UrsaKit LogoUrsaKit
UrsaKit cover asking whether the AI bubble will burst and which signals are worth watching

On 9 October 2026, a revenue memo did what two years of think pieces had not: it put the phrase "AI bubble burst" into the same sentence as a falling Nasdaq. Futurism reported that internal memos sent to OpenAI investors in late September pointed to annualised revenue approaching $50 billion, roughly $20 billion below the $70 billion figure that had leaked earlier. The Nasdaq closed down 1.25% that Thursday and Nvidia fell nearly 3%.

The concept itself is not new. Wikipedia's AI bubble entry traces it to a stock market bubble growing since 2025, and the bond market came to dominate how the build-out is financed — alongside roughly $3 trillion in off-balance-sheet commitments. Whether that ends in a crash is a different question from whether the spending is real, and the difference is worth understanding before you act on a headline.

01What the phrase "AI bubble" actually claims

A bubble is a price pattern, not a moral judgement: valuations rise because buyers expect growth, and the price depends on that growth continuing. The AI version adds one specific complaint. Investment is flowing between the same small group of companies — chip supplier, model developer, cloud provider — in ways that inflate reported demand as well as reported revenue.

That circular pattern is why the AI debate sounds different from an ordinary tech rally. Nvidia's letter of intent with OpenAI, announced in September 2025, commits up to $100 billion of investment in OpenAI, paid progressively as each gigawatt of Nvidia hardware is deployed — hardware that OpenAI buys from Nvidia. Money and orders travel in a loop, and a loop is harder to read than a normal customer relationship.

02The money going in: $1.4 trillion, 10 gigawatts, one $300 billion contract

Three numbers describe the scale. OpenAI has committed to spending about $1.4 trillion over eight years on data centres, much of it built around the 10 gigawatt Nvidia partnership. TechCrunch reported a $300 billion compute contract with Oracle, spread over about five years and starting in 2027, on top of the $500 billion Stargate programme with SoftBank and Oracle.

That spending is no longer funded mainly from cash flow. Early AI investment came out of operating profits; by 2026 the bond market had become the dominant source, which is the part that makes a slowdown contagious rather than contained.

WhoWhat the number saysWhere it comes from
OpenAI~$1.4 trillion of data centre commitments over 8 yearsLeaked and reported financials summarised on Wikipedia's AI bubble page
Nvidia + OpenAIUp to $100 billion of Nvidia investment, 10 GW of systemsNvidia newsroom, 22 September 2025
Oracle + OpenAI~$300 billion of compute over ~5 years from 2027TechCrunch, 10 September 2025
Data centres overall~$3 trillion of global spending 2025–2028, half via private creditMorgan Stanley estimate, cited by Wikipedia

Committed spending is not the same as money already spent

Data centre commitments are multi-year obligations. They can be renegotiated, delayed or cancelled — which is exactly why a fall in expected revenue shows up in share prices before it shows up in cancelled construction.

03The money coming out: the revenue gap at the centre of the story

The build-out is sized to a revenue curve, and that curve keeps moving down. In late September 2026, memos to OpenAI investors pointed to revenue approaching $50 billion instead of the $70 billion that had been circulating, a gap of roughly $20 billion. The market read it as a demand signal rather than an accounting detail.

The same week, Bain & Company's 2026 technology report put a number on the other side of the ledger. The AI industry would need to reach roughly $6 trillion in annual revenue by 2031 to justify the spending on data centres, according to Futurism's report. That is a benchmark for the whole industry, not a prediction, but it makes the $50 billion question look small.

There is also a productivity puzzle underneath. Wikipedia's summary cites a National Bureau of Economic Research study published in February 2026 in which 90% of firms reported no measurable impact from AI on workplace productivity, while executives still projected a 1.4% productivity gain. Optimism in the boardroom is not the same as a return in the accounts, and investors eventually price the difference.

04What central banks and other markets are already saying

The Bank of England has been warning about this for a year. In December 2025, BBC News reported that the Bank's financial stability report flagged a possible "sharp correction" in the value of major tech companies. It called AI-focused valuations "particularly stretched" and compared US equity levels to the period before the dot-com crash. The Bank's own analysis, All chips in!, works through what a fall in AI-related asset values would mean for the wider financial system.

Korea showed what a concentration problem looks like in practice. In late June and July 2026 the KOSPI fell 44% in 40 days and erased about $2.18 trillion in market value as Samsung Electronics and SK Hynix sold off. Those two companies had made up more than half the index. The trigger was doubt about short-term returns on AI infrastructure and memory demand.

For a sense of how much is riding on a handful of names, look at the concentration figures from late 2025. The five largest US companies accounted for around 30% of the S&P 500 and 20% of the MSCI World index, the highest concentration in half a century. AI-related companies were responsible for roughly 80% of US market gains that year.

05The strongest arguments that this is not a bubble

The bear case is not uncontested, and the counter-arguments are specific rather than hopeful. Goldman Sachs's chief equity strategist, Peter Oppenheimer, has argued that the gains in large-cap AI companies are backed by actual profit growth, and that forward price-to-earnings ratios remain well below dot-com levels. Morgan Stanley analysts have called bubble fears premature, pointing to cash flow and capital reserves at the largest US firms that are roughly three times higher than in previous bubble periods.

