Are We in an AI Bubble? A 20-Minute Monday Checklist for Investors

The week the boom meets a rate rise#

The Federal Reserve's rate-setting committee sits down tomorrow morning and announces on Wednesday, 15 and 16 September. Until recently the argument was about how fast rates would come down. It isn't any more. The Bureau of Labor Statistics reported on Friday that consumer prices rose 0.4% in August and 3.4% over the year, and by the close futures markets were putting roughly a 90% chance on a rate rise this week, up from about 70% before the figures landed. The 10-year Treasury yield finished near 4.97%, its highest intraday level since October 2023, according to one market review of Friday's session.

Rates matter here for a specific reason. The single largest investment programme in the world right now, the build-out of AI data centres, has stopped being funded out of profits and started being funded out of borrowing. Oracle showed as much on Thursday evening, in results I'll come back to. Borrowing is getting more expensive.

Most people asking whether this is a bubble don't want a market call. They want to know whether their pension, their tracker fund or their savings are exposed to something they've never actually looked at. That question is answerable, not with a forecast but with four checks you can run yourself before the Fed speaks, using documents that cost nothing.

What a bubble is, and the terms you need to follow the argument#

A bubble isn't the same thing as an expensive market. It is a price that only makes sense if profits arrive that the businesses underneath are unlikely to deliver. That makes it a question of arithmetic.

The IMF ran the sums in its April 2026 Global Financial Stability Report. To justify prices at the time, the S&P 500 would have needed earnings growth of roughly 30% a year through 2027, and the Nasdaq around 35%. Analysts' own forecasts were nowhere near that. The distance between those two numbers is the whole debate, compressed into a line.

Three terms carry most of the weight in what follows. Capital expenditure, or capex, is money spent on long-lived physical things: buildings, chips, power connections. Free cash flow is what's left after that spending, and when capex exceeds it the difference has to come from borrowing or selling new shares. Remaining performance obligations, usually shortened to RPO or simply called backlog, is the value of contracts signed but not yet delivered.

None of this implies the technology is fake. The Bank for International Settlements, the institution central banks bank with, reviewed the evidence in its Annual Economic Report published on 29 June 2026 and found documented productivity gains of 20% to 50% in task-level studies. The argument is about price and financing, not about whether the technology works.

Check one: how much of the boom do you already own?#

Start with what you hold, not with what you read.

The Bank of England's July 2026 Financial Stability Report found that AI-related companies made up around half of the S&P 500's market value by June 2026, against roughly a quarter at the end of 2022. If you own a broad US index fund, or a global tracker, you own that proportion. It isn't a view you took. It arrived by default, through an index that weights companies by size.

That tends to surprise people who think of trackers as the cautious option. They remain the cheap and sensible way for most households to hold shares. But "diversified" describes the number of holdings rather than the number of bets, and on current weightings a US tracker is one large position in AI infrastructure with several hundred smaller ones attached. The Federal Reserve's May 2026 Financial Stability Report added that the equity risk premium, the extra return investors demand for owning shares rather than government bonds, was well below its historical average, with hedge fund leverage near record levels.

Open your fund's monthly factsheet, find the top ten holdings, add up the percentages. Five minutes, and more useful than any forecast.

Check two: who is actually paying for the data centres?#

For most of this boom, the big cloud companies paid for data centres out of operating cash flow. That has quietly stopped being true, and the shift is documented.

FactSet's analysis of Alphabet, Amazon, Meta, Microsoft and Oracle found incremental debt rose from 9% of capital spending in fiscal 2024 to 32% over the twelve months to mid-2026, with aggregate capex expected to pass $690 billion in FY26. The BIS put combined spending by the five largest hyperscalers at roughly $1 trillion across 2025 and 2026.

