The AI Trade Has Started Financing Itself, and That Is the Warning Investors Should Not Ignore

We should say plainly what the last two weeks of trading have been trying to tell us: the artificial intelligence rally has begun to finance itself, and a market that funds its own customers has stopped proving that demand exists and started manufacturing the appearance of it. That is the thesis, and everything that follows is an attempt to make it uncomfortable to dismiss. The boom may yet be vindicated by the technology. But the structure now sustaining it is one that markets have learned, repeatedly and expensively, to distrust.

Circular Financing, Dressed as Growth

Consider the arrangement that defines this cycle. In September, Nvidia and OpenAI announced a strategic partnership under which OpenAI would build and deploy at least ten gigawatts of Nvidia systems, and Nvidia said it intended to invest up to one hundred billion dollars in OpenAI, disbursed progressively as each gigawatt was deployed. Read that sequence again. The supplier of the chips is providing the capital that its customer will use, in substantial part, to buy the supplier's chips. Nvidia's own newsroom described the first gigawatt as arriving in the second half of 2026 on its Vera Rubin platform.

Bloomberg has mapped the wider pattern, showing how cloud-computing companies and chipmakers, Nvidia foremost among them, have helped fund the leading AI developers, which in turn became some of their largest customers. The result, in Bloomberg's phrase, is an increasingly interconnected web of dependencies. The polite term for this is circular financing. The honest term is that revenue and investment have begun to chase each other around a closed track, and from the outside it can be difficult to tell genuine demand from an accounting echo.

None of this is illegal, and little of it is even novel in isolation. Vendor financing is an old instrument. What unsettles is the scale and the concentration. When a single dominant supplier underwrites the balance sheets of the very firms whose orders validate its valuation, the ordinary discipline of the market, the willingness of an independent buyer to pay with independent money, is precisely the discipline that goes missing.

Figures Beneath the Enthusiasm

The bullish story is not fantasy, and it deserves a fair hearing. The first half of 2026 was extraordinary for equities, with the S&P 500 up roughly 9.6 percent and the Nasdaq up more than 12 percent, and semiconductor stocks surged more than 80 percent over that stretch, according to market coverage reported by CNBC. Real products exist. Real revenue is booked. The demand for computation is not imaginary.

Yet the same coverage recorded the market's own second thoughts. In the holiday-shortened week ending July 3, the Philadelphia Semiconductor Index fell 6.3 percent on Wednesday and 5.4 percent on Thursday as investors took profits and questioned whether AI optimism had pushed valuations beyond reasonable levels. A rally that can shed more than eleven percent from its most emblematic sector across two sessions, on no fundamental catastrophe, is a rally whose participants are not certain what their holdings are worth. That uncertainty is the tell.

The valuation math invites the doubt. Nvidia has traded at roughly forty-one times forward adjusted earnings, a substantial premium to most technology peers, which means the price already embeds years of flawless execution. At that multiple there is no margin for a customer that trims its capital budget, no cushion for a deployment that slips, no forgiveness for demand that merely meets expectations rather than exceeding them.

Accounting Question That Lingers

Then there is the quieter argument, the one made not with slogans but with depreciation schedules. Michael Burry, the investor known for his wager against mortgage securities before the 2008 crisis, has contended that the large cloud operators are depreciating their Nvidia processors over five to six years when the economic life of the hardware is closer to two or three. By his arithmetic, reported by CNBC and elsewhere, that mismatch understates depreciation and overstates profit by roughly one hundred seventy-six billion dollars across 2026 to 2028, with the profits of certain firms potentially overstated by more than twenty percent by the end of that window.

One need not accept every figure to grant the force of the point. If the useful life of the chips is shorter than the books assume, then reported earnings across the sector are flattering, and the multiples investors are paying rest on a foundation that will erode faster than the spreadsheets admit. This is not a moral accusation so much as a structural warning: the accounting that makes the boom look profitable may be the accounting most exposed to revision.

Washington Enters the Cap Table

The most revealing development of early July was not a price move at all. According to the Financial Times, reported by Bloomberg and CNBC, OpenAI has held preliminary, conceptual discussions about giving the United States government a five percent stake, with the notion floated that Washington might hold five percent of each leading American AI developer through a sovereign-fund vehicle modeled loosely on the Alaska Permanent Fund.

Whatever its industrial logic, an invitation of that kind is not the behavior of an industry confident in the sufficiency of private capital and private demand. It is the behavior of an industry seeking to widen the circle of stakeholders who cannot afford to let it fail. We have seen this instinct before in sectors that grew too large and too interlinked to be allowed to correct on their own terms.

A market that must recruit its suppliers, its customers, and eventually its government into the same financing loop is not demonstrating strength. It is distributing the consequences of a possible mistake.

The distinction matters because it changes what a correction would mean. Deeper links between AI firms and credit markets, and denser interconnections among the firms themselves, imply that a fall in valuations would not stay contained to equity holders. The Bank of England's Financial Policy Committee has named corrections to AI-related equity valuations as one of its three principal sources of financial-stability risk, alongside sovereign and private-credit markets, and has judged US equity valuations close to the most stretched they have been since the dot-com era.

Caution Has Its Own Cost

Intellectual honesty requires stating the strongest rebuttal. The internet was a genuine revolution wrapped inside a genuine bubble, and the bubble's collapse did not falsify the technology; it merely repriced it. Perhaps AI is the same, and perhaps the winners on the far side of any correction will justify prices that look reckless today. Perhaps the circular deals are simply how capital-intensive frontiers get built when the incumbents have the balance sheets and the newcomers have the demand.

We take the point, and we do not counsel panic. But the lesson of prior manias is not that the underlying technology disappoints. It is that the financial architecture built around the technology fails first, and fails hardest for those who mistook the architecture for the technology. The railways were real; many of the railway companies were not. Fiber optics were real; a great many of the firms that laid them went to zero.

• The demand-proving mechanism is compromised when suppliers finance their own buyers.
• The reported profitability is contingent on depreciation assumptions that a credible skeptic has already challenged with numbers.
• The valuations leave no room for ordinary disappointment, only for perfection.
• The push to socialize the risk, from vendor financing to a mooted government stake, suggests the participants themselves sense the fragility.

Discipline, Defined for This Moment

Our conclusion is not that the AI era is a fraud or that the machines will fail to deliver. It is narrower and, we think, harder to wave away. The present rally has substituted the circulation of capital for the verification of demand, and until an independent buyer pays with independent money at these prices, the market has not actually been tested. Investors who wish to participate should do so with their eyes open to that substitution, sizing positions for the possibility that the loop breaks before the technology matures.

The productive question is no longer whether artificial intelligence will change the economy. It plainly will. The question is whether the financing structure erected on that conviction can survive its own first serious disappointment. On the evidence of a market that funds its customers, books optimistic depreciation, and now contemplates handing equity to the state, the prudent answer is to assume it cannot, and to invest accordingly.