The AI bubble bursts when the debt comes due, not when the hype ends

Ask when the AI bubble will burst and you are really asking three separate questions at once. The first is whether current AI valuations exceed what future cash flows will justify. The second is what event or mechanism would force a repricing. The third is a date. The first question has a defensible answer in mid-2026. The second has a shortlist of plausible candidates. The third does not exist, and anyone selling you a date is selling you their positioning, not their analysis.

The question everyone asks and the answer markets refuse to give

The evidence for the first question has hardened considerably over the past twelve months. Worldwide AI spending is forecast to reach roughly $2.5 trillion in 2026, a 44 percent jump over the prior year, according to Gartner. The five largest hyperscalers alone plan capital expenditures of roughly $725 billion in 2026, up from about $410 billion in 2025. Set against that, the revenue actually generated by AI products remains a fraction of the investment. OpenAI, the largest pure-play AI company by revenue, runs at roughly $25 billion annualized while carrying compute commitments that reach into the hundreds of billions of dollars. The gap between what is being spent and what is being earned is the entire bubble debate compressed into one arithmetic problem.

The second question — what pops it — has moved from theory to live observation. June 2026 gave markets a preview. The Magnificent Seven lost roughly $2.3 trillion in combined market value in a single month, with Microsoft down about 20 percent, as investors began demanding proof that AI capex converts into revenue on a timeline shorter than a decade. That was not a burst. Semiconductor suppliers rallied through the same period, the Philadelphia Semiconductor Index gained 6 percent in June, and the S&P 500 did not enter a bear market. It was a repricing of one leg of the trade while another leg strengthened, which tells you the market has not made up its mind. It has only started asking harder questions.

The third question, the date, deserves honesty rather than a headline. Prediction markets currently price roughly a 16 percent probability that a recognized burst arrives by the end of 2026, which means the same traders assign an 84 percent chance it does not. Economists at the Bank for International Settlements compare today’s buildout to railway manias and the dot-com era without committing to timing, because timing is the one thing those precedents never repeated. The dot-com bubble was called early by credible people for three full years before March 2000. Alan Greenspan’s “irrational exuberance” speech landed in December 1996; the Nasdaq more than tripled afterward before it broke. Being right about the diagnosis and wrong about the date bankrupted more short sellers than being wrong about the diagnosis ever did.

This article takes the question seriously on all three levels. It works through the spending data, the debt structures now financing the buildout, the depreciation accounting that skeptics call the soft underbelly of the boom, the productivity evidence from the first serious academic studies, and the historical base rates from every comparable infrastructure mania since the 1840s. It also treats the counterargument with the respect the numbers demand, because the companies funding this buildout are the most profitable enterprises in the history of capitalism, and that fact alone distinguishes 2026 from 1999 in ways that matter for both timing and severity.

The framing that survives contact with all of this evidence is simple to state. The AI boom has a financing timetable, and the financing timetable — not sentiment, not model releases, not viral skepticism — is what sets the window for a burst. Debt service begins on specific dates. Compute contracts come due on specific dates. Depreciation charges hit income statements on specific quarters. Earnings seasons force disclosure four times a year. Each of these creates a scheduled collision between promises and cash, and each collision is a candidate trigger. The rest of this analysis maps those collisions, weighs the scenarios, and ends with the most honest timing answer the evidence supports — which is a probability distribution across 2026 through 2028, not a date, and a shape of deflation that may look less like an explosion and more like a long, grinding argument between capex and cash flow that individual companies win or lose one earnings call at a time.

One more framing note before the numbers. A bubble bursting and a technology failing are different events, and conflating them is the most common analytical error in this debate. The internet did not fail in 2000. Railways did not fail in 1847. The technology can be real, world-changing, and still wildly overcapitalized at a moment in time. Both things were true then. The evidence below suggests both things are plausibly true now.

A working definition of an AI bubble worth arguing about

The word bubble gets used loosely enough that the debate often collapses into people talking past each other. A useful definition has to be specific enough to be falsifiable. The one used throughout this analysis comes closest to the framing the Bank for International Settlements applied in its 2026 annual report: a bubble exists when a genuine technological breakthrough attracts capital in excess of what commercial returns can, in the end, justify, and it bursts when the reversal of that investment becomes self-reinforcing — falling asset prices forcing spending cuts, which validate the pessimism, which forces further cuts.

Under that definition, several things commonly cited as bubble evidence are not, on their own, decisive. High valuations are not a bubble; they are a forecast. Nvidia trading at a premium multiple is a claim about future demand for accelerated computing, and claims about the future can be aggressive without being irrational. Heavy capital spending is not a bubble either; every general-purpose technology in history required infrastructure built ahead of demand. Even hype is not a bubble. Hype is a constant of technology markets and has accompanied genuine revolutions and genuine frauds with equal enthusiasm.

The definition turns on the relationship between invested capital and achievable returns, which is why the honest analytical work is unglamorous. It means comparing capex schedules to revenue trajectories, checking how the spending is financed, and asking what happens to the financing if the revenue arrives two or three years late. Late is the operative word. Most bubbles are not stories about demand that never materialized. They are stories about demand that materialized on a slower schedule than the capital structure could survive. The telecoms that died in 2001 and 2002 built fiber that carried the video internet a decade later. They were right about demand and dead anyway, because debt does not wait for the future to arrive.

Applied to AI in mid-2026, the definition produces a split verdict, and the split is where all the interesting analysis lives. On the equity side, the companies doing the heaviest spending — Microsoft, Alphabet, Amazon, Meta — earn combined net income in the hundreds of billions of dollars annually and funded the early phase of the buildout from operating cash flow. That is categorically unlike the dot-com pattern of revenue-free companies burning equity raises. On the financing side, however, the picture changed materially during 2025 and 2026. AI-related companies and projects tapped debt markets for at least $200 billion in 2025, hyperscaler bond issuance ran to roughly $121 billion in that year alone — more than four times the five-year average — and private credit exposure to AI infrastructure went from near zero to over $200 billion in a few years. The moment debt entered the structure at scale, the boom acquired the one feature that turns corrections into cascades: fixed obligations that do not flex when revenue disappoints.

A second component of the definition deserves emphasis: the burst is a process, not a moment. Popular imagination compresses the dot-com crash into a single day, but the Nasdaq peaked on March 10, 2000 and did not bottom until October 2002, an 83 percent decline spread across thirty-one months with multiple violent rallies inside it. People who lived through it describe a slow bleed punctuated by false dawns, not an explosion. If the AI trade deflates, the base-rate expectation from every historical precedent is the same shape: an initial sharp repricing, a partial recovery that convinces many the worst is over, and then a longer grind as the fundamental gap between investment and returns closes from both directions — spending falling, revenue slowly rising — until the two lines meet.

That definition also clarifies what would falsify the bubble thesis entirely. If AI revenue across the hyperscalers and model companies compounds fast enough that the 2026 capex cohort earns an acceptable return within the depreciation life of the hardware, there was no bubble, only an unusually front-loaded investment cycle. Concretely: if by 2028 the hyperscalers can attribute several hundred billion dollars of annual revenue directly to AI services at healthy margins, the 2026 spending will look prescient. The evidence on whether that trajectory is plausible — enterprise adoption data, consumer willingness to pay, productivity measurements — occupies the middle sections of this article, and it is genuinely mixed rather than damning. The bubble question is live precisely because the revenue race is still runnable. The financing question is whether the runners get to finish.

The numbers behind the 2026 boom

The scale of the current buildout has outgrown every adjective, so the honest move is to let the figures stand on their own and then explain what each one implies.

Global AI spending in 2026 is projected at approximately $2.52 trillion, a 44 percent increase year over year, per Gartner, which further projects that AI will account for nearly all IT spending by 2030. The five largest hyperscalers — Microsoft, Amazon, Alphabet, Meta, and Oracle — are set to spend more than a trillion dollars on AI-related capex in 2026 by the BIS’s estimate, with the four largest alone guiding to roughly $725 billion in total capital expenditures, up 77 percent from about $410 billion in 2025. Microsoft has guided to $190 billion in calendar 2026 capex, Amazon to roughly $200 billion, Alphabet to between $175 billion and $185 billion, and Meta to as much as $140 billion. Goldman Sachs expects cumulative AI spending by the four largest hyperscalers to reach $5.3 trillion by fiscal 2030, and estimates of total global data-center capital expenditure through 2030 run to $6.7–7.0 trillion.

Those are the outflows. The inflows tell a different story. OpenAI, the revenue leader among model companies, reached a roughly $25 billion annualized run rate by mid-2026 — up from $21.4 billion at end-2025 and only $3.7 billion in 2024 — split roughly into $17 billion from ChatGPT subscriptions, $6.5 billion from API consumption, and about $1.5 billion from Sora and licensing. Anthropic’s revenue has grown fast enough to take substantial enterprise share, with Menlo Ventures putting its enterprise LLM share at 40 percent against OpenAI’s 27 percent by late 2025. Add the AI-attributable cloud revenue at the hyperscalers — real but not separately disclosed in most cases — and the most generous industry-wide estimate of direct annual AI revenue in 2026 sits somewhere in the low hundreds of billions of dollars against $2.5 trillion of annual spending. Even conceding that much of that spending builds durable capacity rather than covering annual costs, the ratio between money in and money out is the widest of any technology cycle on record.

The market capitalization built on top of that ratio is wider still. Nvidia crossed $5 trillion in market value on October 29, 2025, after a twelve-fold rise since ChatGPT’s launch in late 2022. The Magnificent Seven — even after June 2026’s decline — represent roughly 35 percent of the S&P 500, matching the concentration of the top seven names at the dot-com peak. US equity market capitalization stands at nearly twice US GDP, materially above where it stood in March 2000. The Shiller cyclically adjusted price-to-earnings ratio exceeded 40 during 2025, a level previously reached only in the run-up to the dot-com crash, and stood at 39.8 in December 2025 against a historical average of 17.7.

Key figures of the 2026 AI investment cycle

The table condenses the tension that runs through this whole analysis: record profitability and record spending coexisting with a revenue base that has not yet caught up, valuations at generational extremes, and a rapidly growing debt layer underneath. None of these figures alone proves a bubble. Together they establish that the margin for disappointment has become unusually thin.

One number in that table deserves to be pulled out and held up to the light, because it is the one that changed the debate in 2026: free cash flow for the five big hyperscalers is expected to fall roughly 91 percent in 2026 to about $16 billion, even as their combined net income rises about 25 percent to roughly $506 billion. Profits are booming on paper while cash is being consumed by the buildout almost in its entirety. That divergence is precisely the pattern that preceded trouble in prior infrastructure manias — reported earnings holding up while the cash conversion collapses — and it is why the market’s patience, not the technology’s promise, became the binding constraint in June.

June 2026 as the first real stress test of the AI trade

The month of June 2026 will be studied either as the beginning of the deflation or as the correction that reset the trade for another leg higher, and it is worth recording exactly what happened while the details are fresh, because the pattern of the selloff says a great deal about how a genuine burst would propagate.

Over the course of June, roughly $2.3 trillion was erased from the combined market value of the Magnificent Seven, with the CNBC Magnificent 7 Index falling 10 percent for the month and the group down more than 13 percent from mid-May. The losses were sharply uneven. Microsoft fell about 20 percent, Nvidia about 13 percent, Apple and Amazon roughly 8 percent each — while Alphabet ended June up roughly 12 percent year to date. Counting Broadcom and Oracle alongside the seven, the combined June loss reached approximately $2.7 trillion. Goldman Sachs Asset Management noted that dispersion among the Magnificent Seven had widened to 52.3 percent since the end of Q3 2025 and wrote that “the once-cohesive Magnificent 7 narrative is being reconsidered.”

The proximate cause was not a model failure, a demand collapse, or an external shock. It was arithmetic. First-quarter earnings had confirmed the $725 billion capex figure, Microsoft’s CFO Amy Hood pointed to rising memory and component costs pushing the number higher, and the free-cash-flow math above became impossible to ignore. As one widely circulated framing put it, the biggest AI names stopped trading on the promise of future revenue and started trading on the cost of actually delivering it. Meta’s July 9 announcement that it would begin manufacturing its own AI chip in September 2026 and expand computing capacity to 14 gigawatts in 2027 was sold on the news — investors read a demand signal as evidence that spending has no ceiling.

The most analytically important feature of June was the divergence between AI spenders and AI suppliers. While the hyperscalers sold off, the Philadelphia Semiconductor Index gained 6 percent in June and stood up roughly 90 percent for the year; the Roundhill Memory ETF had surged 166 percent in 2026. Micron jumped nearly 8 percent in a single session in early July. The market was not rejecting AI. It was repricing who captures the economics — paying suppliers who book revenue today while discounting buyers whose returns arrive later, if at all. Michael Burry read the same divergence in the opposite direction, disclosing on July 1 a short position in Micron at $1,051.87 after a near-700 percent one-year rally, framing the supplier rally itself as the late-stage phase of the mania: “fear of missing out, the greater fool theory, and public commitment bias.”

Two readings of June coexist, and both have serious backers. The bearish reading holds that June was the first leg of the deflation — that markets have begun the multi-quarter process of demanding cash returns the spenders cannot yet show, and that each earnings season from here ratchets the pressure. Nigel Green of deVere Group argued the selloff will not be the last reckoning and predicted the Magnificent Seven will effectively narrow to a “Magnificent Three” as markets separate companies that capture AI economics from companies that merely consume AI infrastructure. The bullish reading, voiced by Jim Cramer and Wedbush’s Dan Ives, holds that June was a “gut check” inside an intact cycle: Ives called it an “air pocket,” Cramer argued the entire supplier-over-spender trade reverses the moment one hyperscaler raises guidance on the strength of its AI products, and Barron’s noted that Meta exited June trading at roughly 16 times earnings, below the S&P 500 average of 21 — an odd valuation for a bubble’s epicenter.

The synthesis that fits the evidence: June demonstrated that the AI trade now has a working transmission mechanism for bad news. Capex guidance moves trillions of dollars of market value within weeks. That mechanism did not exist in 2023 or 2024, when higher spending announcements were rewarded. A market that punishes spending is a market that has started, however tentatively, to price the bubble scenario — and the July 2026 earnings season, underway as this is written, is the first full test of whether the punishment continues or reverses.