Oracle's first quarter, reported after the close on Thursday 10 September, shows what that looks like on one page. Revenue rose 30% to $19.3 billion, cloud infrastructure revenue more than doubled to $7.4 billion, and operating cash flow reached a record $23 billion. Free cash flow was still negative, at roughly minus $5 billion, because capital spending in the quarter came to $28 billion. Oracle guided to $90 billion to $95 billion of capex for the year, issued $20 billion of new shares during the quarter, and carries about $125 billion of debt on which quarterly interest expense rose 55% to $1.4 billion, according to a detailed breakdown of the results.

Where the debt sits matters as well. The Bank of England, citing OECD data, reports that the share of AI investment financed by private credit, meaning lending by investment funds rather than banks, rose from 9% in 2024 to 34% in 2025. Private credit discloses less and is repriced less often, so strain there shows up later.

Your version takes about a minute per company. Find the cash flow statement in any annual or quarterly report and compare the capital expenditure line with cash generated from operations. Growth funded from profits can be throttled back when conditions change. Growth funded by lenders and share issuance depends on those lenders and buyers still being there next year.

Check three: whose money sits on the other side of the contract?#

Backlog has become the headline number of this boom, and it deserves a harder look than it usually gets.

Oracle's remaining performance obligations reached $664 billion, up $209 billion in a year. Roughly half of that, by the same reporting, sits with one customer: OpenAI, itself loss-making and funded by outside capital. A signed contract is worth what the counterparty can pay across its full term.

The BIS flagged the pattern in June, describing a "complex web of private arrangements" including circular equity stakes and long-dated contracts, and noting that such terms are "typically poorly disclosed". Public deals show the shape of it. Nvidia agreed on 17 August 2026 to guarantee up to $105 billion for an OpenAI data centre campus in Pike County, Ohio, reported by Fortune as a sharp cut from the $250 billion figure that circulated in July. When the larger number first appeared, Nvidia shares fell about 4.5% intraday, which suggests investors already attach some risk to a supplier underwriting demand for its own products. That reading is interpretation, not fact.

One more figure from Nvidia is worth carrying into the next results season. In the quarter to 26 July 2026 the company reported revenue of $96.2 billion, up 106%, with $89.0 billion from data centres, filed with the SEC. Accounts receivable, money billed but not yet collected, stood at $63.1 billion against $38.5 billion at the previous year end. That can reflect demand booked late in the quarter, or customers paying more slowly. We won't know which until the next filing.

Underneath everything sits the slowest-moving number of all, which is whether businesses actually use the stuff. The US Census Bureau's Business Trends and Outlook Survey, summarised in May 2026, put AI use among American businesses at 17% to 20%, rising to 37% at firms with 250 or more employees and staying below 20% at the smallest, with no significant movement between December 2025 and May 2026.

Check four: what would you actually do if it fell?#

This is the check people skip, and it's the only one entirely within your control.

The Bank of England's stress scenario models a 45% fall in US equities over six quarters, credit spreads widening by 350 basis points, and UK GDP dropping 2.2 percentage points. It's a deliberately severe test of whether banks can absorb a shock, not a prediction of one, though it makes a useful number to hold a plan against.

Write that plan down now, while nothing is happening. How many months of spending do you hold outside the market? Will you need to sell shares in the next three years for something specific, such as school fees, a van, a deposit? If prices dropped 30% in a quarter, what would you actually do, as opposed to what you'd like to imagine you'd do?

Worth remembering too that bubbles don't only end in crashes. They can deflate through years of flat prices while earnings catch up, which is duller and harder to time. The plan is the same either way.

The checklist on one page#

Key takeaways#

1. No official body has declared an AI bubble. The Bank of England, the IMF and the Federal Reserve all describe valuations as stretched and the risk of a sharp correction as material, which is a statement about vulnerability rather than timing.
2. Ownership is the first check, because it's the one you control. AI-related companies are roughly half the S&P 500's value, so a standard tracker is more concentrated than it looks.
3. The financing mix has already shifted from profits to debt and equity issuance. Oracle spent $28 billion on capital projects in a single quarter and still ran negative free cash flow.
4. A backlog is a set of promises rather than money received. Around half of Oracle's $664 billion of signed contracts rests on one loss-making customer.
5. Adoption is the slow variable that decides the argument. Between 17% and 20% of US businesses use AI, and the smallest firms haven't moved since December 2025.

