Is the AI Boom Just Companies Paying Each Other? Inside the Circular Deal Loop

Trace any dollar in the AI economy long enough and it tends to boomerang back to where it started. Nvidia invests in OpenAI. OpenAI commits hundreds of billions to Microsoft cloud and Nvidia chips. Anthropic pledges tens of billions to Azure. Microsoft and Oracle sell compute to the very frontier labs they're bankrolling. At each handoff the headline numbers grow larger, the press releases more triumphant, and the underlying circuit tightens. The question keeping a small but vocal cohort of analysts up at night isn't whether AI will transform the economy. It's whether the current boom is, in meaningful part, a closed loop of companies paying each other with money they largely got from each other.

Drawing the Loop: How the Money Actually Moves

The circular-deal mechanics are not secret. Every leg of the loop has been disclosed, reported on, or acknowledged in earnings calls. What's been underappreciated is how tightly the legs connect and what a stumble in any single node does to the whole structure.

Start with the chip maker at the center. Nvidia has pledged up to $100 billion into OpenAI and approximately $10 billion into Anthropic, according to Bloomberg and Global Finance reporting. That capital flows into frontier model training, which means it flows almost immediately back out as compute spending. OpenAI has committed roughly $250 billion to Microsoft cloud and tens of billions more in chip deals with AMD and Nvidia directly. Anthropic, for its part, has committed approximately $30 billion to scale Claude on Azure. Azure is Microsoft's cloud. Microsoft is one of OpenAI's largest investors. The loop is not metaphorical: it is contractual.

The Loop Visualised: Nvidia → OpenAI → Microsoft/Oracle/Nvidia → Anthropic → Azure

Walk it step by step. Nvidia sells GPUs and invests equity into OpenAI and Anthropic. OpenAI uses that capital (plus Microsoft equity financing and bond proceeds) to buy Nvidia and AMD chips and to pay Microsoft for Azure compute at a multi-hundred-billion-dollar scale. Anthropic spends its Google and Amazon investment dollars on Azure compute (Microsoft), on AWS (Amazon), and on Nvidia hardware. Microsoft's Azure revenue flows back to Microsoft shareholders, whose largest single non-index holding in many portfolios is Nvidia. Oracle sits adjacent, signing its own multi-billion compute agreements with OpenAI. CoreWeave, the hyperscaler spun out of an Nvidia-adjacent GPU-mining operation, holds long-term GPU leases funded by private credit and charges the frontier labs a premium for "on-demand" access to hardware they helped finance into existence.

At every node, the same silicon changes hands or the same dollars pay invoices across related parties. That isn't inherently fraudulent: integrated supply chains exist in semiconductors, pharmaceuticals, and aerospace. But it does mean that the headline revenue and investment figures in each company's disclosures are not purely independent demand signals. Some meaningful fraction is intra-ecosystem recirculation.

Bull Case vs. Bear Case: Two Serious Readings of the Same Data

The circular nature of the deals is not disputed by either bulls or bears. What's disputed is what it means. Both sides have coherent, internally consistent arguments, and both deserve a fair hearing before you anchor to a conclusion.

Michael Burry's Cisco Thesis — and Why It Deserves Attention

The most publicised bear trade against this structure belongs to Michael Burry, who holds Nvidia put options covering approximately 1 million shares at a $110 strike, expiring 2027, alongside short positions in Palantir, Oracle, the SOXX semiconductor ETF, and the QQQ. He has drawn an explicit parallel: Nvidia is "a Cisco at the center of it all," per Fortune and TradingKey reporting.

The Cisco analogy is worth unpacking precisely. Cisco didn't make a bad product. Its routers and switches were genuinely essential infrastructure for the internet. The error was that Cisco's revenue was inflated by vendor financing (extending credit to buyers who used it to purchase Cisco gear) and by the circular demand created when startups raised venture money, spent it on internet buildout equipment, and reported those purchases as "enterprise demand." When the venture spigot turned off, the downstream equipment orders collapsed. Cisco's peak-to-trough decline was 89 percent. Twenty-four years later it had not returned to its dot-com peak.

Burry's argument is not that AI is fake or that Nvidia's GPUs are like Pets.com. It is the more precise claim that Nvidia's current revenue multiple is priced as if every dollar of current AI capex represents durable end-customer demand, when some unknown fraction of it is circular financing that will compress when the funding round carousel slows. The put structure (long-dated, out-of-the-money) reflects this: he is not calling an imminent crash but pricing optionality on a multi-year mean-reversion.

The Cascade-Risk Chain: How a Single Disappointment Propagates

What makes the circular structure particularly fragile is that its nodes are not just financially linked: they are informationally linked. Each company's reported metrics feed the next company's fundraising narrative. Analyst framing via Global Finance has sketched the cascade chain explicitly:

Notice that nowhere in that cascade does something need to go catastrophically wrong. A single earnings miss at step one is sufficient to initiate steps two through five. The architecture is not brittle in the way a Ponzi scheme is brittle: it doesn't require intentional fraud to unwind messily. It only requires demand to come in below the assumptions baked into the commitment schedule.

Private Credit: The Hidden Leg That Nobody Is Watching

The most underappreciated systemic feature of the current AI-infrastructure boom is how much of it is financed not by public equity markets (which are visible, priced daily, and scrutinised by regulators) but by private credit funds. These funds have poured capital into GPU-backed infrastructure lenders and compute-cloud platforms at scale. The loans are collateralised against GPU hardware that depreciates rapidly, against long-term customer contracts that have varying degrees of enforceability, and against utilisation projections that are sensitive to the same demand signals that drive everything else in the loop.