The January warning shots that markets chose to forget

June was not the first tremor. The AI trade has now produced two January shocks, one in 2025 and one in 2026, and both were absorbed and forgotten with a speed that itself belongs in the diagnostic file.

The January 2025 event was DeepSeek. The release of a Chinese model claiming frontier-adjacent performance at a fraction of the training cost wiped nearly $1 trillion from global technology market value within days, with Nvidia suffering the largest single-day dollar loss in stock market history at the time. The logic of the selloff was straightforward: if frontier capability could be replicated cheaply, the pricing power justifying hundreds of billions in GPU spending was in doubt. The recovery was equally instructive. Within months the market had reframed cheap models as demand-expanding rather than capex-destroying — the Jevons paradox argument, that efficiency gains increase total consumption — and the buildout accelerated rather than slowed. Whatever one thinks of that reframing, the episode proved that the AI trade could absorb a direct hit to its core thesis and reprice upward anyway. Bubbles historically display exactly this property in their late-middle phase: bad news gets metabolized into bullish narrative.

The January 2026 event was sharper and more domestic. In a single session during the January selloff, Microsoft alone lost $440 billion in market value — a one-day loss larger than the entire market capitalization of all but a few dozen companies on Earth — as investors reacted to the widening gap between AI capex guidance and cloud revenue growth. The broader tape stabilized within weeks, helped by strong Q4 earnings elsewhere and by the steady drumbeat of new model releases, but the episode established the scale of value that can evaporate when the market briefly applies bear-case assumptions to a single hyperscaler. Analysts at the time called it a preview, and June 2026 validated the label: the January single-day mechanism, extended across a month and applied to the whole complex, produced the $2.3 trillion drawdown.

The pattern across DeepSeek, January 2026, and June 2026 is one of shortening intervals and broadening scope. The first shock hit the chip layer, the second hit one spender, the third hit all the spenders at once. Each recovery has been narrower than the last — post-June, the recovery has been concentrated in suppliers while the hyperscalers remain marked down. Students of prior cycles will recognize the sequence. The dot-com market produced comparable warning shots in the summer of 1998 and October 1999, both fully recovered, both retrospectively obvious rehearsals. Markets rarely break on the first stress; they break after the pattern of stress-and-recovery has trained participants to buy every dip, so that when the dip finally does not recover, the positioning is maximally wrong.

None of this proves June 2026 was the final rehearsal rather than the opening act. It does establish something useful for the timing question: the AI trade’s shock absorbers are measurably weaker than they were eighteen months ago. Each recovery has required a stronger narrative to overcome weaker cash-flow facts, and the supply of narrative upgrades — new frontier models, sovereign AI deals, agentic-workflow promises — meets an audience that has now heard several rounds of them while the free-cash-flow line kept falling. When shock absorption fails entirely, the historical record says it fails quickly.

OpenAI’s balance sheet as the bubble’s central exhibit

Every bubble era selects one company whose finances become the referendum on the whole thesis. In 1999 it was arguably Cisco or Pets.com depending on which half of the argument you were making. In 2026 it is OpenAI, and the company’s disclosures ahead of its planned IPO have given the debate an unprecedented volume of hard numbers to fight over.

The growth case first, because it is genuine and extraordinary. OpenAI reached a roughly $25 billion annualized revenue run rate by mid-2026, up from $21.4 billion at end-2025 and $3.7 billion in 2024 — a near-sevenfold expansion in eighteen months, among the fastest revenue scalings in business history. ChatGPT serves on the order of 900 million weekly users. Enterprise and Team seats grew from about $1 billion annualized at the start of 2025 to more than $7 billion by mid-2026. The company filed a confidential S-1 with the SEC in mid-2026, with Goldman Sachs, Morgan Stanley, and JPMorgan leading a listing that Reuters reported could value the company above $1 trillion, and CFO Sarah Friar has signaled strong retail demand with plans to reserve shares for individual investors.

Now the other column. OpenAI’s Q1 2026 disclosures to shareholders showed $5.7 billion in quarterly revenue against $3.7 billion of quarterly cash burn — both roughly tripling year over year, which means the business is not yet showing economies of scale in cash terms. Financial Times reporting put full-year 2025 spending at $34 billion and the net loss at $38.5 billion, nearly eight times the 2024 loss of $5.1 billion. Internal projections reported by The Information anticipate roughly $14 billion in losses for 2026 on some measures, with projected burn of approximately $27 billion in 2026 and $63 billion in 2027 under the renegotiated Microsoft agreement, cumulative losses of up to $115 billion through 2029, and cash-flow positivity arriving no earlier than 2029 and more plausibly 2030. The company is, by the most cited figure, losing about $1.22 for every $1.00 it earns.

Against those losses stand the commitments. OpenAI has locked in roughly $600 billion in future data-center spending by the Wall Street Journal’s accounting — Q1 shareholder documents put long-term compute procurement commitments at $665 billion — accumulated through the Stargate joint venture with SoftBank, Oracle, and MGX (headline figure $500 billion, committed equity closer to $52 billion, the rest dependent on debt not yet raised), a $300 billion Oracle contract, a $100 billion Nvidia partnership, a $38 billion Amazon arrangement, a $22 billion CoreWeave deal, and a $250 billion Azure purchase commitment embedded in the restructured Microsoft relationship.

April 2026 is when these columns collided in public. The Wall Street Journal reported that OpenAI had missed internal targets — including the goal of one billion weekly ChatGPT users by end-2025 — and that Friar had privately warned company leaders that revenue might not grow fast enough to cover the compute contracts. Board directors were reported to be questioning why CEO Sam Altman kept pursuing additional capacity amid the slowdown. Shares of OpenAI’s publicly traded partners — Nvidia, Oracle, Microsoft, CoreWeave — fell on the news, a live demonstration that a private company’s internal targets can now move public markets. Meanwhile Anthropic overtook OpenAI on the Forge Global secondary platform, trading at an implied roughly $1 trillion against OpenAI’s approximately $880 billion, and Bridgewater’s Greg Jensen described OpenAI’s prospective IPO pricing as “priced for a monopoly outcome that does not yet exist.”

The reason OpenAI matters beyond itself is structural. Its compute commitments are other companies’ booked revenue expectations. Oracle’s market value, CoreWeave’s debt service, a sizable slice of Nvidia’s order book, and the economics of the Stargate build all assume OpenAI pays what it has promised. If revenue scales roughly six-fold by 2029 while compute costs scale four-fold, OpenAI plausibly becomes one of the most profitable companies ever created; if the ratios invert, a single counterparty failure propagates through the entire AI supply chain at once. The IPO, expected in late 2026, converts this private referendum into a public one with a daily price — and a disappointing debut or a broken first-year chart would do more to date the bubble’s top than any macro indicator, exactly as a handful of failed listings dated the top in the spring of 2000.

The compute commitment problem

Beneath the headline valuations sits a lattice of long-dated purchase obligations that has no precedent in technology history, and understanding it is necessary for any serious timing analysis, because these contracts are where the bubble’s abstract risk becomes a payment schedule.

The structure works like this. Model companies — OpenAI foremost, but also Anthropic, xAI, and others — sign multi-year commitments to purchase compute capacity from cloud providers and neoclouds. Those providers, in turn, sign their own commitments: to chipmakers for GPUs, to developers and REITs for data-center shells, to utilities for power, often with take-or-pay features and terms stretching five to fifteen years. Each layer borrows or raises against the layer above it. CoreWeave’s bonds are underwritten by its OpenAI and Microsoft contracts. Oracle’s capex is underwritten by its OpenAI backlog. Data-center developers’ construction loans are underwritten by hyperscaler leases. The BIS flagged the pattern in its 2026 annual report, warning that firms attempting to lock in future capacity through long-dated contracts “further expose them to any disappointments in demand.”

The aggregate numbers are stark even by this cycle’s standards. OpenAI’s disclosed long-term compute procurement commitments of roughly $665 billion stand against annualized revenue of about $25 billion — a ratio of more than 26 to 1 between contracted future outflows and current annual income. OpenAI plans roughly $600 billion in computing spend by 2030 against revenue it projects will exceed $280 billion by then; skeptics note the company’s own revenue projections have already been revised downward once, with one restructuring-related disclosure cutting a 2030 revenue target from approximately $85 billion to approximately $39 billion in the relevant scenario. Anthropic carries its own multibillion-dollar commitments to Google Cloud and Amazon, though at a materially smaller multiple of revenue. xAI has borrowed at fixed rates up to 12.5 percent to fund its buildout — a coupon that tells you exactly how the credit market prices the risk.

The commitment lattice creates three specific fragilities. First, it converts a demand forecast into a legal obligation: a model company that overestimated 2028 inference demand cannot simply spend less, the way a hyperscaler trimming discretionary capex can. It must pay, renegotiate, or default. Second, it concentrates counterparty risk in entities that are themselves unprofitable. Sub-investment-grade neocloud tenants have become a recognized underwriting category, with lenders demanding parent guaranties or hyperscaler credit “wrappers” precisely because the tenant’s standalone credit does not support the lease. Third, it synchronizes the stress. Because so many obligations were signed in the same 2024–2026 window with similar tenors, the heavy payment years cluster — which means a demand shortfall arriving in 2027 or 2028 would hit many counterparties in the same quarters rather than being spread across a decade.

Signs of strain at the edges are already documented. Microsoft canceled data-center projects representing about 2 gigawatts of capacity across the US and Europe, citing oversupply — the observation Michael Burry seized on as evidence that even the deepest-pocketed buyer sees more capacity coming than demand justifies. SoftBank’s CFO conceded that Stargate was “taking longer than anticipated,” with committed equity of $52 billion standing against the $500 billion headline and the balance dependent on debt financing that has been slow to materialize. Sarah Friar’s private warning that OpenAI may be unable to pay for future computing contracts if revenue growth fails to accelerate is the same fragility described from the inside.

For the timing question, the commitment lattice matters because it fixes the calendar. Unlike sentiment, which can turn any day or never, contracted payments arrive on schedule. The heaviest step-ups in the OpenAI complex land in 2027 and 2028 as Stargate phases come online and the Azure and Oracle contracts scale. The window in which promised revenue must appear is therefore not open-ended — it is roughly the next eight quarters, and every earnings season inside that window is a checkpoint at which the market gets to compare the payment schedule against the income actually arriving.

Circular deals and money that travels in loops

The single feature of this cycle that most unsettles veterans of previous ones is circularity: the same dollars appearing on multiple companies’ books as both investment and revenue, round-tripping through the ecosystem in ways that make the underlying demand signal genuinely hard to read.

The canonical example is the Nvidia–OpenAI arrangement. Nvidia has committed up to $100 billion to OpenAI, structured around bringing 10 gigawatts of Nvidia systems online, and OpenAI’s CFO acknowledged plainly that the money “will go back to Nvidia” in GPU purchases. Nvidia invests in OpenAI; OpenAI spends the investment on Nvidia chips; the spending books as Nvidia revenue; the revenue supports Nvidia’s valuation; the valuation funds further investments. Nvidia is likewise a prominent investor in CoreWeave, which buys Nvidia GPUs to rent to OpenAI, which is funded partly by Nvidia. Microsoft invested tens of billions in OpenAI, much of it consumed as Azure compute — Microsoft’s own cloud revenue. The restructured 2026 agreement includes OpenAI’s $250 billion commitment to purchase Azure capacity, which is simultaneously OpenAI’s largest cost line and one of Microsoft’s most cited forward-revenue assets. Oracle’s roughly $300 billion OpenAI contract drove a historic single-day surge in Oracle’s stock when announced — a valuation gain built on the creditworthiness of a customer that loses more than a dollar per dollar earned.

Circular structures are not inherently fraudulent or even unusual in infrastructure buildouts; vendor financing has existed as long as railroads. The problem is informational. When Nvidia reports blowout data-center revenue, some unknowable fraction of it traces to purchases funded, directly or indirectly, by Nvidia’s own capital or by debt raised against contracts with unprofitable counterparties. Bulls read the revenue as proof of organic demand; bears read it as the ecosystem billing itself. Both readings fit the same disclosures, which is precisely the epistemic condition — unfalsifiable optimism — in which manias extend furthest. The dot-com era’s closest analogue was telecom vendor financing: Lucent and Nortel lending customers the money to buy Lucent and Nortel equipment, booking the sales as revenue, and collapsing when the customers did. Lucent’s peak-to-trough decline exceeded 99 percent, a reminder that the vendor side of a circular structure is not insulated from its customers’ failure.

Michael Burry’s formulation of the same point has become the skeptic shorthand: chip stocks rise because big tech spends heavily; equipment makers rise because chipmakers add capacity; investors treat each new spending announcement as independent confirmation of demand. When Samsung and SK Hynix announced a vast chip hub in Korea, Burry told the Wall Street Journal, “I see that as the beginning of the end” — capacity expansion read not as demand evidence but as the supply-side overshoot that ends every capex mania. Alphabet, Amazon, Meta, Microsoft, and Oracle raised $255.34 billion through debt and equity in 2026 while planning roughly $750 billion in data-center spending by year-end, per Axios — meaning even the most cash-rich companies in history are now externally financing a growing share of purchases that flow to suppliers whose valuations depend on those purchases continuing.

The analytical takeaway is narrow but important for timing. Circularity does not tell you demand is fake. It tells you that reported revenue overstates the amount of independent, end-user money entering the system, by an amount nobody outside the flows can compute — and that when the cycle turns, the unwind runs backward through the same loop with the same amplification. The suppliers whose rallies have led 2026 would, in a demand disappointment, discover that a portion of their order books was a reflection of their customers’ financing rather than their customers’ customers. That discovery process, in prior cycles, is what converted corrections into crashes.

Hyperscaler capex against free cash flow

The strongest single argument against the bubble thesis has always been the identity of the spenders, so the deterioration in their cash generation during 2026 deserves careful, unhurried treatment — it is the pivot on which the whole debate turned.

