The AI Bubble in 2026: What the Data Says and When It Could Burst

The AI bubble by the numbers: $1 trillion and still climbing

The AI bubble debate has moved from social media arguments into central bank reports, Fed press conferences, and IPO roadshows. For US founders, engineers, and operators who build on AI infrastructure, the question is no longer whether the technology works. It is whether the money financing it can keep flowing at this pace, and what happens to pricing, funding, and vendor stability if it cannot.

So, is AI a bubble? The data supports two readings at once. On one side, spending, revenue, and chip demand are all setting records. On the other, cash flow at the biggest spenders is turning negative, interest rates are rising, and most companies that have bought AI tools cannot yet show a financial return.

Start with the size of the bet. Goldman Sachs Research estimates global AI-related investment will pass $1 trillion in 2026, including $581 billion in the US, and that cumulative spending since 2022 will reach about $1.8 trillion by year-end. The four largest US hyperscalers (Alphabet, Amazon, Meta, and Microsoft) now guide to roughly $720 billion to $745 billion of capital spending this year, after spending about $410 billion in 2025.

Nvidia sits at the center of that flow. It reported $96.2 billion of revenue for its fiscal second quarter on August 26, up 106% from a year earlier, with data center sales of $89 billion. It guided to $108 billion for the current quarter, above the $104.2 billion analysts expected. Its CFO said capital spending from the top five hyperscalers is expected to reach $1.3 trillion next year, up from $800 billion in 2026.

That scale sets the bar. Spending at this level only pays off if AI revenue grows into it. The rest of this report tests that assumption against the latest cash flow, revenue, interest rate, and forecast data available as of September 20, 2026.

Why analysts now describe a late-stage AI bubble

The bear case got its most detailed public airing this month. In a September 10 report, Capital Economics screened eight groups of market indicators and found most sitting at or near levels that preceded past market peaks. Among its flags: a cyclically adjusted price-to-earnings ratio close to the dot com bubble peak, expected S&P 500 earnings growth matching dot com highs, market value concentrated in a few stocks, a surge in new equity issuance, and free cash flow at the top hyperscalers projected to turn negative in 2027.

The firm's forecast is specific. It expects the S&P 500 to end 2026 near 8,250, then fall 21% to 6,500 by the end of 2027, with the AI bubble starting to burst next year. Fortune reports the firm also sees an eventual peak-to-trough decline of at least 30%, which would rank among the worst crashes of the past century. Capital Economics analyst James Reilly told CBS News that AI will prove transformative but that its profits are unlikely to reach what analysts now expect.

Interest rates are the new variable. On September 16, the Fed raised its target range by 25 basis points to 3.75% to 4%, its first hike in more than three years, in a unanimous vote. Chair Kevin Warsh then said broad financial conditions were not restrictive and declined to give forward guidance, and stocks slid. The Dow closed down 1.2% at 51,461.90, the S&P 500 fell 0.45% to 7,551.81, and the 10-year Treasury yield hovered near 5% after touching its highest level since 2007 the day before.

That matters because the dot com bubble unwound after a tightening cycle. Between June 1999 and May 2000, the Fed raised rates from 4.75% to 6.5%, and the S&P 500 then fell by nearly half over about two and a half years. Ruchir Sharma, chairman of Rockefeller International, has argued that AI mega projects get harder to fund if the 10-year yield decisively breaks above 5%, the top of its range since the dot com era. Capital Economics had forecast a milder unwind because it did not expect the Fed to tighten this cycle, a premise the September hike now tests.

Market behavior already looks unsettled. On July 30, Microsoft's market value rose by $450 billion in a single day, and Apple lost $360 billion the next day. A dispersion index tracked by Acadian Asset Management's Owen Lamont hit its third-highest reading in more than 2,850 trading days. Investors are split on whether this is a stock market bubble near its end or a healthy repricing. Morgan Stanley Wealth Management's Lisa Shalett told clients to stay calm through what she called the September and October silly season.

Where an AI bubble burst would show up first: Big Tech cash flow

Capital spending has grown faster than the cash the underlying businesses throw off, and second-quarter earnings made that visible. According to TMT Finance's tally of earnings-call figures, Alphabet posted negative free cash flow of $5.9 billion, its first negative reading since its 2004 IPO. Amazon's trailing twelve-month free cash flow swung to a $7.6 billion outflow, its first negative reading since 2023. Meta generated just $784 million. Microsoft was the outlier at $19.6 billion and says it will stay free cash flow positive through its 2027 fiscal year.

