds of billions, the key things to watch are how large the gap is and what fills it.
Understanding the “structure” of this web is more practical than rushing to judge “whether it’s a bubble.” To follow up on which of these companies is actually making money and which is burning it, read the AI Stock Money-Flow Map; for the full bubble-judgment framework, read Is AI a Bubble?.
Key Takeaways
Circular investing is about this: chipmakers and cloud operators invest in model companies like OpenAI and Anthropic, and the model companies spend the money right back on those investors’ chips and clouds, forming a money loop.
Bulls call it the normal strategic investment and pre-purchasing of a capital-intensive industry; bears say it can make demand look hotter than it really is, like the vendor financing of the dot-com era. Both views have a point. The more pragmatic approach is to keep each deal’s “equity vs. procurement,” “official vs. media,” and “intent vs. contract” clearly apart, and then judge for yourself.
This article only describes the industry’s money structure; it does not compile beneficiary stocks, does not rank individual stocks, and does not constitute investment advice.
To see who is actually making money on this chain, read the AI Stock Money-Flow Map; to see whether this counts as a bubble, read Is AI a Bubble?; to understand the new clouds that rent out GPUs, read neocloud compute rental; and to see the physical hardware supply chain, head back to The AI Hardware Supply Chain, End to End.