Through 2023, 2024, and most of 2025, the bulls held an unanswerable point: Microsoft, Alphabet, Meta, and Amazon funded the AI buildout overwhelmingly from operating cash flow. Combined net income across the group ran to hundreds of billions of dollars annually, net margins for the Magnificent Seven exceeded 25 percent against the S&P 500 average of 13 percent, and balance sheets held cash reserves comparable to mid-sized national economies. Companies spending their own profits on productive assets do not fit the classic bubble template of speculative capital chasing revenue-free stories. Goldman Sachs and J.P. Morgan built their non-bubble case on exactly this foundation: monetization had begun, the buyers were paying enterprise customers, and the growth was funded by earnings rather than by hope.

That foundation eroded in measurable stages. Capex grew 77 percent year over year into 2026 while revenue at the same companies grew at a fraction of that pace — capex scaling roughly 50 percent faster than revenue by mid-2026. The consequence is the number already flagged in the data section, worth restating because it anchors everything: projected 2026 free cash flow for the five big hyperscalers of roughly $16 billion, down about 91 percent, against net income projected to rise 25 percent to about $506 billion. The gap between those two figures is the buildout. Every incremental dollar of reported profit, and then some, is going into the ground as GPUs, shells, cooling, and power.

Accounting mechanics widen the optical gap further. Capex does not hit the income statement when spent; it depreciates over the asset’s useful life. A company can therefore report record earnings in the exact quarters it is consuming record cash, with the income-statement reckoning deferred to future depreciation charges. The hyperscalers have collectively lengthened assumed server lifespans to as much as six years, which spreads those charges thinner — and which becomes its own controversy, taken up in the depreciation section below. For now the point is simpler: reported earnings currently flatter the economics of the buildout, and cash flow tells the unflattered version. Markets historically tolerate that divergence for a long time and then, at some hard-to-predict threshold, stop — June 2026 was the first month in this cycle that looked like the stopping.

The individual company pictures diverge in ways the single-trade framing hides. Alphabet’s Q1 2026 print — revenue of $109.9 billion, up 22 percent, with Google Cloud growing 63 percent — is the closest thing the bulls have to proof that AI capex converts to revenue, and the market rewarded it with year-to-date gains while the rest of the complex fell. Meta has kept operating margins around 41 percent despite heavy spending, monetizing AI invisibly through advertising performance rather than through separately billed products. Microsoft, by contrast, carries the largest capex number, the deepest OpenAI entanglement, and the June-worst 20 percent monthly decline. Amazon booked only $43.2 billion of its projected $200 billion by March 31, a pacing gap that either signals discipline or foreshadows a second-half spending surge, depending on the reader. The market is learning to price these as different businesses with different AI payback periods, and Goldman’s dispersion data confirms the single-trade era is ending.

The forward question is the sustainability of the spending itself. Free cash flow near zero leaves the buildout’s continuation dependent on either rising operating cash flow — plausible if cloud AI revenue accelerates in the second half of 2026 as several banks expect — or external financing, which is already happening at scale and is the subject of the next two sections. A third possibility is the one the market fears most and the suppliers’ shareholders have not priced: the hyperscalers simply slow down. Capex guidance cuts of even 15–20 percent for 2027, announced in an October or January earnings call, would preserve the spenders’ cash flow and devastate the supplier complex whose 2026 rallies assume the trillion-dollar trajectory continues. In every prior capex mania, the buyers’ rationality was the sellers’ catastrophe.

Debt enters the story and changes the risk profile

If a historian of this period had to pick the single structural change that most altered the bubble’s risk profile, it would not be a model release or a valuation milestone. It would be the migration of the buildout’s financing from retained earnings to debt across 2025 and 2026 — because equity bubbles bruise investors, while debt-financed bubbles break economies.

The scale of the migration is documented across every credit channel at once. AI-related companies and projects raised at least $200 billion in debt during 2025 — an acknowledged undercount given how many deals are private — with projections in the hundreds of billions for 2026 alone. Hyperscalers issued approximately $121 billion in bonds in 2025, more than four times their five-year average, with AI-related investment accounting for roughly 30 percent of net issuance in the US dollar investment-grade market. Total data-center debt issuance nearly doubled to $182 billion. Alphabet, Amazon, Meta, Microsoft, and Oracle together raised $255.34 billion through debt and equity in 2026. Bank of America’s head of global credit described the volumes with unusual candor for a financing banker: “The numbers are like nothing any of us who have been in this business for 25 years have seen. You have to turn over all avenues to make this work.”

The pricing of the riskier tranches maps the credit spectrum’s honest opinion. xAI’s fixed-rate debt carries coupons of 12.5 percent. CoreWeave refinanced in the high-yield market at around 9 percent. One documented GPU-collateralized facility carried a variable rate averaging roughly 11 percent with repayments beginning in January 2026 — just as the collateral’s market value was declining sharply, a sentence that could have been lifted from a 2007 mortgage post-mortem with the nouns changed. Spreads on investment-grade bonds of the major AI issuers widened by as much as 40 basis points relative to comparable credits between September 2025 and early 2026, an early and classical sign of investor discomfort with sector concentration.

Structure compounds scale. Meta’s $27.2 billion data-center financing with Blue Owl combined features of asset-backed securities, commercial mortgage-backed securities, and investment-grade debt in an off-balance-sheet vehicle — engineering that keeps obligations out of headline leverage ratios while preserving them in economic reality. Oliver Wyman’s analysts drew the uncomfortable parallel directly: in 2008, banks discovered they owned far more US housing risk than their internal reports suggested, and the off-balance-sheet character of much AI financing recreates the conditions for the same discovery. The Federal Agricultural Mortgage disclosure cited earlier in the credit chain projects nearly $1.8 trillion of new debt financing flowing into US data centers through 2030, across syndicated loans, private credit, ABS, and CMBS channels — with the existing data-center ABS market of roughly $25 billion facing projected refinancing take-out needs approaching $300 billion.

Debt changes the bubble calculus in three mechanical ways. It introduces fixed payment dates, converting “revenue eventually” from a hope into a covenant. It creates forced sellers in a downturn — leveraged holders who must liquidate at the bottom rather than wait for recovery, which is what turns 30 percent corrections into 80 percent ones. And it builds contagion channels to holders far outside the technology sector: insurers, pension funds, retirees in private-credit vehicles, and banks with warehouse exposure. An equity-only AI correction would have impoverished technology shareholders and stopped roughly there. The 2026 version, with more than a trillion dollars of debt in and around the buildout, has acquired the transmission mechanisms of a credit cycle — and credit cycles are what the BIS, whose warning anchors a later section, exists to worry about.

The timing implication is the most concrete in this article. Debt schedules are public, dated, and unforgiving. Heavy AI-related maturities and repayment step-ups cluster in 2027 and 2028; refinancing needs scale each year through 2030. A boom financed by earnings can outwait skepticism indefinitely. A boom financed by debt has appointments to keep.

Private credit, securitization, and the shadow side of the buildout

The fastest-growing financing channel of the buildout is also the least transparent, and the combination of speed and opacity in credit markets has a historical track record bad enough to warrant its own section.

Private credit funds — principally Blackstone, Blue Owl, Apollo, Pimco, and BlackRock — now originate most data-center debt, through off-balance-sheet special-purpose vehicles and direct lending facilities to neoclouds and developers. Outstanding private-credit loans to AI-related companies surged from near zero to over $200 billion within a few years, and Morgan Stanley projects private credit will supply an additional $800 billion in data-center financing over just the next two years. Oliver Wyman expects over $1 trillion of total AI debt to come from private credit — stacked on top of roughly $3 trillion in existing global private credit outstandings, meaning AI infrastructure alone could grow the asset class by a third.

Private credit’s defining features cut both ways here. Its loans are unlisted and rarely marked to market, which dampens the daily volatility that public bond markets transmit — a stabilizer in mild stress. But the same opacity means deterioration accumulates invisibly. There is no daily price discovering that a neocloud’s GPU collateral has lost 40 percent of its resale value, no index forcing funds to acknowledge that a tenant’s parent guaranty is weaker than underwritten. Insurance brokers told CNBC that data centers have become a “stress test” for their industry, with one veteran noting that insuring a $20 billion campus was nearly impossible in 2023 and has become a weekly conversation in 2026, and describing the sector as “almost going back to the same cycle where there’s almost no transparency about the financing structures — the scale is astronomical.” When the people paid to price catastrophe describe your financing structures as opaque at astronomical scale, the observation deserves weight.

The collateral question sits underneath everything. A large fraction of the new lending is secured, directly or effectively, by GPUs and the cash flows they generate. GPU-backed lending assumes the chips hold resale value across the loan’s life — an assumption Nvidia’s own product cadence attacks annually, as each new generation improves performance per watt enough to make prior generations uneconomic for frontier workloads well before their loans amortize. Legal analysts have already begun cataloguing the coming disputes: Quinn Emanuel’s early-2026 client alert on data-center financing litigation reads as a preview of the claims — misrepresented collateral values, disputed depreciation assumptions, guaranty enforcement, inter-creditor fights in structures that blended ABS, CMBS, and corporate credit into instruments no single doctrine cleanly governs. Bankruptcy practitioners have separately flagged that a distressed data center may struggle to obtain debtor-in-possession financing in a tightened credit environment, meaning failures could resolve as liquidations rather than reorganizations, dumping capacity and collateral onto falling markets.

The systemic question is who ends up holding the losses. Private credit’s investor base runs through pension funds, insurers, sovereign funds, and, increasingly, semi-liquid vehicles marketed toward individuals. The channels from an AI demand disappointment to a retiree’s statement now exist and are widening; whether they carry crisis-scale losses depends on numbers — recovery rates on GPU collateral, true leverage inside SPVs, concentration within individual funds — that will not be knowable until tested. The honest statement is conditional: the private-credit layer converts an AI correction from a technology-sector event into a test of the least-examined corner of the modern financial system, and no one, including its participants, can currently size the exposure with confidence.

The BIS warning and the view from the central banks’ bank

Institutional warnings deserve sorting by the incentives of the institution issuing them. Short sellers profit from fear, banks from reassurance, consultancies from complexity. The Bank for International Settlements profits from none of these, which is why its 2026 annual report — the most direct bubble warning yet issued by a body of its standing — anchors the official-sector portion of this analysis.

The BIS placed the current AI investment boom in an explicit lineage: the canal and British railway manias of the 1800s, the electrification exuberance of the 1920s, and the dot-com boom of the 1990s. Its summary sentence is worth reproducing in full because of who wrote it: all of these episodes “shared one common trait: a genuine technological breakthrough that attracted capital in excess of what commercial returns could ultimately justify.” The report continues that these episodes “ended with an eventual reversal in investment, inducing economy-wide recessions,” and that the scale and pace of the current boom, with its expectations of large productivity payoffs, “bear resemblance to these precedents, highlighting potential downside risks in the near term.” Central-bank prose does not get more pointed.

The report’s specific concerns map onto the structures documented above. It estimates the five largest hyperscalers will spend more than $1 trillion on AI-related capex in 2026 and observes that these commitments now outpace earnings and free cash flow, pushing firms into debt issuance — the migration this article has traced channel by channel. It flags widening credit spreads and warns that circular investing increases contagion risk. It identifies a “supply side roadblock” of electricity availability, chip shortages, and grid-connection bottlenecks, warning both that data centers are pressuring energy prices with “potential spillovers to inflation” and that temporary shortages “may also amplify over-investment, as firms attempt to lock in future capacity through long-dated contracts that further expose them to any disappointments in demand” — as precise a description of the commitment lattice as exists in print. And it closes the loop with the macro-financial scenario: should inflation spike or AI-led investment collapse, existing financial vulnerabilities could amplify the consequences, with policy tightening potentially precipitating “a sharp pullback in asset prices after a prolonged period of exuberant risk-taking, triggering disruptive macro-financial feedback loops.”

Buried in that scenario is the report’s most underappreciated point: the policy response would be constrained. The AI boom is itself inflationary at the margin — through energy prices, memory prices, construction costs, and the sheer GDP weight of the buildout — which means the classic remedy for a bursting bubble, aggressive rate cuts, could be unavailable or delayed if inflation is elevated when the bust arrives. The 2000 and 2008 crashes were cushioned by central banks with room to ease. A 2027 AI bust arriving alongside buildout-driven inflation would find the fire brigade already occupied. Seeking Alpha’s summary of the report put the limitation bluntly: an AI bubble burst and private-credit stress “with limited policy response capability.”

The IMF’s assessment, for balance, runs milder: a burst is possible but unlikely to trigger a 2008-scale financial crisis, with damage sizable but concentrated. The difference between the two institutions is largely about the private-credit unknowns — the IMF weights the banking system’s limited direct exposure; the BIS weights the untested channels. Both agree on the diagnosis of capital outrunning demonstrated returns. Neither offers a date, and the BIS’s own historical framing explains why: in every precedent it cites, the reversal came one to four years after credible institutions first warned, a lag that in the current cycle brackets everything from late 2026 through 2029.

Depreciation schedules, chip lifespans, and Michael Burry’s accounting case

The most technical strand of the bear case concerns depreciation, and it deserves patient treatment because it is the argument that, if correct, means the boom’s reported profitability — the bulls’ central exhibit — is partly an accounting artifact with a scheduled expiry.

The mechanics first. When a hyperscaler buys a GPU cluster, the cost is capitalized and expensed gradually over the asset’s assumed useful life. Lengthen the assumed life and each year’s depreciation charge shrinks, lifting reported earnings without changing a single cash flow. Across recent years, the major tech companies extended assumed server lifespans on their books to as long as six years — moves each justified individually by improved utilization and software efficiency, and each of which flattered earnings in the year it was made.

Burry’s Scion analysis, laid out in the Substack newsletter he launched after deregistering his fund, attacks the six-year assumption at its root. Nvidia’s product cadence has compressed to roughly annual: Hopper, then Blackwell, with Vera Rubin sampling to customers in 2026. Each generation delivers step-change gains in performance per watt, and since power is the binding constraint of the entire buildout, a chip two generations old is not merely slower — it is uneconomic to run for frontier workloads, occupying scarce megawatts that newer silicon would use several times more productively. Functional obsolescence, on this argument, arrives in two to three years while the books amortize over six, meaning the industry is systematically understating the true annual cost of its compute fleet and overstating current earnings by the difference — with the reckoning arriving either as future write-downs or as permanently elevated replacement capex that never tapers the way the models assume.

