Is AI a bubble? The Nvidia-OpenAI circular revenue map

Nvidia invests ten billion in OpenAI. OpenAI announces a hundred-billion-dollar chip purchase from Nvidia. Nvidia's revenue lifts on the chip purchase. Nvidia's stock lifts on the revenue. The lifted stock funds the next round of cross-inve

Nvidia invests ten billion in OpenAI. OpenAI announces a hundred-billion-dollar chip purchase from Nvidia. Nvidia's revenue lifts on the chip purchase. Nvidia's stock lifts on the revenue. The lifted stock funds the next round of cross-investment. If you have been watching this and wondering whether it is a real growth story or financial engineering with extra steps, you are asking the right question.

The pattern has a name. It is called AI circular revenue, and the AI industry runs on a more concentrated version of it than the headline numbers suggest — Anthropic, AMD, Oracle, CoreWeave, the hyperscalers, all wired into the same ring. Hundreds of billions of dollars cycle between roughly a dozen companies, and that ring is doing a non-trivial fraction of the work currently driving US economic growth.

AI circular revenue is real vendor financing on a real infrastructure build, AND bubble-shaped capital structure with bubble-shaped warning signs — which leg of that statement matters more depends on which question you are asking.

The deal structure, in plain English

A circular deal is one where a supplier takes an equity stake in a customer, the customer commits to buying the supplier's product, and the resulting revenue feeds back into the supplier's valuation, which funds the next round of cross-investment.

In AI, the canonical example: Nvidia commits an equity investment into OpenAI; OpenAI commits a multi-year purchase of Nvidia chips; the chip purchase shows up as revenue on Nvidia's books; Nvidia's stock prices in the revenue; the higher market cap makes the next equity round trivial. The same shape applies between Nvidia and Anthropic, between Microsoft and OpenAI, between Amazon and Anthropic, between AMD and Oracle, between hyperscalers and basically every model lab.

Two things to internalise about this shape. It can be vendor financing — when the seller of an expensive thing lends the buyer capital to buy it, accelerating demand that would otherwise wait on the buyer's own cash flow. Cisco and Lucent did this through the late-1990s telecom build-out. It is also legally and structurally round-tripping — when revenue is created by colluding parties cycling the same capital, with no underlying demand to justify the transaction.

The mechanism is identical. The judgment is downstream of whether the demand is real.

Why this looks like vendor financing — the bull case

The bull argument is that this is a build-out, not a scam.

Real product demand exists. ChatGPT and Claude have hundreds of millions of monthly users. Enterprise AI contracts are signed and being expanded. The inference layer — the part that actually runs the models for users — has real load, growing.

The hyperscalers funding the bulk of the capex (Microsoft, Amazon, Google, Meta) have substantial standalone cash flow. Their AI spending is mostly being financed off their own balance sheets, not by a circular ring. If their AI bets don't pay off perfectly, they survive. The datacenters they are building are physical assets with multi-decade economic lives.

Vendor financing accelerates a build-out when capital is the bottleneck. That is what is happening here on the model-lab side. If OpenAI and Anthropic eventually generate enough inference revenue to service their own capex, the loop closes legitimately, the way good vendor financing always does. The capacity gets built faster than the buyers' organic cash flow would have allowed; the supplier carries the risk of the spread; the spread closes when the demand catches up.

For a fuller treatment of where AI's real applications and limits actually sit, the practical version is more boring than the bull narrative — but the underlying demand is real enough that the build-out is not pure imagination.

Why this looks like a bubble — the bear case

The bear argument is that the capital structure has bubble-shaped pathologies and the demand verification is missing.

No frontier AI company has demonstrated profitability at scale. OpenAI, Anthropic, xAI, Mistral — all running on continued external capital, all spending more on training and inference than they earn from selling access. The economics of inference at frontier-model scale are not yet proven.

The widely-cited MIT NANDA study found roughly 95% of enterprise AI pilots produced no measurable revenue impact in their first wave of deployment. The enterprise demand side of the picture — the side that is supposed to absorb the projected three trillion dollars of datacentre spend through 2028 — is not showing up at the speed the capex narrative requires.