Financing has shifted to match. Alphabet's debt rose from about $16 billion a year ago to roughly $100 billion, and in June it announced an $80 billion equity raise that included a $10 billion private placement with Berkshire Hathaway. Meta took on $24.9 billion of new debt and stopped buying back shares. Oracle plans to raise about $40 billion in debt and equity in its 2027 fiscal year. Capital Economics adds that bond issuance from the top hyperscalers has more than doubled over the past year.

The Bank for International Settlements raised the same concern in June. Its annual report warned that debt-fueled data center spending and opaque financing ties between AI companies, shadow banks, and data center builders could threaten financial stability in an AI bust. AETW covered that report in June. If hyperscalers slowed their capex, the BIS said, borrowers across the supply chain would struggle to replace lost revenue and service their debt.

There is also an accounting question. Microsoft extended the assumed useful life of its data centers and office buildings from 15 to 25 years, which moves more future leases off the capex line. CFO Amy Hood said spending plans are unchanged in economic terms. Amazon CEO Andy Jassy argues that data centers earn revenue for 30-plus years and that servers and networking equipment pay back in under three years against a five- to six-year useful life. Investors will test whether the hardware lasts as long as those assumptions imply.

Customer concentration adds another layer. Microsoft's contracted backlog grew 84%, but only 25% excluding its OpenAI contract. Oracle's remaining performance obligations rose 363%, and OpenAI's reported cloud deal with Oracle is about $300 billion over roughly five years. The gap shows how much of the new backlog rests on commitments from a single AI lab, which is why that lab's financial health matters to the whole chain.

Revenue is surging, but proof of payoff is still thin

Demand is real, and the growth rates are unlike anything in enterprise software. Anthropic told investors its annualized revenue run rate reached $65 billion at the end of July, up more than sevenfold from year-end 2025, and Axios reported preliminary second-quarter revenue above $11.5 billion, more than 14 times the same quarter a year earlier. OpenAI closed a $122 billion round at an $852 billion post-money valuation in March, says enterprise now makes up more than 40% of its revenue, and has been reported at a run rate above $40 billion as of August.

Read those numbers with care. A run rate annualizes a recent period instead of reporting audited revenue, and Axios notes the two companies may not measure it the same way. FutureSearch puts OpenAI's recognized second-quarter revenue at $6.7 billion, which annualizes closer to $27 billion. Costs are heavy too. Outside estimates of OpenAI's 2026 loss or cash burn range from about $14 billion (HSBC, as cited by Value Add VC) to about $27 billion (Sacra), and its disclosed multi-year cloud commitments exceed $500 billion.

Anthropic has confidentially submitted a draft S-1 to the SEC and, per reports, plans to list this fall, with coverage citing valuations that could approach $2 trillion. It reportedly posted positive adjusted operating income in the second quarter. OpenAI filed confidentially on June 8 but is reportedly leaning toward a 2027 listing. Capital Economics treats the wave of new equity as a classic late-stage signal. Supporters read it differently: public markets funding companies whose revenue is compounding faster than any comparable business. Either way, the listings should give public investors their first audited look at AI lab economics.

Outside the labs, the payoff is harder to find. PwC's 29th Global CEO Survey, which polled 4,454 chief executives across 95 countries, found that 56% had seen neither higher revenue nor lower costs from AI, 33% had seen one of the two, and only 12% had seen both. The widely cited 2025 MIT Project NANDA study reached a harsher figure, finding that 95% of enterprise generative AI pilots produced no measurable profit-and-loss impact, though it measures pilots rather than mature deployments. On the other side, a 2024 IDC study sponsored by Microsoft found an average return of $3.70 per $1 invested. The studies measure different companies, which is why they can all be true.

The macro data has not confirmed a boom yet. The Bureau of Labor Statistics reported second-quarter productivity growth of 1.4%, the third straight weak quarter, according to Dean Baker of the Center for Economic and Policy Research. Baker argues that current valuations only make sense if productivity growth reaches 4% to 5%, well above the 1.5% to 2% of recent years. That is one economist's framing, but it points at the test that matters: whether earnings arrive on the schedule the market has priced.