Nvidia has pushed back, arguing its hardware stays productive far longer than critics claim because CUDA software tuning keeps extracting value from older generations, which can cascade down to less demanding inference work. The rebuttal is partly persuasive — a thriving secondary market in Hopper-class inference capacity exists — but Burry and others have seized on the tension inside it: Nvidia simultaneously markets each new generation as categorically superior, to justify upgrade purchases, and defends the enduring economics of the old ones, to protect customers’ depreciation schedules and its own demand narrative. Both claims can be partially true; they cannot both be fully true. One of those defenses, as the critics put it, has to give.

The sums involved move the question from accounting trivia to market-level materiality. With $450 billion or more of 2026 hyperscaler capex directed at AI infrastructure and the GPU fleet’s replacement cycle disputed by a factor of two or three, the annual earnings difference between the bull and bear depreciation assumptions runs to tens of billions of dollars per company — enough to swing whether the AI segments are profitable at all on true economics. GPU-collateralized lending imports the same dispute into credit: a loan underwritten against six-year collateral life is mispriced if the honest number is three, which is how the depreciation argument connects to the private-credit fragilities above. The documented case of a GPU-backed facility whose repayments began in January 2026 exactly as its collateral’s market value slid is the argument in miniature.

Burry’s own positioning traces the argument’s evolution. His late-2025 disclosures targeted Nvidia and Palantir. By mid-2026 his short book had broadened to the trade’s outer rings: Tesla, Applied Materials, Caterpillar — the last a construction-equipment maker shorted purely as AI-infrastructure exposure, evidence of how far he believes the mania has spread — and, as of July 1, Micron at $1,051.87 after its near-700 percent one-year run, with the semiconductor index it belongs to trading at 76 times earnings and 13 times book. His stated logic for the Micron short distills the whole accounting case: a rally driven by “fear of missing out, the greater fool theory, and public commitment bias” rather than by durable end-demand economics. Whether his timing proves better than it was in early 2023 — when he covered bearish positions before a historic rally — is a separate question from whether his arithmetic is right, and the two have been confused before, in both directions.

The bears, name by name

Skeptical positions differ enough in mechanism and timeline that lumping them as “the bears” obscures more than it reveals. The serious versions, individually:

Michael Burry runs the accounting-and-circularity case just detailed: depreciation schedules overstating earnings, capacity announcements masquerading as demand signals, supplier rallies as the terminal phase. His public shorts — Nvidia, Micron, Tesla, Caterpillar, Applied Materials — express a view that the unwind hits the whole supply chain, not just the model companies. May 2026 warnings under his name helped anchor the Polymarket repricing.

Jeremy Grantham, whose bubble calls span Japan 1989, dot-com, and housing, has called the current US market the most expensive in history and predicts a major correction if not a crash. His framework is valuation reversion at the index level rather than AI mechanics specifically — the Shiller CAPE near 40 does the work — and his well-documented weakness is earliness, often by years, which he acknowledges as the cost of the method.

Bill Gurley has drawn the late-stage dot-com comparison from the venture side, focusing on private-market dynamics: the 60 percent of all US venture funding flowing to AI startups in 2025 (from 23 percent in 2023), valuation-to-revenue multiples in the private rounds, and the return of structures — SPVs stacked on SPVs, secondaries at escalating marks — that characterized 1999’s final phase.

Ruchir Sharma supplies the macro trigger: the bubble bursts if US rates rise on persistent inflation, since extreme price-to-sales stocks are valued almost entirely on discounted future earnings and reprice violently when the discount rate moves. He has pointed at 2026 specifically as the vulnerable year under that scenario — the one bear thesis with a stated timeframe, contingent on a policy path.

Richard Bernstein has called an AI bubble the likely outcome from a market-structure standpoint, emphasizing concentration: five companies at 30 percent of the S&P 500, AI-adjacent names driving roughly 80 percent of 2025’s US equity gains, and passive flows mechanically amplifying whatever the largest names do, in both directions.

The BIS, covered above, is the institutional bear — historically grounded, mechanism-focused, timing-agnostic.

Alongside them sits a figure the label fits awkwardly: Warren Buffett, whose Berkshire Hathaway has amassed roughly $325 billion in cash, the largest reserve in its history. Buffett has issued no AI-specific warning; the position is the statement. The investor most famous for deploying capital when others are fearful has spent the boom accumulating the means to do so, which markets have long read as a valuation opinion expressed in treasuries rather than words.

Setting the theses side by side exposes the pattern relevant to timing: the bears agree on diagnosis and diverge completely on trigger — accounting recognition for Burry, rate policy for Sharma, valuation gravity for Grantham, private-market exhaustion for Gurley, flow reversal for Bernstein, credit stress for the BIS. Six mechanisms, six different calendars, one shared conclusion. Historically that configuration — consensus on the disease, dissensus on the date of death — has described every major bubble roughly one to three years before its peak, which is precisely as unhelpful for market timing as it sounds and precisely why the scenario section later in this article works in probabilities rather than predictions.

The bull case rests on real cash flows

An analysis that only steelmans the bears would be advocacy, not analysis. The case against the bubble thesis — or at least against the catastrophic version of it — is substantive, made by serious institutions, and correct on several points where the bears are loose.

The foundation is the one already conceded: the profitability of the core spenders. The dot-com bubble’s fatal population was companies with no revenue and no viable business model trading at absurd multiples. Microsoft, Alphabet, Meta, and Amazon generate real profits in the hundreds of billions of dollars, hold cash reserves at nation-state scale, and — the June free-cash-flow squeeze notwithstanding — retain the ability to throttle capex at will. A company spending its own earnings can stop; a company spending borrowed hope cannot. Goldman Sachs and J.P. Morgan rest their fundamentally-justified verdict here: the chips are being bought by paying enterprise customers, monetization has begun, and early AI applications in enterprise software, cybersecurity, and specialized infrastructure already carry high margins.

The revenue evidence, while thinner than the spending, is not imaginary. OpenAI’s climb from $3.7 billion to a $25 billion run rate in eighteen months is the fastest software revenue scaling ever recorded; it reached $12 billion annualized faster than Google or Facebook reached comparable milestones by years. Alphabet printed 63 percent cloud growth in Q1 2026. Meta funds its buildout from an advertising machine whose AI-driven performance gains are measurable in its 41 percent operating margins. Anthropic’s enterprise share gains show a genuine, competitive, paying market for frontier models rather than a single-vendor story. Fortune 500 adoption of AI tools is near-universal at 92 percent for at least one vendor’s products. None of this resembles Pets.com.

Valuation, the bears’ favorite exhibit, is more contestable than the CAPE headline suggests. Howard Marks and Larry Fink — neither known for pom-poms — argue that while prices are elevated, they lack the irrational-mania signature of an imminent top: today’s leaders are profitable businesses with real cash flows, and any correction may therefore be more gradual and less severe than 2000. Nvidia’s forward price-to-earnings ratio has actually fallen even as its stock price rose, because earnings have grown faster than the shares — the exact inverse of bubble mechanics, where prices detach from earnings. After the June selloff, Meta traded at roughly 16 times earnings against the S&P 500’s 21; Barron’s called the group attractively valued. A bubble whose epicenter trades below the market multiple is, at minimum, an unusual bubble.

The demand-side rebuttal is the strongest forward-looking element. Every measurable indicator of compute demand remains supply-constrained: Nvidia’s Blackwell generation sold out, cloud GPU capacity sold out, Micron’s earnings confirming memory shortages, UBS noting AI supply-chain bottlenecks showing no sign of easing, and Jensen Huang reporting Q1 FY2027 revenue of $81.6 billion, up 85 percent, with the claim that hyperscaler compute spend is pushing toward $3–4 trillion. The bulls’ point is simple: overbuilding manias end with unsold inventory and idle capacity, and this cycle exhibits the opposite — every chip made is bought, every megawatt energized is consumed. Sovereign AI programs tripled to over $30 billion in fiscal 2026, adding a state-backed demand layer no prior tech cycle had. Dan Ives’s formulation is that the buildout’s cost will one day look like the cost of building Las Vegas in the 1950s — enormous, mocked, and retrospectively obvious — once “AI consumer hardware, physical AI deployments, and enterprise use cases explode at scale.”

The bull case’s honest weakness is that most of it defends the technology and the spenders while the bear case attacks the financing and the timeline — the arguments pass each other. Supply constraint today does not prove the demand is end-user demand rather than the circular, commitment-driven variety. Profitable spenders do not immunize the leveraged neoclouds, the GPU-collateralized lenders, or the suppliers whose order books assume the capex trajectory holds. And “more gradual and less severe” — the Marks–Fink formulation — is not actually a rejection of the bubble thesis. It is a claim about the shape of the deflation, which may be the most accurate framing available and is treated as its own scenario later in this article.

Dot-com parallels that hold and the ones that break

The dot-com comparison does so much work on both sides of this debate that it deserves to be disassembled into its components, tested one at a time.

Parallels that hold. Market concentration: the top seven names constitute roughly 35 percent of the S&P 500, matching the top-seven share at the 2000 peak, with the top ten at 35 percent against 25 percent then — today’s market is more concentrated than the dot-com extreme. Valuation at the index level: the Shiller CAPE above 40 in 2025 matched a threshold crossed only once before, in 1999–2000. Capital flooding a single theme: 60 percent of US venture funding to AI in 2025 mirrors the internet share of 1999 venture allocation. Infrastructure overbuild rhetoric: Gartner’s 2026 forecast that AI will approach the entirety of IT spending by 2030 rhymes uncomfortably with its February 2000 projection of $7.29 trillion in B2B e-commerce by 2004 — issued one month before the Nasdaq’s peak, and wrong by more than half. Vendor financing and circularity: Lucent and Nortel lending customers the purchase price of their own equipment maps directly onto Nvidia’s investment-linked GPU sales. Even the Davos vibes reproduce: one 2026 attendee compared the mood to January 2000, “right before the Internet bubble burst,” when it “was similarly ebullient.”

Parallels that break. The profitability of the leaders, exhaustively covered above — Cisco earned real money in 2000, but the median dot-com did not, whereas today’s median mega-cap AI spender is among the most profitable firms ever. The equity-versus-debt starting point: the dot-com bubble was overwhelmingly equity-financed, which is why its collapse, brutal as it was for shareholders, produced only a mild 2001 recession; the AI cycle’s trillion-dollar debt layer inverts this comparison in the direction of greater systemic risk, not less — the parallel breaks against the optimists here. Demand visibility: 1999’s build was ahead of any measurable usage for much of the fiber laid, while 2026’s compute is fully utilized the moment it energizes. And scale relative to the economy: US equity capitalization near twice GDP exceeds the 2000 peak, AI investment accounts for roughly half of US GDP growth, and the buildout has been called the largest peacetime investment project in history — meaning both the boom’s real-economy contribution and the hole its reversal would leave are larger than the dot-com equivalents.

The timeline lesson, applied with discipline. The dot-com chronology that matters for the timing question: credible warnings from December 1996; acceleration through 1997–1999 with two fully-recovered crashes along the way; a blow-off top in the final five months when the Nasdaq nearly doubled; the peak in March 2000 triggered by nothing in particular — a confluence of rate hikes already delivered, a failed mega-IPO window, and insider lockup expirations; then thirty-one months to the bottom, minus 83 percent. Mapping crudely: if ChatGPT’s launch in November 2022 plays the role of Netscape’s 1995 IPO, mid-2026 corresponds to roughly 1998–1999 — the phase of recovered warning-shot crashes and institutional alarm, one to two years before the top, not after it. If instead the January 2025 DeepSeek shock was this cycle’s 1998 and June 2026 its October 1999, the peak window falls in late 2026 through 2027. The mapping is a heuristic, not a law; its honest use is to establish that nothing in the dot-com precedent supports confidence that the burst is imminent, and nothing in it supports confidence that the trade has years of safety either. The precedent’s one firm teaching is about the aftermath: the technology’s believers were vindicated a decade later, and the investors who financed the vindication were not the ones who profited from it.

Railways, canals, and electricity as the older precedents

The BIS reached past the dot-com era for its comparisons, and the older manias reward the detour, because each isolates a mechanism the internet precedent blurs.

The British railway mania of the 1840s is the purest precedent for infrastructure-led overinvestment by fundamentally sound actors. Railways rebuilt the economy around themselves — they repriced land, labor, and time across an entire economy — and the promoters were frequently established, profitable enterprises rather than frauds. At the mania’s 1846 peak, railway investment absorbed several percent of British GDP, a share comparable to what AI-related investment now contributes to US growth. Parliament authorized thousands of miles of track that duplicated existing routes, because each promoter’s projections were individually plausible and collectively impossible — the exact logic of five hyperscalers simultaneously building capacity for the same enterprise AI demand. The bust arrived in 1847 not through demand failure but through a financing shock: the Bank of England tightened amid a commercial crisis, calls on partly-paid railway shares came due, and leveraged holders liquidated into a falling market. Rail traffic kept growing through the entire collapse. The infrastructure was right; the capital structure died anyway — the single most transferable sentence in the historical record for anyone holding AI-adjacent credit in 2027.

The canal mania of the 1790s adds the technological-succession risk. Canals were genuinely superior to road haulage and genuinely profitable — the early ones paid dividends for decades. The late-mania canals were destroyed not by demand shortfall but by railways, a successor technology their fifty-year payback assumptions never priced. The AI translation is uncomfortable for the GPU-collateral complex specifically: infrastructure with multi-decade financing amortization built for a technology whose architecture — transformer inference on Nvidia-style accelerators — could itself be succeeded within the loan term, whether by radically cheaper inference silicon, by algorithmic efficiency gains of the DeepSeek variety compounding for a decade, or by architectures not yet published.

The 1920s electrification boom supplies the market-structure warning. Utilities were the growth stocks of the decade, their expansion real and their output demand genuine — electricity consumption grew straight through the 1930s. What collapsed was the financial architecture built on top: pyramided holding companies, notably Samuel Insull’s, in which thin layers of equity controlled vast leveraged asset bases, and which unwound catastrophically when credit tightened after 1929, wiping out retail investors who believed they owned the safest infrastructure in America. The modern rhyme is the SPV-and-holding-layer architecture of data-center finance — off-balance-sheet vehicles, wrapped credits, securitized leases — in which the distance between the ultimate investor and the ultimate asset has widened exactly as it did in 1928, with exactly the effect on risk visibility.