JPMorgan went further and concluded that AI does not meet the classic criteria for a financial bubble, using a five-factor diagnostic. Federal Reserve chair Jerome Powell has drawn a similar distinction, noting that AI companies generate real revenue and that data centre spending contributes to broader growth rather than merely to share prices.

Both sides can be right at the same time: the companies may be profitable while the surrounding capital structure is fragile. That is why the next section is about plumbing rather than prediction.

06What a burst would actually look like

The likeliest version of a burst is a sequence, not an event. It starts with revenue guidance that falls short of plan, then meets a cost that cannot be cut quickly — multi-year contracts, already-ordered hardware, debt raised against future cash flow.

StageWhat you would seeWhy it spreads
\1. Expectations resetGuidance below the plan, share prices fall firstMarkets price future cash flow, not current construction
\2. Spending freezesNew orders pause, delayed sites are shelvedSuppliers lose revenue before they lose staff
\3. Credit feels itBonds issued by data centre operators repriceMorgan Stanley estimated data centre debt could exceed $1 trillion by 2028, much of it BBB or junk rated
\4. Local effectsConstruction slows, power contracts and tax deals get re-examinedCommunities absorb both the building boom and the bust

The debt layer is what turns a correction into a crisis. Private credit covered roughly half of the $3 trillion of data centre spending projected between 2025 and 2028, according to a Morgan Stanley estimate summarised on Wikipedia, and a meaningful share of the bonds involved sits at BBB or below. Lenders, not just shareholders, end up holding the risk.

Meanwhile the companies at the centre of it are planning for a public reaction. Axios reported on 9 October 2026 that executives at Anthropic, OpenAI and other AI firms are privately gaming out "day after" scenarios. They expect a political and public revolt after a catastrophic AI event, most likely a cyberattack that interrupts financial services, connectivity or utilities. Whether or not that risk materialises, a company rehearsing the aftermath is describing how much of its value depends on public permission.

07How to think about your own exposure

Most readers will not decide whether the bubble bursts. They will decide whether they understood what they owned. Two questions make that concrete: how much of your retirement or index money sits in the same handful of companies, and how much of a drop you could tolerate without changing your plan.

The arithmetic of a drawdown is asymmetric, which is worth doing once with real numbers:

const drops = [0.2, 0.3, 0.4];
for (const drop of drops) {
  const needed = 1 / (1 - drop) - 1;
  console.log(`drop ${(drop * 100).toFixed(0)}% -> need +${(needed * 100).toFixed(2)}% to get back`);
}

The output is the reason a 30% fall feels so much worse than a 30% rally feels good:

drop 20% -> need +25.00% to get back
drop 30% -> need +42.86% to get back
drop 40% -> need +66.67% to get back

Run against your own numbers, that calculation is a stress test rather than a forecast. If the answer would change your plan, the plan is the part worth revisiting — not the headline that prompted the question.

Watch the funding line, not the vocabulary

The words "bubble" and "correction" generate noise. The two numbers that move first are revenue guidance from the companies buying compute, and the price of the debt funding it.

08Sources

Frequently asked questions

Is AI a bubble?

Smart money disagrees. Goldman Sachs and JPMorgan argue that today's AI leaders earn real revenue and that valuations are below dot-com levels, while the Bank of England and the IMF warn that AI-focused valuations are stretched and could correct sharply. A useful position is to treat the bubble question as a set of measurable signals rather than a label.

When will the AI bubble burst?

Nobody knows, and anyone giving you a date is guessing. The observable triggers to watch are the ones described in this article: revenue guidance that keeps falling below the financing plan, debt that has to be refinanced into a cooler market, and capital spending that gets cut after it has already been committed.

Why did OpenAI's revenue forecast move the market so much?

Futurism reported that internal memos sent to investors in late September 2026 pointed to revenue approaching $50 billion rather than the $70 billion that leaked earlier, a gap of roughly $20 billion. That matters because the entire build-out is sized to a revenue curve; the same week, the Nasdaq fell 1.25% and Nvidia fell nearly 3%.

Does the AI bubble affect people who own no tech shares?

It usually reaches them indirectly. Index funds hold the same five companies that dominate the S&P 500, suppliers depend on AI orders for jobs and wages, and data centers affect local power and water. The exposure is smaller and slower, but it is not zero.

What is the $6 trillion number?

It comes from Bain & Company's 2026 technology report, which calculated that the AI industry would need to reach roughly $6 trillion in annual revenue by 2031 to justify the amount being spent on AI data centers. It is a benchmark, not a forecast.

How much gain does it take to recover from a 30% drop?

About 42.9%. A portfolio that falls 30% has to rise 42.86% just to return to where it started, because the recovery is calculated on the smaller base. That asymmetry is the practical reason drawdowns deserve more attention than rallies.

What is the most useful thing a non-investor can do?

Check your indirect exposure and avoid decisions based on a headline. Look at what your pension or index fund holds, then decide whether that concentration matches the risk you thought you had taken.

Back to articles

Related articles

Related tools