Frequently asked questions#

So are we in an AI bubble or not?

Nobody with the authority to say so has said so. What the official reports show is a market whose prices require earnings growth well above what analysts forecast, financed increasingly by debt. Whether that resolves through a fall in prices or a rise in earnings is not knowable in advance.

Is this article telling me to sell?

No. It doesn't know your age, income, debts or timescale, and none of it is investment advice. A regulated financial adviser is the right source for a decision about your own money.

I only own an index fund. Am I exposed?

Yes, to the extent that AI-linked companies make up around half of the S&P 500's value. That isn't an argument against index funds, which remain cheap and sensible for most households. It's an argument for knowing what's inside yours.

What happens if the AI bubble bursts?

The Bank of England's modelled scenario involves a 45% fall in US equities over six quarters and a 2.2 percentage point hit to UK GDP. That's a stress test designed to be severe, not a forecast. The honest answer is that the size and speed would depend on how much of the debt sits with lenders who are forced to sell.

Why does a rate rise matter to AI shares specifically?

Two reasons. Higher rates raise the cost of the borrowing now funding data centre construction, and they reduce the present value of profits expected far in the future, which is most of what an AI valuation rests on.

What's the cheapest signal to follow if I only track one thing?

The Census Bureau's Business Trends and Outlook Survey. It's free, published fortnightly, and measures what businesses actually do rather than what they expect.

Doesn't a $664 billion backlog prove demand is real?

It proves contracts were signed. Whether it proves demand depends on the customers' ability to pay across the full contract term, which is why concentration in a backlog matters as much as its size.

Glossary#

Capital expenditure (capex) Money spent acquiring or upgrading long-lived physical assets, here mostly data centres, chips and power infrastructure.

Free cash flow Cash from operations after capital spending. When negative, the shortfall is met by borrowing or issuing new shares.

Remaining performance obligations (RPO) The value of contracts signed but not yet delivered, often described as backlog. Revenue that is promised rather than received.

Basis point One hundredth of a percentage point. A 350 basis point move equals 3.5 percentage points.

Equity risk premium The extra return investors expect from shares over government bonds. A low premium means investors are accepting little compensation for taking risk.

Private credit Lending by non-bank institutions such as investment funds, usually outside public bond markets and with less public disclosure.

Hyperscaler An operator of very large cloud computing infrastructure, typically Alphabet, Amazon, Meta, Microsoft and Oracle.

Accounts receivable Money a company has billed customers but not yet collected. Growing faster than revenue, it can indicate slower payment.

References#

1. Bank of England, Financial Stability Report, July 2026.
2. International Monetary Fund, Global Financial Stability Report, April 2026.
3. Board of Governors of the Federal Reserve System, Financial Stability Report, May 2026 (Overview).
4. Board of Governors of the Federal Reserve System, FOMC meeting calendar, 2026.
5. Bank for International Settlements, Annual Economic Report 2026, published 29 June 2026.
6. US Bureau of Labor Statistics, Consumer Price Index, August 2026, released 11 September 2026.
7. Oracle Corporation, first quarter fiscal year 2027 earnings announcement date, results released 10 September 2026.
8. ERP Today, "Oracle Q1 FY27 results: $664B backlog, negative cash flow".
9. NVIDIA Corporation, results for the second quarter of fiscal 2027, filed with the US Securities and Exchange Commission.
10. Fortune, "OpenAI data center deal with Nvidia comes in $145 billion lower than reported", 18 August 2026.
11. US Census Bureau, "Large Firms With at Least 20 Employees Biggest AI Users", May 2026.
12. FactSet, "Hyperscalers Tap External Financing as AI Capex Outruns Cash Flow".
13. eOption, Market Review, 11 September 2026.