Across all three precedents plus dot-com, the BIS’s structural summary holds without exception: genuine breakthrough, excess capital, eventual investment reversal, economy-wide recession. The timing distribution across the four episodes is the useful extract. Measured from the first year of institutional-scale warnings to the investment peak: roughly one year (canals), two years (railways), three-plus years (dot-com, from 1996), and about two years (utilities, from mid-decade margin concerns to 1929). Institutional-scale AI warnings date from roughly late 2024 to mid-2025. The historical base rate therefore brackets the investment peak between late 2026 and 2028 — a range, not a date, and a range whose width is itself the honest finding.

The productivity evidence so far

Every dollar of the $2.5 trillion rests, in the end, on one empirical claim: that AI raises the productivity of the people and firms using it by enough to pay for itself. Through 2025 that claim ran on anecdote and vendor benchmark. In 2026 the first wave of serious measurement arrived, and its findings — genuinely mixed, resistant to both camps’ talking points — are the most important underreported inputs to the bubble question.

The adoption base is now beyond dispute. Coordinated surveys by research teams at the Federal Reserve Bank of Atlanta, the Bank of England, the Deutsche Bundesbank, and Macquarie University — nearly 6,000 CEOs, CFOs, and senior finance managers surveyed between November 2025 and January 2026 — found 69 percent of firms across the US, UK, Germany, and Australia reporting some current AI use, with the US highest at 78 percent. Text generation via large language models is the most common application at 41 percent of firms. Individual usage studies cluster around 50 percent of people using AI, roughly half of that for work. Firm-level adoption in the US Census Bureau’s business trends survey — a stricter measure — ran at 18 percent of businesses as of December 2025, rising toward 21 percent, a reminder that “adoption” spans everything from an enterprise deployment to one employee’s chatbot habit.

The productivity findings are positive, modest, and slow — three adjectives that matter in that order. The NBER executive survey of nearly 750 corporate leaders found labor productivity gains that are real, concentrated in high-skill services and finance, and expected to strengthen in 2026, driven mainly by revenue-side total factor productivity rather than capital deepening. The multi-country central-bank study found the same directional result at the macro level: industries with higher AI adoption have shown faster recent productivity growth, statistically significant and similar in magnitude across Europe and the US, with causality unproven. Employment effects, per the same literature: near zero in aggregate — expected decline of less than 0.4 percent attributable to AI in 2026, large firms shedding modestly while small firms add. And the barriers are prosaic: over 40 percent of companies invested nothing in AI during 2025, citing technological immaturity (42 percent), untrained workforces (36 percent), and privacy concerns (36 percent).

Against that steady-but-slow academic picture stands the finding that became the bear case’s favorite citation: MIT’s “GenAI Divide” report, which examined 300-plus deployments, 52 organizational interviews, and 153 executive surveys and concluded that 95 percent of enterprise generative AI pilots deliver no measurable P&L impact, despite $30–40 billion of enterprise investment. The report’s internals are more interesting than its headline. The failures trace not to model quality but to integration — generic tools that do not learn workflows, budgets misallocated to sales and marketing pilots when the measured ROI concentrates in back-office automation, and a “learning gap” between tools and organizations. The successful 5 percent show large, concrete returns: BPO cost reductions of $2–10 million annually, 30 percent agency-spend cuts, externally partnered deployments succeeding at twice the rate of internal builds. Critics of the study, including its more careful readers, note that “no measurable P&L impact within a pilot window” is a standard early-cycle finding for general-purpose technologies — the same instruments would have failed 95 percent of 1996 corporate website pilots — and that a thriving “shadow AI economy” of unofficial employee usage generates value the P&L attribution never captures.

The February 2026 NBER finding of minimal real-world productivity impact despite executive optimism was, notably, cited by Polymarket traders as a factor in bubble-market pricing — evidence that the academic literature now feeds directly into the financial debate. The synthesis the full literature supports: AI’s measured productivity effect in mid-2026 is real, positive, concentrated, and at least an order of magnitude too small — so far — to justify $2.5 trillion of annual spending on any near-term payback arithmetic. The bulls’ rejoinder that electricity took twenty years to show up in productivity statistics is historically accurate and financially useless, because the debt documented in this article does not have twenty years.

Enterprise adoption against enterprise value

Between the macro productivity studies and the market’s daily verdicts sits the enterprise software market itself — the place where AI spending must eventually convert into recurring, contracted, high-margin revenue if the valuations are to survive. Its 2026 condition merits a close look, because it is both the bulls’ best evidence and the source of a competitive force that could deflate parts of the trade without any macro trigger at all.

The genuine strength: enterprise AI spending is real, growing, and increasingly contracted rather than experimental. OpenAI’s enterprise and team revenue climbing from roughly $1 billion annualized at the start of 2025 to over $7 billion by mid-2026 is the single cleanest series. Anthropic’s rise to roughly 40 percent enterprise LLM share by Menlo Ventures’ measure, with its implied valuation crossing $1 trillion on secondary platforms, demonstrates a competitive market with multiple credible vendors — a structural improvement over any single-vendor boom. The 92 percent Fortune 500 penetration figure, whatever its definitional generosity, reflects procurement decisions by the most process-heavy buyers in the economy. And the deployment pattern the MIT study’s successful 5 percent reveals — back-office automation, document-heavy analysis, coding — matches where the models are objectively strongest, suggesting the market is slowly learning to buy what works.

The structural worry inside the same data: intensifying competition is compressing the economics before scale arrives. OpenAI’s enterprise share falling from 50 percent to 27 percent in two years while Anthropic’s rose, Google’s Gemini 3 topping consumer benchmarks and triggering Altman’s internal “Code Red” in December 2025, Microsoft shipping in-house MAI models to reduce OpenAI dependence, Meta and xAI giving frontier-adjacent capability away — this is a market where the product is improving faster than pricing power is forming. Inference costs per token have collapsed generation over generation, which is wonderful for adoption and corrosive for revenue-per-use. The enterprise software historian will recognize the pattern: markets that commoditize at the model layer push the durable margins up the stack, to application, workflow, and distribution layers — which is an argument for Microsoft’s and Google’s eventual capture and against the pure-play model companies’ current valuations, a distinction the AI trade prices only partially.

The churn question is the one enterprise datum nobody outside the vendors can verify. Annualized run rates reported at moments of rapid growth reveal nothing about seat utilization, renewal rates, or the fraction of pilots the MIT study would classify as failed that nonetheless renew for a year out of inertia. The one public natural experiment — the shadow-AI finding that employees route around failed official deployments with consumer tools — cuts both ways: it confirms demand for the capability while undermining the enterprise contract value that the capitalizations discount.

The link back to the bubble’s timing is direct. The scenario in which enterprise revenue accelerates through late 2026 and 2027 — Cramer’s rally trigger, where one hyperscaler raises forecasts because of its AI products — is the scenario in which the June selloff marks a correction inside a longer boom. The scenario in which enterprise growth decelerates while competition compresses margins is the one in which the capex trajectory loses its justification from below, no macro shock required. The Q2 2026 earnings season now underway is, by common agreement across the bull-bear spectrum, the first clean read.

The consumer side of the ledger

Consumer AI is the boom’s most visible face and, financially, its most puzzling component: usage at civilization scale, payment at rounding-error scale.

The usage numbers have no precedent. ChatGPT alone serves on the order of 900 million weekly users — with the caveat that it missed its internal one-billion target for end-2025, the miss that catalyzed April’s partner selloff. Adding Gemini’s distribution through Android and Search, Meta AI’s through its social platforms, and the rest, some large fraction of the connected human population now touches generative AI weekly, four years after the category’s consumer debut. No prior technology — not the web, not smartphones, not social media — reached comparable penetration this fast.

The payment numbers are the puzzle. Roughly 3 percent of American households currently pay for AI services. OpenAI’s roughly 20 million paid subscribers against 900 million weekly users is a conversion rate near 2 percent. ChatGPT subscriptions at about $17 billion annualized constitute the largest consumer AI revenue stream on Earth — and amount to less than a single quarter of Netflix-scale subscription economics applied to a user base many times Netflix’s size. The gap admits two readings. The optimistic one: conversion is early, ARPU compounds as agents and premium tiers mature, and the free tier is a deliberately underpriced acquisition engine for the largest audience ever assembled. The pessimistic one: consumers have efficiently priced a service whose free tier is nearly as good as its paid one and whose alternatives are numerous and improving — meaning consumer willingness to pay, not consumer interest, is the binding constraint, and it binds hard.

Advertising is the historically proven escape from exactly this trap — it is how search and social converted free usage into two of the most profitable business models ever built — and the strategic behavior around it is revealing. OpenAI delayed advertising and shopping-agent initiatives during the December 2025 “Code Red” to concentrate on model quality, a defensible priority call that nonetheless deferred the one revenue mechanism sized to its audience. Meta, meanwhile, monetizes AI invisibly through ad-performance gains inside its existing inventory, at 41 percent operating margins, without asking consumers for a dollar — the quiet counterexample suggesting the consumer AI profit pool may accrue to whoever already owns the attention and the ad infrastructure, rather than to whoever built the model. Google’s position is the same argument with a search box.

For the bubble arithmetic, the consumer ledger’s contribution is straightforward: at 3 percent household payment penetration, consumer subscriptions cannot carry the buildout on any timeline the financing permits, so the valuations depend on enterprise revenue, advertising, or agentic commerce scaling — each real, none proven at the required size. The consumer numbers guarantee AI’s cultural permanence. Cultural permanence and capital returns are different assets, as every newspaper shareholder learned two decades ago.

Power, chips, and the physical limits of the buildout

The AI boom is the first technology mania whose constraint is physical before it is financial, and the physics complicates the bubble analysis in both directions at once.

The constraints are documented at every layer. Electricity first: the buildout’s projected tens of gigawatts of new demand collide with grid-connection queues measured in years, transformer lead times measured in multiples of years, and generation additions that cannot legally or physically match the data-center construction pace. The BIS flagged electricity availability, chip shortages, and grid bottlenecks as a “supply side roadblock,” warning that data centers are already pressuring energy prices and input costs with “potential spillovers to inflation.” Morgan Stanley’s 2026 energy outlook frames investor positioning around natural gas, storage, nuclear, and off-grid generation precisely because utility timelines fail the buildout’s schedule — developers now plan temporary on-site generation to bridge delayed utility deliveries, and utility counterparty risk has become a standard underwriting line item. Chips and memory second: Blackwell sold out, cloud GPU capacity sold out, high-bandwidth memory in structural shortage with the Roundhill Memory ETF up 166 percent in 2026 on the pricing consequences, and advanced packaging capacity at TSMC rationed among the largest buyers. Academic work on the financing side, including Van Nieuwerburgh’s 2026 analysis, draws the direct credit implication: if hardware cannot be delivered at scale, debt collateralized by data-center leasing revenue faces shortfalls because actual compute capacity falls below contracted capacity.

Read one way, the physical constraints are the bubble’s best defense. Overbuilding manias die of glut — unsold fiber, dark capacity, empty office towers — and a buildout that cannot physically proceed fast enough to overshoot demand is structurally protected from the classic ending. Scarcity underwrites pricing power up and down the supply chain, which is exactly what Micron’s earnings and the semiconductor complex’s 2026 rally reflect. On this reading, the binding constraint on AI is megawatts, not demand, and the trade’s suppliers are collecting economic rent on a shortage with years to run.

Read the other way, the same constraints are accelerants. The BIS’s second-order warning is the sharper one: temporary shortages “may also amplify over-investment, as firms attempt to lock in future capacity through long-dated contracts that further expose them to any disappointments in demand.” Scarcity panic is how the commitment lattice got built — take-or-pay power agreements, decade-long capacity reservations, prepaid chip allocations — and every contract signed to beat the shortage becomes a fixed obligation if demand plateaus. The 45-gigawatt projected capacity gap by 2028 that bulls cite as demand insurance is, from the credit side, a description of how much capital is being committed against forecasts. Shortage-driven price inflation in memory, power, and construction also raises the cost per unit of delivered compute, worsening the ROI arithmetic the June selloff repriced — Microsoft’s CFO explicitly attributed part of the 2026 capex increase to rising memory and component costs. And the inflation spillover closes the macro loop: buildout-driven price pressure is one of the identified paths to the higher-rates scenario in the next section, meaning the boom’s physical intensity could summon its own financial trigger.

The synthesis: physics rules out the fastest bust scenarios — there will be no 2001-style discovery of vast dark capacity in 2026, because the capacity cannot be built that fast — while simultaneously loading the slower scenarios with more fixed obligations per quarter of delay. The constraint that prevents a glut today is manufacturing the contract structure of a glut in 2028, if and only if demand disappoints by then. Once again the analysis lands on the same eight-to-twelve-quarter window, approached this time from the supply side.

Interest rates as the most likely pin

Every survey of bubble history returns the same finding about proximate causes: bubbles are rarely punctured by revelations about the asset. They are punctured by the price of money. The 1847 railway crash followed Bank of England tightening; the 1929 unwind followed 1928–29 rate rises; the Nasdaq’s March 2000 peak followed six Fed hikes; the 2022 rehearsal — when the Nasdaq lost roughly 33 percent in a year as rates rose — demonstrated the modern AI-adjacent complex’s specific sensitivity. The mechanism is arithmetic, not psychology: growth assets are claims on distant cash flows, and the discount rate is the gravity acting on distance.

Ruchir Sharma has made this the explicit core of the one bear thesis carrying a timeframe: the AI bubble bursts if US rates rise on persistent inflation, with 2026 the vulnerable year under that path. The n-tv analysis of the tech rally’s financing sensitivity spells out the transmission: higher rates raise capital costs across a buildout that now borrows hundreds of billions annually, cut the present value of the future-growth expectations that constitute most of the complex’s market value, and pull flows toward newly attractive lower-risk assets. Each channel operates independently; a rate shock fires all three at once.

