Find this research useful? Add aibubblequestion.com to your Google Preferred Sources, one click, affects only your own results.

The AI Bubble Question
All pages

Weighing It

The Case Against the Bubble

Not a token counterpoint. If you have read the rest of this site and find the bubble case persuasive, this page is the strongest evidence that you, and it, might be wrong.

Updated

1. Who is spending: fortress balance sheets, not burning startups

The 1999 boom was financed by unprofitable startups spending other people's money; when funding stopped, they died. Today's build-out is led by Microsoft, Alphabet, Amazon, and Meta, companies generating hundreds of billions of dollars of operating cash flow a year and funding most of their capital spendingMoney a company spends on long-lived physical things: buildings, machines, chips, data centers. Spent now, paid back (hopefully) over years.Full definition in the glossary from it. BofA forecast in December 2025 that combined operating cash flow at five profitable hyperscalers would grow from $378 billion (2023) to $577 billion (2025). FY2025 is now a completed year for all five; reading their own 10-Ks (Microsoft, Alphabet, Meta, Amazon and Oracle, each on its own fiscal-year-end) gives an actual combined operating cash flow of $577.0 billion for FY2025, $136.16B, $164.71B, $115.80B, $139.51B and $20.82B respectively, landing almost exactly on the forecast. The near-exact match is not a coincidence to be suspicious of: BofA published its forecast in December 2025, and by then Microsoft's FY2025 (ended 30 June 2025) and Oracle's FY2025 (ended 31 May 2025) had already closed and been filed, so two of the five inputs were already-known numbers, not forecasts, when the forecast was made. This is the site's own arithmetic summing five named filings, not a BofA figure; the debt-to-cash ratio claim was not independently recomputed this pass and is dropped rather than restated on the forecast's authority alone (Microsoft, Form 10-K, FY2025). A company that can absorb a failed bet without touching solvency is a categorically different risk than one that dies when the music stops. Structurally, this looks less like the South Sea Company and more like profitable industrials building railways or electric grids, booms that hurt investors but built economies.

2. The revenue is real, and compounding violently

Pets.com had no revenue. Microsoft's AI business runs at $37 billion annualized, up 123% year over year (GeekWire); OpenAI went from nearly nothing in 2022 to $13 billion of 2025 revenue and a run rateTake the most recent month's revenue and multiply by 12. It shows how fast a company is growing right now, but it is a projection, not money already earned.Full definition in the glossary that has since surpassed $40 billion as of mid-August 2026, up from about $25 billion in early 2026, reported via internal company communications, not an audited figure; enterprise spending on generative AI tripled in a year (AI Index). Meanwhile the cost of using AIRunning a trained AI model to answer questions or generate text. Training builds the model once; inference is the ongoing cost of using it, every single time.Full definition in the glossary falls roughly 10x per year for fixed capability (LLM Token Price Index), and technologies whose price collapses that fast, chips, storage, bandwidth, have historically grown demand by far more than the price fell. Adoption curves look like the early web and smartphones, not like VR or 3D TV.

3. The general-purpose-technology argument

If AI is a general-purpose technology on the order of electricity, then front-loaded overinvestment is not waste; it is how economies restructure. Goldman's capex model reaches $1.6 trillion a year by 2031 in its baseline scenario, explicitly labeled "a scenario-based framework... not a forecast of future spending," sensitive enough to assumptions that they "could shift figures by hundreds of billions of dollars" (Goldman Sachs, "Tracking Trillions"). The oft-paired GDP figure is a different document from a different year: a 2023 Goldman research note put the level effect at "7% (or almost $7 trillion)" of global GDP growth cumulative over a 10-year period, not $7 trillion added per year, and not the same analysis as the 2031 capex model (Goldman Sachs, "Generative AI could raise global GDP by 7%"). Against numbers like those, even trillions of cumulative capex earn respectable returns. A systematic review of the bubble literature concludes the most likely path is overvaluation, correction, and consolidation around a real technology, not a wipeout (InnovaPath review). A bull-case voice entered this same window: Goldman Sachs Research's Joseph Briggs modeled global AI-related investment at roughly $1 trillion in 2026, including $581 billion in the US, a derived estimate, not a disclosed figure, calling it "consistent with the 2%-5% of GDP peak investment impulses observed in prior general-purpose technology buildouts" (Goldman Sachs Research, Joseph Briggs, "Global AI Investment Is Forecast to Exceed $1 Trillion in 2026"). The same note flags the widely cited $794 billion hyperscaler-capex figure as itself mismeasured, by roughly $200 billion in each direction, globally and for the US, Goldman's own reconciliation, not an audited one. This is a bank's model, not a filed fact, and it is weighed here on the same terms as every other estimate on this page. Note what this argument does not claim: that prices are right. It claims the spending is rational at the system level, which was also true of the railways, for everyone except the shareholders.