Private credit does not mark to market daily. Stress in private credit portfolios can be hidden for quarters behind lagged valuation methodology, only surfacing when a covenant is tripped or a refinancing attempt fails. This creates an information asymmetry: by the time private credit fragility shows up in observable signals, the downstream effects on compute pricing, hyperscaler capex, and chip demand may already be in motion.

How This Post Sits Alongside Our Other Analysis

The circular-deal mechanics analysed here are distinct from (but connected to) the valuation and market-psychology dynamics covered in our companion piece OpenAI's Valuation and the AI Bubble Psychology. That post covers how private market valuations are set, what historical bubble patterns look like from a behavioural finance perspective, and whether OpenAI's $300B-plus markup reflects fundamental value or narrative momentum. This post, by contrast, is specifically about the structural mechanics of money flow between the nodes: the commitments, the counterparty relationships, the cascade risks, and the private-credit leg that sits beneath the public-market headlines. They are companion reads: one explains the psychology and valuation math; this one explains the plumbing.

For context on the most visible shock this loop has already absorbed, see our earlier analysis of DeepSeek R1 and Nvidia's single-day market cap loss. That event (where a Chinese lab published a model that appeared to achieve GPT-4-class performance at a fraction of the training cost) was the first public demonstration that the compute-intensity assumption underpinning the entire loop is not a physical constant. If frontier capability can be achieved with dramatically less hardware, the utilisation projections embedded in every private-credit facility and multi-year cloud commitment become subject to revision.

What "Circular" Does Not Mean

It's worth being precise about what the circular-deal critique is not saying, because the argument is sometimes caricatured into a claim it doesn't make.

The loop does not mean AI is useless. GPT-4, Claude 3.5, and Gemini Ultra are genuinely capable systems with documented productivity applications. Code generation, legal document review, drug discovery screening, customer service deflection: these are real, measurable productivity improvements, not marketing copy. The question is not "does AI work?" but "does the current spend level reflect the pace at which those productivity improvements translate into enterprise revenue for the infrastructure providers?"

The loop also does not mean all the headline numbers are fictitious. The commitments are real legal agreements. The compute is real silicon. The revenue is real cash. What the circular structure implies is that the demand signals are partially endogenous: some of the "demand" for Nvidia chips is downstream of Nvidia's own investments, not an independent validation from the end-user market. In a valuation framework that uses forward demand as a key input, that distinction matters.

Finally, circularity does not mean the boom ends badly. The bull case scenario (where real enterprise demand grows into the supply, the compute commitments prove prescient rather than excessive, and the private-credit vehicles refinance successfully) is entirely plausible. If Microsoft's AI revenue grows 40 percent annually for three years on the back of Copilot and enterprise model deployments, the loop becomes a feature: a fast-moving, tightly integrated supply chain that correctly anticipated demand. Cisco is Cisco. But Amazon in 2001 also looked like a dot-com cautionary tale, and it turned out to be something else entirely.

How Builders Should Navigate the Uncertainty

Whether the circular-deal structure resolves bullishly or bearishly, the practical implications for software teams and product companies are similar: the macro financing environment around AI infrastructure is more uncertain than the product capability curve, and building a business that depends on specific infrastructure pricing staying stable is a fragile position.

Inference costs have been falling and will likely continue to fall, but the pace of that fall is not guaranteed, and temporary supply-side shocks (a slowdown in new data-center buildout, a tightening of GPU availability due to private-credit restructuring) could create pricing volatility that destroys the unit economics of margin-thin AI applications. The teams building durable businesses are designing for a range of infrastructure cost scenarios rather than assuming the current cost trajectory continues uninterrupted.

They are also building applications where the value is in the workflow integration, the proprietary data layer, and the user experience, not in raw model access that any competitor can replicate by swapping API providers. In a world where the circular loop eventually tightens, the companies that survive are the ones whose moat does not depend on the next funding round flowing through the system.

Conclusion: The Loop Is Real — Its Resolution Is Not Yet Known

The AI circular-deal loop is not a conspiracy theory or a fringe short-seller narrative. It is a structural feature of the current investment landscape that is documented in earnings calls, regulatory filings, and Bloomberg reporting. Nvidia funds OpenAI and Anthropic. They spend the money on Microsoft, Oracle, and Nvidia. Microsoft and Oracle report AI revenue. Private credit funds the hardware beneath it all. Each node validates the next node's valuation narrative.

The cascade risk (Microsoft disappoints, Azure cuts, Nvidia misses, CoreWeave wobbles, OpenAI fundraising tightens) is not hypothetical. It is the structural consequence of building a tightly coupled financing system in which every participant's forward revenue depends on every other participant continuing to invest and spend at the current pace. Michael Burry's Cisco parallel is the most succinct expression of this risk: not "AI is fraud" but "the infrastructure at the center may be priced assuming demand that is partly endogenous to the financing rather than purely exogenous from end customers."

The 2026-specific dimension (escalating private-credit fragility, the DeepSeek compute-efficiency signal, and the first enterprise ROI renewal cycles) means this is not the same conversation it was in 2024. The numbers are bigger, the commitments are longer-dated, and the gaps in observable verification are wider. That is precisely why it deserves careful, sober analysis rather than either reflexive bullish dismissal or dramatic bubble proclamation.

The most defensible position, for investors and builders alike, is to hold the uncertainty without forcing a resolution that the data doesn't yet support. The loop exists. Whether it tightens into a virtuous supply chain or unwinds into a cascade is a question the next two to three years of enterprise AI adoption data will answer. Plan accordingly.