What makes the rate scenario more than generic macro caution is the reflexive loop this cycle has built into it. The AI buildout is itself inflationary through the channels just documented — energy prices, memory and component costs, construction labor, and its outsized weight in GDP growth, which the BIS and others put at roughly half of US growth. A boom that generates inflation invites the tightening that kills booms: the snake is positioned to eat its own tail. And the same loop constrains the rescue. The post-2000 and post-2008 recoveries were engineered with aggressive easing; a bust arriving amid buildout-driven inflation would find that tool politically and economically jammed, which is the “limited policy response capability” the BIS warning turns on. The debt layer sharpens everything further: floating-rate private credit facilities — including the documented GPU-backed loans at variable rates averaging 11 percent — transmit any tightening to borrower cash flows within a quarter, no market repricing required.

The scenario’s probability is genuinely uncertain and honest analysis says so. As of mid-2026 the rate path is contested, inflation prints are the market’s most-watched macro series precisely because of this loop, and nothing obliges the tightening scenario to occur — rates could as easily fall into a slowing economy, extending the trade’s financing conditions for years. The claim here is narrower and better supported: among all candidate pins — earnings disappointment, credit event, counterparty failure, capex capitulation — a rate shock is the only one with a four-for-four record across the historical precedents, and this cycle has wired itself to that pin more tightly than any predecessor, by financing with debt what earlier manias financed with equity and by generating, through its own physical appetite, the inflation that moves rates. Anyone maintaining a probability distribution over burst timing should weight the quarters following any sustained upward inflation surprise far more heavily than the calendar-uniform prior.

Concentration risk in an index built on seven stocks

The bubble question is usually framed as a question about AI companies. For most people’s actual savings, it is a question about index funds, because the concentration statistics have quietly converted every passive portfolio into a leveraged AI position.

The figures: the top ten stocks represent roughly 35 percent of the S&P 500, against 25 percent at the dot-com peak. The five largest account for about 30 percent. The Magnificent Seven’s total return since January 2020 is nearly eightfold while the remaining 493 stocks have not doubled, and AI-adjacent names drove approximately 80 percent of US equity gains in 2025. Microsoft, Alphabet, Amazon, and Meta alone carried more than $10 trillion in market value and 17 percent of the index as of April. US equity capitalization stands near twice GDP. Every one of these concentration measures exceeds its March 2000 reading.

The mechanics of concentration matter more than the level. Passive vehicles now hold a majority of US equity fund assets, and passive flows are mechanically pro-cyclical at the security level: every 401(k) contribution buys the largest names in proportion to their size, amplifying whatever the market has already decided. The amplifier runs identically in reverse — sustained outflows or even a rotation sell the largest names hardest — and June 2026 provided the live demonstration, when institutional money rotating toward small-caps and value stripped $2.3 trillion from seven companies in twenty trading days while the equal-weighted index barely noticed. The wealth-effect channel scales with the same numbers: at twice-GDP capitalization, a 25 percent decline in the top names erases wealth on the order of a quarter of US GDP, with the consumption consequences that implies, before any credit transmission begins.

Geographic exposure follows the concentration. The United States bears the overwhelming share of direct AI equity risk — its technology names dominate global AI investment and drove the 2025 gains. Analyses of other markets, including a Motilal Oswal report on India, find relative insulation through limited pure-play exposure; European indices are likewise less directly exposed, though no market of consequence would escape a US-led correction’s second-round effects through trade, credit, and sentiment. For the European reader, and specifically for anyone whose pension or index savings are globally allocated from Slovakia or its neighbors, the practical translation: a world-index fund is now roughly one-fifth Magnificent Seven by weight, and the diversification it promises is thinner against this specific risk than against any risk in the product’s history.

Concentration also reshapes the burst’s plausible mechanics. A market this top-heavy does not require the bubble narrative to break for the index to fall — it requires only that two or three mega-caps disappoint in the same season, which June demonstrated. Conversely, dispersion inside the seven (Alphabet up 12 percent in the same year Microsoft fell 22 percent, per the deVere and Goldman data) means the index can mask a rolling, name-by-name deflation for quarters — the “Magnificent Three” path, in which the bubble deflates as a sorting rather than a crash and the index-level damage arrives in slow motion. Both paths run through the same seven tickers, which is the point: whatever the AI bubble does, the S&P 500 has been structured to do it too.

Prediction markets and the honest math of timing

For the timing question specifically, one dataset deserves separate treatment because it is the only one where people answer with money rather than commentary: the prediction markets.

Polymarket’s “AI bubble burst by…?” contract, launched November 19, 2025 and carrying $2.9 million in cumulative volume, prices the leading outcome — burst by December 31, 2026 — at roughly 16 percent, with the by-2025 outcome expired at zero. Read carefully, the market is saying three things. First, a near-term burst is a live minority scenario, not a fringe one: 16 percent is roughly the probability assigned to a US recession in an average year, priced here for a specific asset-class event. Second, the complement dominates: 84 percent against a 2026 resolution means traders expect the trade to survive the year, the June selloff notwithstanding — the market moved on the May Burry warnings and the NBER productivity findings and still declined to make the burst the base case. Third, and most subtly, the contract’s own structure concedes the definitional problem this article opened with: “burst” requires resolution criteria, and reasonable observers will dispute in real time whether any given drawdown qualifies. June 2026 subtracted $2.3 trillion from seven companies and the contract did not resolve, which is itself a data point about how high the bar sits.

The limitations are real and worth stating. $2.9 million in volume is thin relative to the question’s importance; prediction markets embed the same herding as any market; and the trader population skews toward exactly the technology-adjacent demographic most saturated in the debate’s narratives. The academic literature on prediction markets nonetheless finds them at least as calibrated as expert forecasters on comparable horizons, and their one unambiguous virtue applies fully here: the price moves when information arrives, without an editor, a book to talk, or a client letter to justify.

Situating the market price among the other timing signals produces the fairest available summary. Historical base rates from four manias bracket the investment peak one to four years after institutional warnings began, pointing at late 2026 through 2028. The financing calendar concentrates obligations in 2027–2028. The dot-com mapping, for whatever a single precedent is worth, places mid-2026 at roughly the 1998–1999 phase. Sharma’s rate-contingent thesis names 2026 conditional on an inflation path that has not yet resolved. Analyst consensus across the factually.co synthesis and the bank commentaries clusters on 2026–2027 as the highest-risk window with the pain concentrated rather than systemic — the IMF’s reported position. And the money says 16 percent by year-end. No honest method — base rates, financing schedules, precedent mapping, or market prices — produces a date. Every honest method produces the same window: material and rising risk from late 2026, peaking through 2027 and 2028. The next section converts that window into scenarios with probabilities attached, which is the strongest claim the evidence permits.

Three scenarios for the deflation

Forecasting a date is astrology; forecasting scenarios with probabilities is analysis. Three deflation paths cover the plausible space, drawn from the mechanisms this article has documented and weighted by the historical base rates and current market pricing. A fourth, non-deflation path is included because intellectual honesty requires it. The probabilities are this analysis’s own synthesis, stated so they can be disagreed with precisely.

Scenario one: the sharp burst — roughly 20 percent through end-2027. The Oliver Wyman “hybrid scenario turbocharged by debt.” A trigger — a sustained inflation surprise forcing rates up, an OpenAI IPO failure or post-IPO collapse, a major counterparty default in the neocloud or Stargate complex, or a hyperscaler capex capitulation — initiates equity repricing that the debt layer converts into forced selling. GPU collateral values gap down, private-credit marks are finally taken, refinancing windows shut for the $300 billion ABS take-out pipeline, and the reversal propagates through the circular deal structure in reverse. Equity drawdowns in the AI complex reach 50–70 percent; the S&P 500, at current concentration, falls 30–40 percent; S&P Global’s scenario work puts job losses at up to 2.5 million in a full burst; AI capex, at roughly half of GDP growth, drags the US into recession with the BIS’s constrained-policy caveat governing the response. The dot-com aftermath is the template for duration: a multi-year bottom, not a quarter.

Scenario two: the grinding deflation — roughly 40 percent, spread across 2026–2028. The Marks–Fink shape, and the modal outcome by this analysis’s weighting. No single trigger; instead, the June 2026 mechanism repeats each earnings season — capex scrutinized, payback dates pushed, one or two of the seven punished per quarter — while dispersion does the sorting Nigel Green predicts. The Magnificent Seven resolve into winners that capture AI economics (on current evidence Alphabet, Meta, and Nvidia are the market’s candidates) and consumers of AI infrastructure that derate to utility multiples. Pure-play valuations compress 40–60 percent over two years without any single crash day; weaker neoclouds and overleveraged data-center vehicles restructure quietly through the private-credit workout machinery; capex growth decelerates to match revenue growth around 2028. The index grinds sideways-to-down as its largest components offset; the economy avoids recession but loses its principal growth engine. Nobody ever agrees on the date the bubble burst, because there is none — which is the historically common outcome hiding behind the dramatic precedents.

Scenario three: the delayed burst — roughly 25 percent, 2028 or later. The trade re-accelerates first. Q2–Q4 2026 earnings deliver the Cramer trigger — a hyperscaler raises guidance on AI products — capital floods back, the OpenAI IPO prices above $1 trillion and trades well, and the final leg carries the complex through 2027 in the blow-off pattern of 1999’s last five months, with retail participation via the reserved IPO allocations playing its historical role. The financing calendar then does its work: the 2027–2028 commitment step-ups and refinancing walls arrive into whatever demand actually materialized, and the reckoning occurs at higher altitude with more debt aboard. This is the scenario the 1846 and 1929 precedents fit best — the burst that follows the euphoria that follows the warnings — and it is the worst of the four for eventual severity, because every quarter of extension adds fixed obligations.

Scenario four: no bubble after all — roughly 15 percent. The revenue arrives. Enterprise AI spending compounds through 2027 at rates that validate the capex; agentic products open per-task revenue at scale; the hyperscalers’ AI segments disclose returns that justify the trillion-dollar trajectory; the 2026 vintage of spending earns its cost of capital within the hardware’s life. The Shiller CAPE stays elevated for years — as it did through the late 1990s and again through the 2010s — and the June 2026 selloff joins the January episodes in the file of rehearsals for a crisis that never premiered. The probability assigned here is genuine, not a courtesy: the fastest revenue scaling in software history is underway, and forecasters who bet against American mega-cap execution have a long record of losing.

The weightings imply a summary sentence: roughly 85 percent probability that a material deflation of the AI trade occurs, against perhaps 20 percent that it takes the catastrophic form, with the center of mass on the grinding, multi-year, name-sorting variant beginning — arguably having already begun — in 2026 and running through 2028. The June selloff is compatible with scenarios one through three and is most naturally read as the opening quarters of scenario two, subject to revision by exactly the signals the next section catalogues.

Deflation scenarios and their assigned probabilities

The table compresses the full scenario analysis into its decision-relevant skeleton: an aggregate 85 percent weight on some material deflation, a modal path that is slow rather than cinematic, and a genuine minority case in which the spending is vindicated. The probabilities are a synthesis judgment, stated numerically so that each quarter’s evidence can revise them explicitly rather than rhetorically.

The signals worth watching quarter by quarter

A probability distribution is only useful if it updates. These are the concrete, publicly observable indicators that should move the weights — the dashboard this analysis would maintain, in rough order of information value.

Hyperscaler capex guidance, quarterly. The single highest-signal series. The 2026 figure of roughly $725 billion is confirmed; the variable is 2027 guidance, which begins forming in the October 2026 calls. Guidance growth above 20 percent extends scenario three; flat guidance confirms scenario two; a cut of 15 percent or more from any single company is the scenario-one tripwire, because the supplier complex’s valuations price continuation. Watch the language as closely as the numbers — Amy Hood’s memory-cost commentary moved markets as much as her figures.

AI-attributed revenue disclosure. The bulls’ burden of proof. Any hyperscaler breaking out AI revenue explicitly — or raising company-level guidance on AI products, the Cramer trigger — resets the debate bullishly. Continued non-disclosure through two more earnings cycles is itself evidence, of the unflattering kind. Alphabet’s cloud growth rate (63 percent in Q1) is the current proxy; deceleration below 40 percent would mark the proxy failing.

The OpenAI IPO, expected late 2026. The public prospectus — due roughly 60–90 days after the confidential filings — forces disclosure of the Microsoft restructuring terms, the capex schedule, and the profitability assumptions. Then the market prices it daily. Pricing above $1 trillion with a stable first year is scenario-three fuel; a pulled or repriced offering, or a broken first-year chart, is the single most probable scenario-one trigger on the visible calendar, with Friar’s reported caution against Altman’s aggressive timeline the tell worth remembering.

Credit spreads on AI issuers. The 40-basis-point widening from September 2025 was the early warning; the series to watch is spreads on Oracle, CoreWeave, and the data-center ABS complex, plus any repricing of new GPU-collateralized deals. Private-credit stress will surface here first, indirectly, before any fund gates or marks make news. A second leg of widening concurrent with equity weakness — rather than offsetting it — signals the debt layer engaging, which is the difference between scenarios one and two.

Inflation prints and the rate path. The Sharma trigger. Sustained upward inflation surprises, particularly with identifiable energy and memory components — the buildout’s own signature — shift weight from scenarios two and three toward one, per the four-for-four historical record of rate shocks as pins.

Counterparty milestones in the commitment lattice. Stargate phase completions and their financing closes (the $448 billion debt gap), Microsoft’s data-center cancellation pace beyond the documented 2 gigawatts, any renegotiation of the OpenAI–Oracle or Azure commitments, and the neocloud refinancing calendar. A single public renegotiation normalizes renegotiation, which is how commitment lattices unwind.

The depreciation footnotes. Any hyperscaler shortening server-life assumptions or taking accelerated write-downs on prior GPU generations concedes Burry’s arithmetic in the accounts, with mechanical earnings consequences across the sector. This is the lowest-probability, highest-information signal on the list.

Adoption and productivity series. The Census BTOS adoption rate (18 percent rising toward 21), renewal and expansion data in vendor disclosures, and the successor studies to the 2026 NBER wave. These move slowly and matter cumulatively: the bull case requires the measured-productivity line to bend upward within the financing window.

The disciplined reader will notice the list contains no sentiment indicators, no commentary volume, and no model-release tracker. Capability announcements have been the cycle’s least informative series — every release since 2023 has been read bullishly regardless of financial content. The bubble resolves on cash, contracts, and credit; the dashboard follows accordingly.