The structure compounds the risk. When a dozen companies are interlinked through cross-investment, vendor financing, and shared customer concentration, "diversification" within the ring is illusory. Correlation inside the loop approaches one. If any major node fails to deliver on its commitments, the failure does not stay contained.

This is the pattern that ended the late-1990s telecom build. Vendor financing held until demand didn't catch up; then Lucent's receivables went bad, the telecom carriers couldn't service the gear, and the fiber overcapacity sat dark for a decade. The mechanism is well documented and not new.

The dotcom comparison — what's the same, what's different

The capital structure parallels are strong.

Same. Vendor financing as the lubricant of a capacity build. Aggressive cross-investment among a tight cluster of players. Headline revenue growing on commitments that are funded by the supplier itself. Equity valuations running on the slope of the revenue line, not on free cash flow. A handful of companies driving a disproportionate share of the headline equity-market return.

Different. The underlying product. In 1999, most internet companies had business models that did not yet have customers paying real money at the scale the valuations implied. In 2026, frontier LLM products have millions of paying retail users and substantial enterprise contracts — the demand side is not zero. It is just not yet at the scale that justifies the capex curve.

Different. The infrastructure layer's economic life. Fiber-optic gear had a long useful life but was technically displaceable; the bandwidth glut took a decade to absorb. GPUs depreciate harder (each generation faster than the last), but the buildings, power infrastructure, and cooling stay useful regardless of which silicon goes inside. The infrastructure survives a model-lab-equity correction better than fiber survived the telco correction.

Different. The buyers. The bulk of AI capex is being run through hyperscalers with hundreds of billions of standalone cash flow. The dotcom telcos were credit-fuelled startups carrying receivables they could not service. That is a real structural difference and it matters.

The honest synthesis: 2026 AI rhymes with 1999 dotcom on the capital structure layer and diverges from it on the demand-and-balance-sheet layer. Calling the AI boom "the next dotcom" is wrong by oversimplification. Calling it "nothing like dotcom" is wrong by denial. Both at once is closer to the truth.

The US growth story — concentrated and real, but fragile

US private fixed investment in 2025 and 2026 has been disproportionately driven by AI-adjacent capex: datacentres, chips, power infrastructure, real estate around hyperscaler campuses. Estimates of AI capex's contribution to recent US GDP growth land somewhere in the 1–2 percentage-point range of headline real growth, depending on which sub-categories you assign to the AI bucket.

That contribution is concentrated. A small set of companies. A small set of geographies — Texas, Virginia, the Pacific Northwest, parts of Arizona and Wyoming. A small set of capex categories.

If the equity layer of the circular ring corrects, the capex spigot does not turn off overnight. Hyperscaler self-financed spending continues — their cash flow generation does not depend on whether Nvidia's stock trades at fifty or two hundred times earnings. But the marginal layer — circular-financed model-lab capex, speculative third-party datacentre capacity, vendor-financed customer purchases — could compress quickly. The difference between the headline GDP print and the post-correction GDP print is the size of that marginal layer.

The pattern is not new. The 1999–2001 telecom build added to US GDP for several quarters and subtracted from it for several more. The 2005–2007 housing build did the same on a longer scale. Both episodes finished with infrastructure that retained long-term economic value and equity valuations that did not.

What probably survives if the equity layer corrects

The dotcom precedent is the right anchor for this question.

What survives. The physical infrastructure — datacentres, fibre, power generation, transmission upgrades, cooling capacity. The model capabilities — current frontier weights do not unlearn what they have learned. The user behaviour — consumer and enterprise AI usage is sticky and continues regardless of which specific labs go on. The labour reorganisation — workflows have already moved. The training methodology — what we know about scaling laws and post-training does not evaporate when a stock falls.

What corrects. Equity valuations on AI companies whose revenue is the loop. Equity valuations on chip vendors whose growth was loop-amplified. M&A and IPO multiples across the AI stack. A subset of vendor-financed deals get unwound; receivables get written down; some capacity sits idle for a stretch.

