The AI Bubble Debate Should Be About Debt, Not Demos

The argument over whether AI is a bubble has been conducted mostly in the wrong currency. Commentators trade anecdotes about chatbot adoption and failed pilots, while the sharper warnings this month are about something else: how the build-out is being paid for. Ray Dalio's latest remarks and a recent legal-policy report both point to the same place, the debt and the financial engineering, and that is where the debate should move. This is an analysis piece; the figures below are attributed to the sources that reported them, and several are projections or one person's estimates.

Two warnings, one mechanism

Bridgewater founder Ray Dalio told a Forbes conference in Singapore on Wednesday that surging AI borrowing, rising interest rates and pressure to turn paper wealth into cash are pushing markets closer to a burst, according to a Benzinga report republished by Yahoo Finance. He called AI a "classic bubble" and, in a Bloomberg Television interview, compared the boom to the late 1920s, saying such bubbles "always come together" with major innovation. His specific claim is about sequencing: the pop begins with debt and rates, and "we're in the part of the cycle that is before that but approaching that."

A month earlier, the Regulatory Review published a summary by Alex O'Donnell of a report by Vanderbilt Law School's Asad Ramzanali, "After the AI Crash." It reaches a similar view from a different direction, cataloguing structural risks: circular equity financing among AI, chip and cloud companies, off-balance-sheet borrowing, escalating data-centre subsidies and vertical integration of models, chips and facilities. Notably, neither source argues that the technology is fake. Their concern is that valuations and financing may have run ahead of earnings.

What the numbers do and do not show

The Benzinga piece cites several figures. It attributes to Jamie Dimon an estimate that hyperscaler-ecosystem spending could rise from about $700 billion this year to $1 trillion next year, and to Goldman Sachs an expectation that the five largest hyperscalers will issue about $250 billion in bonds this year and $400 billion in 2027. It also cites iShares data that Amazon, Microsoft, Alphabet, Meta and Oracle issued roughly $200 billion of investment-grade debt in the first half of 2026, almost twice their full-year 2025 issuance. The article says Nvidia, Apple and Microsoft make up more than 21% of the S&P 500, citing Creative Planning data, and that the S&P 500 and Nasdaq hit fresh highs on Tuesday.

The Regulatory Review summary reports Ramzanali's figures: industry investment of about $1.5 trillion in 2025, projected at about $2.5 trillion in 2026, and data-centre costs of $50 billion to $60 billion per gigawatt. It adds that major developers are operating at a loss, with one OpenAI projection of profitability no earlier than 2030, and it cites a 2025 MIT report finding that 95% of generative AI projects fail to produce a return on investment. These are the report's numbers as summarised; we have not verified them against the originals, and the MIT statistic in particular is a contested, narrow measure of pilot projects rather than of the whole market.

None of these figures proves a bubble. Large spending alongside record equity prices is also what a genuine, durable technology shift can look like. What they show is that the financing profile has changed: more borrowing, in bigger amounts, backed by revenue that, per the report, is still some years from covering the outlay.

Why the financing question matters more than the adoption question

Adoption can be real and the financing still fragile

The internet was transformative and the dot-com crash still happened. Dalio's point, as reported, is that the two go together. A technology can change work and still be financed on assumptions that fail. That is why anecdotes about useful products cannot settle the matter, and why sceptics who cite failed pilots miss it too. The relevant test is whether cash flows arrive before the debt comes due.

Cost of capital is the trigger

Dalio's reasoning is explicitly about rates and liquidity: large debts become dangerous when borrowing costs rise, and unrealised gains are hard to spend, so holders must sell assets to raise cash. He argued that wealth taxes and other efforts to cash out gains could help trigger a decline, and said "the bubble usually pricks at that." That is his view, not an established mechanism, and the article offers no rebuttal. Still, it identifies observable variables: bond issuance, spreads, and the pace of new capacity commitments.

Opaque structures hide the exposure

The Ramzanali report's focus on circular financing and off-balance-sheet debt is the part with the most practical bite. If chip makers invest in customers who then buy their chips, or if borrowing sits outside corporate balance sheets, outsiders cannot easily tell how much revenue is organic and how much leverage is hidden. Disclosure, not prohibition, is the modest first step that follows from the concern.

Where the bubble case is weakest

Intellectual honesty requires noting what the warnings leave out. The Benzinga article presents no counterarguments to Dalio, and it pairs his comments with a bearish quote from Michael Burry rather than a dissent. Both warnings come from people with a long record of cautioning about excess, and prediction of timing is notoriously poor: Burry is quoted saying the market is in "denial" and that such a stage lasts six to nine months, a timeline that is itself a guess. Equity markets at record highs are weak evidence for either side. The strongest rebuttal to the bubble thesis would be evidence of rapidly rising, durable revenue, and the sources we reviewed do not supply it either way.

Ramzanali's policy prescriptions are also contestable. They include restricting circular financing, separating AI software developers from physical data infrastructure in the manner of Glass-Steagall, and creating a new federal regulator. Reasonable people will disagree on whether those remedies fit the problem. The diagnosis, though, that nobody outside the largest firms can clearly see the leverage, deserves a hearing independent of the cure.

What a better debate would look like

• Track the debt, not the demos. Bond issuance, maturities and interest cover for the largest spenders tell more about fragility than product launches do.
• Demand disclosure on circular deals. Investors, regulators and journalists should be able to see how much customer revenue is funded by the supplier.
• Separate the technology verdict from the financing verdict. It is coherent to believe AI will matter for decades and that today's capital structure is too stretched.
• Avoid timing claims. Warnings about months-long phases are speculation; the conditions they describe are what can be measured.

The honest summary is uncomfortable for both camps. The boosters cannot cite the technology's promise as proof that the financing is sound, and the pessimists cannot cite the financing as proof that the technology is a mirage. The question that matters is narrower and more answerable: who is holding the debt, and what happens to them if the revenue arrives later than planned?