4. The rebuttals on circular financing

Defenders of the deals mapped on The AI Money Loop make three points. First, scale and transparency: Nvidia's stakes and Microsoft's OpenAI position are disclosed and generally small beside their cash flows, structured as equity plus offtakes rather than the vendor loans of 1999-2001. The comparison point itself needs a correction: a widely repeated "123% of vendors' pretax earnings" loan-book figure very likely describes five North American telecom-equipment vendors collectively in 1999, not Lucent alone (CNET); a same-era, separately reported figure for Nortel specifically ("as much as 130% of equipment cost") shows the era's vendor-financing numbers were reported per-company and varied. The origin article for "123%" was not independently re-reached this pass. Whether today's AI equity stakes are proportionately "small" is itself an assertion, not a measured ratio on either side of this comparison. Second, economic substance: these are strategic stakes securing access to models and capacity, closer to Apple pre-paying key suppliers than to loans propping up customers with no business model; OpenAI and Anthropic have billions in real revenue, where Lucent's financed customers often had none. Third, a different failure mode: equity stakes and offtake agreementsA long-term contract to buy a project's future output, signed before it's built. Lenders rely on these promises, so the buyer's health becomes the project's foundation.Full definition in the glossary are marked to market and renegotiated in daylight, not carried as mispriced "safe" vendor loansA supplier lending customers the money to buy its own products. Inflates the supplier's sales until the customers can't pay.Full definition in the glossary that explode all at once (Reuters). If AI disappoints, on this view, the giants take write-downs and move on; the system doesn't unravel.

5. The track record of bubble calls: early is the norm

This is the part the bubble side most needs to sit with. Prominent, credible, and ultimately correct bubble warnings have consistently run years early:

  • Greenspan, December 1996: the "irrational exuberance" speech. The Nasdaq then gained roughly another 300% before the March 2000 peak (DQYDJ). An investor who exited on the warning missed more upside than the crash took back from an index holder.
  • Housing, 2002-2004: Housing warnings didn't arrive as one signal. Case and Shiller's academic diagnosis in 2002-03 preceded three to four more years of rising prices. Shiller's own plainest warning on the record is much later and much closer to the top: in June 2005, under a year before the actual peak, he called it "a bubble of unprecedented proportions" that would "end ugly" (Shiller, Brookings, June 2005). Treating "2002-04 warnings" as one uniform three-to-four-years-early signal overstates how early Shiller's own strongest warning actually was.
  • The "everything bubble," 2013-2019: near-continuous warnings that zero rates had inflated everything; US equities compounded strongly for the rest of the decade, and the eventual 2020 break came from a pandemic, not valuations.

A specific numeric base rate this page previously carried, that manias persist 3 to 5 years after expert warnings begin and another 12 to 24 months after warnings become consensus, could not be located anywhere in its cited source (Quinn and Turner's *Boom and Bust*) across two independent search passes, including three secondary treatments of the book and a broad open search; it is dropped rather than restated on an author-name citation alone. What holds without it: "it's a bubble" and "get out now" are different claims, and exiting years early has historically cost real money. Dot-com worries were everywhere by 1998; subprime was in the papers by 2005. The opportunity cost is not a footnote: for a diversified, unleveraged investor, exiting 2-3 years early has often destroyed more return than riding the drawdownHow far a price falls from its peak to its lowest point, in percent. The dot-com crash was a 78% drawdown for the Nasdaq.Full definition in the glossary would have. "It's a bubble" and "get out now" are different claims, and history is brutal to people who confuse them.

New sceptical voices entered this same window too, and this page's counterarguments have to hold up against them, not only against the critic profiled in full on The Critics. Michael Burry cited a real, independently retrievable paper, "Intelligence per Watt," to argue frontier GPU demand is overstated relative to what small local models can already handle; Jim Chanos put the moment "closer to a '99-type moment than a '97" in a 31 July 2026 appearance. Neither is a filed fact; both are named, dated positions, tracked in full at that link so the two pages stay in balance.

What would settle it

The strength of this case is that it is falsifiable. The research specifies what 12-24 months of evidence would look like on each side; those criteria are laid out in full on Conclusions, attached to the scenarios they would confirm. In brief: sustained AI revenue scaling toward $200-300 billion a year with healthy margins, positive enterprise ROI at scale, and capex settling below cash flow would falsify the bubble thesis; sector credit events, vendor write-downs of customer stakes, and collapsing used-GPU prices would confirm it.