Sector-by-sector exposure if the correction comes

A deflation of the AI trade would not distribute its damage evenly, and mapping the exposure by sector is where the analysis becomes practical for readers who are neither fund managers nor technologists.

Semiconductors and hardware suppliers. The highest beta in either direction, and — counterintuitively, given their 2026 outperformance — the most exposed to the correction scenarios. The supplier rally rests entirely on the capex trajectory: Micron, the memory complex, networking vendors, and Nvidia itself trade on order books that are downstream of hyperscaler guidance. The semiconductor index at 76 times earnings with a beta above 2 prices continuation; a 15 percent capex cut converts sold-out backlogs into inventory corrections within two quarters, and the 2022 rehearsal (Nasdaq minus 33 percent on a rate move alone) understates the sensitivity now that valuations are higher. The vendor side of the circular structures — the Lucent seat — is occupied here.

Software and cloud. Bifurcated. Companies with AI revenue attached to existing distribution and margins — the Meta advertising model, Microsoft’s seat-based Copilot attach, Alphabet’s cloud — deflate toward their non-AI earnings power, which is substantial; the derating is painful but bounded. Pure-play AI application companies priced on forward multiples of unproven revenue face dot-com-style outcomes, and the private cohort — recipients of 60 percent of US venture funding — faces the standard post-bubble venture winter: down rounds, consolidation, and a fundraising environment that Startup Fortune’s H2 2026 preview already describes as concentrating capital at the top.

Financials. The novel exposure of this cycle. Banks carry syndicated construction loans and warehouse lines; insurers hold data-center ABS and private-credit paper; private-credit funds hold the $200-billion-and-growing direct book. Oliver Wyman’s advice to the sector — rigorous scenario analysis under both equity and debt-driven bursts, on the 2008 lesson that internal reports understated housing exposure — is the professional consensus stated politely. The retail transmission runs through pension allocations to private credit and through semi-liquid vehicles whose gates would become news in scenario one.

Energy and utilities. Asymmetric and partially protected. Power purchase agreements with investment-grade hyperscalers survive a correction; merchant capacity built on neocloud demand does not. The utilities’ regulated buildout — grid upgrades rate-based across all customers — persists regardless, which is why the sector’s AI exposure is more durable than its 2026 correlation suggests. A demand disappointment would strand the speculative fringe of generation projects while leaving the grid investment as the cycle’s most unambiguous residual value, the fiber-optic analogue.

Construction, industrials, and real estate. The Caterpillar short is the thesis: heavy equipment, electrical components, cooling, and data-center REITs have re-rated on buildout demand that is contractually softer than it appears — construction pipelines can be paused at phase boundaries. Commercial real estate exposure concentrates in markets where data-center land banking has repriced industrial acreage; a capex halt reverses those marks with the usual real-estate lags.

Labor markets. Two distinct channels, often conflated. The correction channel: S&P Global’s 2.5-million-job scenario in a full burst, concentrated in construction, technology, and finance — a cyclical shock. The technology channel runs opposite: the NBER evidence shows AI-attributable employment effects near zero in aggregate today (under 0.4 percent in 2026), with large firms shedding and small firms adding, and a burst would slow, not accelerate, whatever displacement the technology eventually causes. The paradox is worth stating plainly: the workers most exposed to the AI bubble bursting are the ones building it, not the ones it is supposed to replace.

Europe and the Slovak-adjacent view. Direct exposure is limited — European indices carry a fraction of the US’s AI weight, and the region’s participation in the buildout runs through suppliers (ASML above all) and power infrastructure rather than through model companies. The transmission to a small open economy like Slovakia is second-round: German industrial demand, euro-area credit conditions, and globally allocated pension savings whose US mega-cap weight was covered above. A US-centered correction of scenario-two shape would reach Central Europe as a growth headwind and a savings drawdown, not as a domestic financial event; scenario one would arrive the way 2008 did, through credit and trade, with the usual lag of two to four quarters.

Marketing, SEO, and the agency business inside the bubble question

This publication’s readers work in search, content, and digital marketing, and the bubble question lands on that industry with unusual directness — because the industry sits simultaneously inside the boom’s spending, inside its product, and inside its distribution channel. The exposure deserves its own accounting.

The channel dependency first. The past two years rewired search distribution around AI surfaces: Google’s AI Overviews and AI Mode, ChatGPT Search, Perplexity, Copilot, and the answer engines have inserted a model layer between content and audience, and the discipline of generative engine optimization — GEO — exists because visibility inside model outputs now carries commercial value the way blue links did. That entire optimization surface is a bet on the AI platforms’ permanence in their current form. Here the bubble analysis delivers its most reassuring finding for the industry: every deflation scenario in this article is a financial event, not a usage event. The 900 million weekly ChatGPT users, the AI Overviews rollout, and the 69 percent firm adoption rate do not reverse if Nvidia’s multiple compresses or a neocloud restructures — the railway traffic grew through 1847, the internet’s usage grew through 2001–2002. GEO as a discipline survives every scenario including the sharp burst; what changes is which platforms survive to be optimized for, and the scenario-two sorting suggests concentrating GEO effort on surfaces attached to profitable distribution (Google, Microsoft, Meta) over standalone answer engines whose funding runs through the venture channel that a correction closes first.

The client-spending exposure is the industry’s real cyclical risk. Marketing budgets are the historically first casualty of any downturn, and the MIT finding that 50–70 percent of enterprise AI budgets flow to sales and marketing pilots — the category with the weakest measured ROI — marks exactly the spending a CFO cuts when the June 2026 scrutiny reaches their own board. Agencies whose 2024–2026 growth came from AI-pilot project work should read that budget line as bubble-correlated revenue and weight retainers, performance-attributed work, and the back-office automation category (where the MIT study found the actual returns: $2–10 million BPO savings, 30 percent agency-spend reductions) accordingly. The uncomfortable detail in that last figure: “agency-spend reduction” is itself one of AI’s proven enterprise ROI categories, meaning the industry is both a seller of AI services and a documented line item in AI’s cost-cutting case. Agencies that internalize the productivity gain — using AI to deliver more output per billed hour — hold the defensible position; agencies whose pricing model depends on the hours AI eliminates do not.

The content-economics angle closes the loop with this publication’s recurring themes. The Pangram-style detection research and the AI-content flood have already compressed the value of undifferentiated text; a financial correction would compress it further by cutting the venture-funded content farms’ runway while leaving the demand for provenance, expertise, and original analysis intact — E-E-A-T as a flight-to-quality trade. The link-economy and digital-PR disciplines face the same sorting. A bubble deflation, in other words, would accelerate the industry’s existing divide between commodity production and differentiated authority rather than create a new one. For a Slovak or Central European agency serving clients in euros while the correction risk concentrates in dollar-denominated mega-caps, the practical summary: the technology stack survives every scenario, the client budgets are cyclical, and the strategic positioning that wins each scenario — provable ROI, authority content, platform-diversified visibility — is the same positioning that wins if no correction ever comes. Few industries get handed a strategy that dominates across the whole probability distribution; this one has.

Regulatory and geopolitical wildcards

The scenarios above assume the bubble resolves on economics. Two exogenous forces — regulation and geopolitics — could reshuffle the timing from outside the financial system, and this cycle has already demonstrated both at scale.

The export-control precedent is the cleanest case study, because it happened. US export controls on advanced AI systems produced, in the first half of 2026, the temporary shutdown and July restoration of frontier model access — an episode this publication covered at length — and the market lesson generalizes: a single policy decision can switch off or restore revenue streams that valuations treat as permanent. Nvidia’s current guidance embeds a zero-China assumption, which analysts frame as pure upside optionality; the same framing read defensively says a material fraction of the addressable market already sits behind a policy wall that moved twice in one year. Any tightening cycle in export controls — new thresholds, cloud-access rules, allied-country tiering — compresses the demand forecasts underwriting the capex; any loosening extends them. The variable is binary, political, and unhedgeable, which is precisely what discounted-cash-flow models handle worst.

Government entanglement runs deeper than trade policy. Preliminary discussions about transferring a 5 percent stake in OpenAI to the US government — potentially extended to other leading AI companies — introduce a governance category with no precedent for a company approaching a trillion-dollar listing. Sovereign AI programs, having tripled to over $30 billion in fiscal 2026, add state buyers whose demand is strategic rather than commercial — stabilizing in a downturn, but also subject to fiscal politics rather than ROI discipline. The Stargate program’s White House launch and the willingness of Gulf sovereign capital (MGX’s $7 billion committed tranche) to anchor the buildout mean any correction now has diplomatic dimensions: a Stargate financing failure would be a foreign-policy event as much as a credit event, which cuts both ways — governments rescue strategic projects that markets would abandon, and governments also abandon projects for reasons markets cannot price.

Litigation and antitrust supply slower-burning risks. The Musk–OpenAI trial that began March 30, 2026 produced discovery — the unsealed Brockman diary entries about the nonprofit conversion, the judge’s finding of “substantial grounds” for believing Musk was misled — that bears directly on the governance story OpenAI must tell public investors months later. Copyright litigation across the model companies remains unresolved at scale, with damages theories that range from nuisance to existential depending on jurisdiction. Antitrust attention to the circular investment structures — Microsoft–OpenAI, Nvidia’s ecosystem stakes — has so far produced inquiries rather than actions, but a structural remedy in any of them would rewrite the commitment lattice this article has treated as fixed. And the EU’s AI Act enforcement calendar phases in through exactly the 2026–2027 window the financing analysis flagged, adding European compliance costs to the model companies’ burn precisely when their revenue must accelerate.

None of these wildcards has a probability this analysis can defend numerically, which is why they sit outside the scenario weights rather than inside them. Their aggregate effect on the timing question is directional: nearly every one — controls tightening, litigation loss, antitrust action, subsidy withdrawal — shortens the runway, while only a few (controls loosening, state rescue) extend it. Exogenous risk, in this cycle, skews toward earlier rather than later.

Open questions the evidence cannot yet settle

Intellectual honesty requires marking the load-bearing unknowns — the questions on which this article’s probability weights would move most if answered, and which no public data currently answers.

The true unit economics of inference at scale. OpenAI’s Q1 2026 gross margin of 39 percent is one datapoint from one company under one pricing regime. Whether frontier inference is a business with software margins at maturity or a commodity with utility margins determines nearly everything downstream — the model companies’ terminal value, the durable price of compute, the hyperscalers’ AI-segment profitability. Falling inference costs help margins and invite price competition simultaneously; the net is unknowable from outside.

The share of reported AI revenue that is circular. The article documented the loops; nobody outside the flows can compute their magnitude. If independent end-user money constitutes 80 percent of the ecosystem’s reported revenue, the demand signal is broadly honest; at 50 percent, the boom is substantially self-billing. The OpenAI S-1’s related-party disclosures will provide the first real visibility, which is another reason the prospectus is the year’s most consequential document.

Private-credit marks and true leverage. The $200-billion-plus AI private-credit book carries no daily price. Recovery rates on GPU collateral have never been tested through a cycle; SPV leverage is disclosed to lenders, not markets. The system’s actual loss-absorption capacity in scenario one is a number that exists and is unpublished.

Agentic revenue. Every 2027-and-beyond revenue projection of consequence — OpenAI’s path to $100–125 billion, the hyperscalers’ AI-segment cases — leans on per-task agentic products that barely exist commercially in mid-2026. The category could be the escape velocity the bulls need or this cycle’s B2B-e-commerce forecast. Two more quarters of enterprise agent deployments will begin to say which.

Demand elasticity below the shortage. Every current demand signal is measured under supply constraint. The counterfactual — what utilization and pricing look like when the 2027 capacity waves land — is the whole question of whether the shortage is masking a demand plateau, and it is unanswerable until the capacity exists. The railway mania’s traffic grew through its crash; the fiber glut’s traffic grew too, at prices that ruined the builders. Both futures fit today’s data.

The successor-architecture risk. DeepSeek proved that efficiency shocks arrive unannounced. A comparable shock in 2027 — algorithmic, or silicon-based like the custom-chip wave (OpenAI’s Jalapeño with Broadcom, Meta’s chip entering manufacturing in September 2026) — could revalue the installed GPU base overnight in either direction: catastrophic for GPU-collateralized credit, liberating for the compute-constrained demand side.

These six unknowns share a property worth noticing: most of them resolve, at least partially, within the same 2026–2028 window the financing calendar defines. The S-1 publishes, the agent products ship or do not, the capacity lands, the credit book gets tested. The bubble question is not permanently unanswerable — it is scheduled to answer itself, in quarterly installments, over roughly the next eight to ten of them.

Practical positioning for businesses and investors

Analysis that ends without consequences is entertainment. What follows is framework, not personalized advice — the risk tolerances, time horizons, and tax situations that determine individual decisions cannot be known from here — but the framework is the part that generalizes.

For investors, the starting move is an honest exposure audit. Concentration statistics mean a standard global or US index allocation is already a substantial AI position: roughly a fifth of a world tracker and a third of an S&P 500 fund sit in seven names whose fortunes this article has spent twenty sections examining. The question is not whether to have AI exposure but whether the exposure held matches the exposure intended. From there, the scenario weights translate into orientations rather than trades. The 85 percent aggregate probability of material deflation argues against adding concentrated AI-complex risk at current prices; the 40 percent modal weight on the grinding, sorting scenario argues for dispersion-aware positioning within the complex — the June lesson that Alphabet and Microsoft are no longer the same trade — over binary in-or-out decisions. The 15 percent no-bubble weight and the 25 percent delayed-burst weight together caution against the opposite error: shorting or exiting entirely has cost skeptics more than the bubble has cost believers for three years running, Burry’s own 2023 record included. Historical severity data suggests the useful stress test: a portfolio that survives its AI-adjacent holdings falling 60 percent while the index falls 30 — the scenario-one shape — without forcing sales is positioned for the whole distribution. Diversification’s unglamorous specifics follow: equal-weight and value exposure as concentration hedges, non-US allocations for their documented insulation, duration and cash as the optionality Buffett’s $325 billion models, and skepticism toward any private-credit or semi-liquid vehicle whose AI exposure cannot be stated in one sentence.