Cisco is the textbook example. The internet survived 2001 — usage continued to compound. Cisco's revenue, in absolute terms, recovered and grew past the 2000 peak within a few years. Cisco's stock did not regain the 2000 peak for over twenty years. The product layer and the equity layer can diverge by an order of magnitude in either direction.

The circular structure is not proof of fraud and is not proof of legitimacy. It is proof that the loop matters, and that any view of where AI capital flows resolve has to model what happens if the loop breaks.

— Internal note on AI capital structure, Tradoki desk

What this means for a retail trader trying to make sense of it

The framework, not a recommendation.

Do not conflate "AI is real" with "AI equities are fairly priced." Both can be true. Neither can be. The two statements are independent.

Do not conflate "this looks like dotcom" with "AI is fake." Pattern matching is useful for orientation. Pattern matching is not prediction. Every bubble has produced a "this rhymes with the last one" narrative, and so has every secular boom that did not pop.

The most useful diagnostic question is mechanical: if the circular layer collapsed tomorrow, which companies in the ring have positive standalone free cash flow without their loop counterparties? The list is short, the answer is informative, and the exercise sharpens any "diversified AI exposure" thesis quickly.

Concentration risk is the underrated leg. If you hold AI exposure across a "diversified" set of names that are all wired into the same circular ring, your correlation in a stress scenario is much higher than your portfolio's headline diversification suggests. This is the same lesson the future-of-retail piece makes about regulatory and macro concentration — applied to a different layer of the stack.

For the broader frame on how to think about AI as an industry over the next several years, Jensen Huang's March 2026 statement that Nvidia's recent equity bets on OpenAI and Anthropic would likely be its last is worth reading literally, then reading sideways. For the analytical view of why vendor financing in AI is not necessarily fraud, Noah Smith's piece on AI circular deals is the cleanest summary of the bull-side structural argument.

The trader who has internalised that the structure of the deal matters as much as the headline number is positioned differently than one who is reading "Nvidia revenue up X%, must mean Nvidia is the trade." The latter is not a strategy. It is momentum chase dressed up as a thesis.

● FAQ

What is AI circular revenue?

A pattern in which a supplier (Nvidia) takes an equity stake in a customer (OpenAI, Anthropic), the customer commits to buying the supplier's product, the supplier's revenue lifts on that committed purchase, the supplier's stock lifts on the revenue, and the lifted stock funds the next round of cross-investment. The capital cycles between a small set of related companies.

Is the AI boom a bubble like the dotcom bubble?

The capital-structure pattern rhymes — vendor financing, concentrated cross-investment, capacity narratives running ahead of cash-flow verification. The demand picture is different — current LLMs have real paying users and substantial enterprise contracts that most 1999 dotcoms did not. The honest read is that 2026 AI is rhyme, not repeat. Pattern matching is informative; it is not prediction.

How does Nvidia's investment in OpenAI actually work?

Reported deal: Nvidia commits a large equity investment into OpenAI; OpenAI commits a large multi-year purchase of Nvidia chips. The headline numbers are in the hundred-billion-dollar range. The mechanism is sometimes called vendor financing — a normal industry practice — and sometimes called round-tripping, which is the same mechanism in a fraudulent costume. Which it is depends on whether underlying demand justifies the capacity.

What happens to US economic growth if the AI bubble pops?

AI-adjacent capital expenditure has been a meaningful contributor to recent US private-investment growth. If circular-financed projects unwind, the marginal capex layer could vanish quickly. Hyperscaler self-funded spending continues regardless. The difference between current headline growth and post-correction growth is the marginal layer the circular structure is funding, which by some estimates is in the 1–2 percentage-point range of US GDP.

What is vendor financing?

When the seller of an expensive thing also provides the buyer with the capital to buy it. Cisco, Lucent, and Nortel did this in the late-1990s telecom build-out. It accelerates customer purchases when capital is the bottleneck. It works when underlying demand catches up to the capacity. It fails when demand does not, leaving the seller holding bad receivables and the buyer with assets it cannot service.