The case against an AI crash: earnings, not just enthusiasm

Bulls have data too, and it is not trivial. Fidelity noted in February that the Russell 3000's aggregate capex-to-free-cash-flow ratio was below 1, versus nearly 4 in 2000, meaning companies were spending what they earned rather than what they borrowed. That cushion has thinned at the top since then, as the second-quarter numbers show, but it explains why many strategists do not treat 2026 as a repeat of 1999.

Goldman Sachs Research argues the buildout sits inside historical norms. It projects AI capex rising from 1.8% of US GDP in 2026 to 2.8% in 2028, within the 2% to 5% peak investment range seen in earlier general-purpose technology buildouts, and says its leading indicators sit near the top of their range since 2022. Morgan Stanley's Shalett notes the S&P 500's forward price-to-earnings ratio fell from 22.5 in January to about 19.5 without major losses, because earnings growth above 30% absorbed the compression. Her committee also calls the buildout mostly rate-insensitive and holds S&P 500 targets of 8,000 for year-end and 8,300 for mid-2027.

Nvidia's latest results back the demand story. CEO Jensen Huang said the company has supply for 70% growth in fiscal 2028 and that demand is higher than that, and Amazon Web Services announced a purchase of 2 million Nvidia GPUs. A post-earnings analysis from RexShares adds a caution: the dollars Nvidia adds each quarter are no longer growing, with a guided sequential step of $11.8 billion versus $14.6 billion the quarter before.

Students of past bubbles offer a third argument. BCG economists Philipp Carlsson-Szlezak and Paul Swartz wrote in Harvard Business Review on September 11 that bubbles can leave transformative infrastructure behind even when investors lose money, and that the more useful question for executives is whether a bust would threaten the wider economy, not when it arrives. Kenneth French of Dartmouth's Tuck School told CBS News there is not enough information to say whether current prices are too high or too low, and that AI could end up more important than people expect.

The bulls still have to answer the price question. Baker of CEPR, citing OpenRouter usage data, argues that Chinese open-weight models now handle a large share of traffic on that platform at token prices one-fifth to one-tenth of US frontier models. OpenRouter covers only its own users, so it is not a market census, but price compression at the model layer would squeeze the revenue that all of the infrastructure spending is meant to earn. An AI crash does not require the technology to fail. It only requires margins to come in below what valuations assume.

When will the AI bubble burst? Forecasts, triggers, and what to do

Nobody can time it, and the more careful economists say so. EY-Parthenon chief economist Greg Daco told CBS News that every technology wave produces some excess and that the word bubble needs careful use. Baker makes a similar point from the bearish side: there is no fixed date, and a bubble keeps inflating as long as investors believe in the payoff. So will the AI bubble burst, and when? The forecasts below are best read as a range of views, not a schedule.

Capital Economics says its best guess is that the bubble begins to burst in 2027, with an S&P 500 correction of at least 20% next year. Morgan Stanley sees 8,000 at year-end and 8,300 by mid-2027. Ed Yardeni cut the odds of his Roaring 2020s scenario from 80% to 70% and raised his bearish odds from 20% to 30%. On rates, the Fed's projections show one more hike this year and then a pause through 2027, but Warsh distanced himself from those forecasts, and LPL Financial's Jeffrey Roach says a cut may not come before 2028.

Several checkpoints will show which view is winning. Watch whether the 10-year yield holds above 5%. Watch the hyperscalers' September-quarter earnings in the October and November reporting round, including whether Meta puts a number on 2027 capex and whether Amazon's 2026 figure climbs beyond about $220 billion. Watch the public S-1 and pricing of the Anthropic IPO, then OpenAI's timetable. And watch Nvidia's next report, since a chipmaker feels any capex cut first.

None of this is investment advice. For US builders and operators, the practical goal is to stay flexible whichever way the data breaks.

• Keep a second model provider warm. Price and capacity terms can shift fast if a lab's financing tightens or a price war starts.
• Tie AI spend to a metric your finance team already tracks. PwC found that companies with strong AI foundations are the ones capturing financial returns.
• Treat run rates as marketing and GAAP revenue as accounting. Ask which one a vendor is quoting before you size a dependency on it.
• If you plan to raise capital in the next 12 months, model higher-for-longer rates. A 10-year yield near 5% changes what investors pay for growth.
• Separate a market correction from a technology reversal. Past bubbles left useful infrastructure behind, as the HBR analysis argues.

Sources