For operating businesses, the deflation scenarios change surprisingly little about the technology decision and a great deal about the vendor decision. The adoption evidence — 69 to 78 percent of firms using AI, measured productivity gains concentrated where deployment is disciplined — holds across every scenario; the MIT failure taxonomy is a checklist for being in the successful 5 percent (back-office focus, workflow integration, external partners, measured P&L attribution) rather than a reason to wait. Vendor risk is the live variable: a correction sorts AI suppliers by balance sheet, and multi-year commitments to venture-funded platforms carry counterparty risk that the platform’s product quality does not reveal. Contracts that would survive a vendor’s acquisition, repricing, or disappearance — data portability, model-agnostic architectures, exit clauses — are cheap insurance priced at zero while the boom runs. Budget discipline follows the CFO logic documented above: AI spending attached to measured returns survives internal corrections; pilot spending attached to strategic FOMO is the first line cut, so building the measurement is building the budget’s survival.

For both audiences, the meta-advice is about calendars. The signals section gave the dashboard; the discipline is checking it quarterly against pre-committed thresholds rather than daily against sentiment. Bubbles impose their worst costs not on the exposed but on the reactive — the investors who sold the June bottom and will buy the next top, the businesses that froze all AI investment in a panic and resumed it at the peak. A written position on what evidence would change your allocation, decided in advance, is worth more than any forecast in this article — including this article’s.

The forecasting record of this cycle, kept for accountability

Before the final answer, one exercise that timing debates almost never perform: scoring the predictions already made and expired, because the pattern of past errors is itself evidence about which future claims to discount.

The failed bear calls come first, because there are more of them. Burst-by-2023 predictions, common after the first valuation run, expired worthless. Burst-by-2024 calls, which multiplied after the initial GPT-4 plateau commentary, did the same. The Polymarket by-2025 outcome resolved at zero. Grantham’s major-correction warnings have run continuously since 2021 against a market that roughly doubled. Burry covered bearish positions in early 2023 before one of the strongest rallies on record, and the SOXX index he is currently short returned 113 percent in the year to June 30, 2026 — his thesis may yet prove right and his entries have been consistently, expensively early. The August 2025 viral moment around MIT’s 95 percent finding produced a wave of imminent-collapse commentary; the twelve months since delivered the largest capex expansion in corporate history and record supplier earnings. The scoreboard’s first lesson: skeptics in this cycle have been reliably right about mechanisms and reliably wrong about dates, at an average cost that has so far exceeded the cost of believing.

The failed bull calls are fewer but instructive. OpenAI’s internal target of one billion weekly users by end-2025 was missed and never announced — the market learned of it from the Journal in April 2026 and repriced four public companies on the news. The company’s 2030 revenue projection was cut from roughly $85 billion to roughly $39 billion in restructuring-related scenarios within a year of being made. Stargate’s $500 billion headline stands against $52 billion of committed equity eighteen months later, with SoftBank’s own CFO conceding delay. Gartner’s trillion-dollar forecasting lineage — the $7.29 trillion B2B call of February 2000, wrong by more than half, issued a month before the peak — is the standing caution against the current $2.52 trillion projection’s precision. The AGI-timeline promises that anchored 2023–2024 valuations have been quietly extended by most of their original authors. The second lesson mirrors the first: the boom’s promoters have been reliably right about capability direction and reliably wrong about monetization schedules — the exact asymmetry the bubble thesis predicts.

Two calls scored well and deserve credit. The Jevons-paradox reading of DeepSeek — that cheaper models would expand rather than destroy compute demand — was mocked as cope in January 2025 and has been vindicated by every subsequent quarter of sold-out capacity. And the analysts who spent 2025 warning specifically about free cash flow rather than valuations — a minority position when net income was the headline — called the June 2026 repricing’s actual mechanism a year in advance.

The synthesis this record supports is the one threaded through the whole article, now with its evidentiary basis explicit. Directional claims about AI economics have been forecastable; dated claims have not, in either direction, by anyone, across four years of maximally incentivized attempts. A reader should therefore weight this article’s own probability distribution as its authors do: the scenario shapes and the mechanism analysis carry the confidence; the window carries less; any implied date carries none. Forecasting humility is not a rhetorical flourish here — it is the best-supported empirical finding the cycle has produced about itself.

An honest answer to the timing question

The question was when the AI bubble bursts. Twenty-plus sections of evidence permit an answer more useful than a date and more falsifiable than a shrug, and it comes in five parts.

First, on whether there is a bubble at all: by the BIS definition — genuine breakthrough, capital in excess of returns the business can eventually justify — the preponderance of evidence says yes, with roughly 85 percent confidence by this analysis’s weighting. The $2.5 trillion annual spend against low-hundreds-of-billions of revenue, the 26-to-1 ratio of OpenAI’s commitments to its income, the 91 percent free-cash-flow collapse at the spenders, the CAPE near 40, and the migration to debt collectively describe capital running ahead of demonstrated returns. The 15 percent residual is genuine respect for the fastest revenue scaling in software history and for the possibility that the returns arrive on schedule.

Second, on the shape: the modal outcome is not the cinematic crash but the grinding, multi-year sorting — scenario two, 40 percent — in which the trade deflates one earnings season and one company at a time, winners separate from spenders, and the “burst” never gets a date because it never happens on one. June 2026 is most naturally read as this process’s opening quarters. The sharp burst (20 percent through end-2027) and the delayed-then-worse burst (25 percent, 2028+) split most of the remainder.

Third, on the window: every independent method converged on the same range. Historical base rates from four manias: one to four years from institutional warning, bracketing late 2026 through 2028. The financing calendar: commitment step-ups and refinancing walls clustering in 2027–2028. The precedent mapping: mid-2026 resembling 1998–1999, not March 2000. The prediction market: 16 percent by end-2026, implying the center of mass beyond it. The window of maximum risk runs from late 2026 through 2028, with 2027 as its center — not because anything is scheduled to break in 2027, but because that is when the promises made in 2024–2026 are contractually due to be kept.

Fourth, on the trigger: unforecastable in identity, constrained in type. The historical record and this cycle’s structure shortlist a rate shock on buildout-driven inflation (the only four-for-four precedent), an OpenAI IPO failure or post-listing collapse, a hyperscaler capex capitulation, and a credit event in the private-lending layer. The dashboard section operationalizes the watch. Anyone claiming to know which fires first, or the month, has exceeded what the evidence supports — this analysis included.

Fifth, on what the burst would and would not mean. It would not mean AI failed, any more than 1847 refuted railways or 2001 refuted the internet; usage, adoption, and the measured productivity gains survive every scenario, and the buildout’s grid and compute infrastructure would be the correction’s residual gift to the 2030s, bought at par by its builders and operated at a discount by its inheritors. It would mean — as every precedent meant — that the capital structure erected around a real technology was built for a schedule the technology declined to keep. The original Slovak question, kedy praskne AI bublina, deserves its answer in the same plain register it was asked: most likely nie naraz, ale postupne — not with a bang but by installments — beginning in the market you are already watching, concentrating between now and 2028, and settled, like every question in this article, not by argument but by the arithmetic of cash against commitments, four quarters a year, until one of them wins.

The AI bubble questions readers keep asking

Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency

This article is an original analysis supported by the sources cited below

How the AI bubble could pop and take down the global economy, according to the BIS Coverage of the Bank for International Settlements’ 2026 annual report comparing the AI investment boom to railway, electrification, and dot-com manias, with its warnings on debt-financed capex, supply bottlenecks, and constrained policy responses.

The BIS Warning: AI Bubble Burst, Credit Event, Inflation, Fiscal Problems Analysis of the BIS report’s private-credit and circular-investment warnings, alongside Jeremy Grantham’s and Michael Burry’s bearish positioning.

Mag 7 value shrinks by $2.3 trillion amid AI spending jitters CNBC’s account of the June 2026 selloff, including the 10 percent monthly index decline, Microsoft’s 20 percent drop, and the divergent semiconductor rally.

The Magnificent Seven just lost $2 trillion and the market is asking a question big tech cannot yet answer Analysis of the June repricing with the hyperscaler free-cash-flow projection of roughly $16 billion for 2026 and the $725 billion capex figure.

Big Tech’s $2.7 Trillion AI Reckoning: What Investors Need to Know Extended breakdown of the June 2026 losses including Broadcom and Oracle, the 3 percent household payment-penetration figure, and the capex-versus-cash-flow gap.

Why The Magnificent Seven Could Become The “Magnificent Three” As Investors Warned Over AI Trade Nigel Green’s dispersion thesis, Goldman Sachs Asset Management’s 52.3 percent dispersion figure, and the contrasting valuation views from Barron’s and Wedbush.

The State Of The $2.52 Trillion AI Bubble, January 2026 Gil Press’s Forbes analysis of Gartner’s $2.52 trillion 2026 forecast, the historical record of trillion-dollar projections, and the Davos 2000 parallel.

AI Bubble 2026: Is the Tech Rally About to Burst? Survey of the bull and bear cases including Ruchir Sharma’s rate-trigger thesis, the Goldman Sachs and J.P. Morgan fundamentally-justified view, and Berkshire Hathaway’s $325 billion cash position.

AI bubble burst by…? Predictions & Odds 2026 The Polymarket prediction market pricing roughly 16 percent probability of a burst by December 31, 2026, with $2.9 million in cumulative trading volume.

OpenAI Revenue 2026: $25B ARR and the Path to Profitability Detailed breakdown of OpenAI’s mid-2026 revenue run rate, subscriber base, cumulative loss projections through 2029, and the compute-commitment gap.

OpenAI revenue, valuation & funding Sacra’s company profile covering the renegotiated Microsoft agreement, projected burn of $27 billion in 2026 and $63 billion in 2027, and the custom-silicon roadmap.

OpenAI Misses Revenue Targets—Bringing Shares Of These Investors Down Forbes’ coverage of the April 2026 Wall Street Journal report on missed internal targets, Sarah Friar’s compute-contract warnings, and the partner-stock selloff.

OpenAI Fell Short of Its Own Targets as Compute Costs Piled Up Reporting on the roughly $600 billion in locked-in data-center spending, board scrutiny of Altman’s capacity pursuit, and Anthropic’s higher implied valuation on Forge Global.

OpenAI Exposes Financial Black Hole on Eve of IPO Analysis of OpenAI’s Q1 2026 shareholder disclosures: $5.7 billion revenue, $3.7 billion cash burn, 39 percent gross margin, and $665 billion in compute procurement commitments.

Facing $14B losses in 2026, OpenAI is now seeking $100B in funding Deep examination of OpenAI’s burn rate, the Stargate equity-versus-headline gap, the Nvidia circular-investment structure, and the December 2025 Code Red.

Michael Burry’s next ‘Big Short’: An inside look at his analysis showing AI is a bubble CNBC’s account of the Scion depreciation thesis, Nvidia’s rebuttal, and Microsoft’s canceled 2-gigawatt data-center projects.

Michael Burry’s newest short reveals what really worries him about AI Detail on Burry’s mid-2026 short book — Nvidia, Tesla, Caterpillar, Applied Materials — and the semiconductor index’s 76 times earnings valuation.

‘Big Short’ investor Michael Burry issues blunt 4-word warning on AI stocks Coverage of the July 2026 Micron short at $1,051.87, the $255.34 billion in 2026 hyperscaler debt and equity raising, and Burry’s self-feeding-trade framing.

Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives NBER working paper surveying nearly 750 executives, finding positive productivity gains concentrated in high-skill services and near-zero aggregate employment effects.

Global Evidence on Business Use of AI NBER digest of the coordinated Atlanta Fed, Bank of England, Bundesbank, and Macquarie surveys finding 69 percent firm adoption across four countries with small measured effects.

Mind the Gap: AI Adoption in Europe and the U.S. Brookings-presented NBER paper documenting transatlantic adoption gaps and the statistically significant link between industry AI adoption and productivity growth.

MIT report: 95% of generative AI pilots at companies are failing Fortune’s coverage of the MIT NANDA “GenAI Divide” study, its integration-gap diagnosis, and the back-office concentration of measured returns.

The $3 Trillion AI Data Center Build-Out Becomes All-Consuming For Debt Markets Bloomberg-sourced reporting on the $200 billion-plus of 2025 AI debt issuance, xAI’s 12.5 percent coupons, and CoreWeave’s high-yield refinancing.

Client Alert: Emerging Litigation Risks in Financing AI Data Centers Boom Quinn Emanuel’s mapping of the data-center capital stack: $121 billion in 2025 hyperscaler bonds, private credit’s $200 billion-plus AI book, and GPU-collateral disputes.

How An AI Bubble Burst Could Shake Global Financial Markets Oliver Wyman’s equity and debt-turbocharged burst scenarios, the twice-GDP market-capitalization observation, and the 2008 hidden-exposure parallel.

Data Center Investment in 2026: AI Demand, Power Constraints, and Private Equity Trends Ropes & Gray’s survey of neocloud tenant credit risk, utility counterparty risk, and the ABS market’s $300 billion projected take-out needs.

Powering AI: Markets Race to Invest in AI Energy Solutions Morgan Stanley’s 2026 energy outlook on the trillion-dollar 2025–2026 spending commitment and the financing role of hyperscaler balance sheets.

Financing the AI Buildout Stijn Van Nieuwerburgh’s academic analysis of data-center financial structures, supply-side constraints, and the risks to debt collateralized by leasing revenues.

Nvidia Earnings: Updates and Commentary May 2026 Coverage of Nvidia’s $81.6 billion Q1 FY2027 quarter, the $725 billion hyperscaler capex revision, tripled sovereign AI spending, and Jensen Huang’s demand commentary.

Major Hyperscalers Just Reported Earnings. Nvidia Was The Winner Motley Fool’s summary of Q1 2026 hyperscaler results, including Microsoft’s $190 billion capex guidance and Amazon’s spending pace.

Is There an AI Bubble — And Will It Burst in 2026? Synthesis of analyst consensus on the 2026–2027 risk window, the IMF’s concentrated-damage assessment, and the near-term signals to monitor.

AI Bubble: Is It Real? When Will It Burst? Data-Backed 2026 Analysis Data compilation covering the Shiller CAPE record, the January 2026 Microsoft $440 billion single-day loss, venture concentration, and geographic exposure.

Jim Cramer says investors are getting the Mag 7 all wrong The bull counter-read of the June selloff: the supplier-over-spender trade, Meta’s chip announcement, and the guidance-raise rally trigger.