AI · Published · Updated

Is AI a bubble? What the spending, cash and revenue numbers say

The AI build-out is now bigger than the dot-com peak as a share of the US economy, more of it is borrowed, and money circles between suppliers and customers. Revenue is growing fast, but it is still small next to the commitments. The evidence, without the share prices.

This is general information, not financial advice. It does not take your circumstances into account. The value of investments can fall as well as rise and you may get back less than you put in. If you are unsure, speak to an adviser regulated where you live.

Most answers to "Is AI a bubble?" argue about share prices. This article looks at the other half: the real economy. How much is being spent, where the money comes from, how much the spending earns so far, and who is financing whom. Every figure comes from a company filing, a government statistics office or an official body such as the US Federal Reserve or the IMF.

The short version

  • The spending is historic. US business investment in IT equipment and software has reached 5.1% of the US economy in 2026, above the 4.4% of 2000, the peak of the dot-com boom.
  • More of it is borrowed. Long-term debt at Amazon, Alphabet and Meta has roughly tripled or more in a year. The US Federal Reserve counts more than $680bn of AI-related debt issued this year, against $130bn in 2023.
  • Money goes round in circles. The IMF found 14 deals in which AI companies invested in each other: about $300bn of shares and $650bn of debt, in exchange for $1.4 trillion of promised future revenue.
  • The revenue is growing fast, but it is small next to the commitments, and almost no company reports AI revenue as a separate figure.
  • Nobody can date a bubble. This article does not predict when or whether anything "bursts". It shows the signals to watch.

What "bubble" would mean here

A bubble, in plain terms, is spending and prices running far ahead of the income they can ever earn. The closest modern example is the telecom boom around 2000. Capital spending by US telecom companies went from $47bn in 1995 to $121bn in 2000, then "by 2002, investment had fallen by over half to $49 billion", according to a Federal Reserve Bank of San Francisco study, which describes the industry afterwards as "littered with the wrecks of overly optimistic expectations".

So the question for AI is not whether the technology is real. It is whether the money being spent now will earn enough, soon enough, to pay for itself and the debt behind it.

The spending is historic

The US Bureau of Economic Analysis measures business investment across the whole economy. Spending on IT equipment (servers, chips, networking) and software has climbed past its dot-com high:

US business investment in IT, share of GDP

  • IT equipment and software
  • IT equipment only
US business investment in IT, share of GDPUS business investment in IT, share of GDP. IT equipment and software: 1990 2.9, 1991 2.9, 1992 3, 1993 3.1, 1994 3.1, 1995 3.3, 1996 3.5, 1997 3.7, 1998 3.9, 1999 4.1, 2000 4.4, 2001 4, 2002 3.6, 2003 3.5, 2004 3.5, 2005 3.4, 2006 3.5, 2007 3.6, 2008 3.6, 2009 3.5, 2010 3.5, 2011 3.6, 2012 3.7, 2013 3.7, 2014 3.7, 2015 3.7, 2016 3.8, 2017 3.9, 2018 4, 2019 4, 2020 4.1, 2021 4.1, 2022 4.2, 2023 4, 2024 4.1, 2025 4.5, 2026 5.1; IT equipment only: 1990 2.2, 1991 2.1, 1992 2.2, 1993 2.2, 1994 2.3, 1995 2.5, 1996 2.5, 1997 2.6, 1998 2.6, 1999 2.7, 2000 2.9, 2001 2.5, 2002 2.2, 2003 2.1, 2004 2.1, 2005 2, 2006 2.1, 2007 2.1, 2008 2.1, 2009 1.9, 2010 2, 2011 2, 2012 2, 2013 2, 2014 2, 2015 2, 2016 1.9, 2017 2, 2018 2, 2019 1.9, 2020 1.9, 2021 1.9, 2022 1.9, 2023 1.7, 2024 1.8, 2025 2.1, 2026 2.6123456%1990199419982002200620102014201820222026IT equipment and software, 1990: 2.9 %IT equipment and software, 1991: 2.9 %IT equipment and software, 1992: 3 %IT equipment and software, 1993: 3.1 %IT equipment and software, 1994: 3.1 %IT equipment and software, 1995: 3.3 %IT equipment and software, 1996: 3.5 %IT equipment and software, 1997: 3.7 %IT equipment and software, 1998: 3.9 %IT equipment and software, 1999: 4.1 %IT equipment and software, 2000: 4.4 %IT equipment and software, 2001: 4 %IT equipment and software, 2002: 3.6 %IT equipment and software, 2003: 3.5 %IT equipment and software, 2004: 3.5 %IT equipment and software, 2005: 3.4 %IT equipment and software, 2006: 3.5 %IT equipment and software, 2007: 3.6 %IT equipment and software, 2008: 3.6 %IT equipment and software, 2009: 3.5 %IT equipment and software, 2010: 3.5 %IT equipment and software, 2011: 3.6 %IT equipment and software, 2012: 3.7 %IT equipment and software, 2013: 3.7 %IT equipment and software, 2014: 3.7 %IT equipment and software, 2015: 3.7 %IT equipment and software, 2016: 3.8 %IT equipment and software, 2017: 3.9 %IT equipment and software, 2018: 4 %IT equipment and software, 2019: 4 %IT equipment and software, 2020: 4.1 %IT equipment and software, 2021: 4.1 %IT equipment and software, 2022: 4.2 %IT equipment and software, 2023: 4 %IT equipment and software, 2024: 4.1 %IT equipment and software, 2025: 4.5 %IT equipment and software, 2026: 5.1 %IT equipment only, 1990: 2.2 %IT equipment only, 1991: 2.1 %IT equipment only, 1992: 2.2 %IT equipment only, 1993: 2.2 %IT equipment only, 1994: 2.3 %IT equipment only, 1995: 2.5 %IT equipment only, 1996: 2.5 %IT equipment only, 1997: 2.6 %IT equipment only, 1998: 2.6 %IT equipment only, 1999: 2.7 %IT equipment only, 2000: 2.9 %IT equipment only, 2001: 2.5 %IT equipment only, 2002: 2.2 %IT equipment only, 2003: 2.1 %IT equipment only, 2004: 2.1 %IT equipment only, 2005: 2 %IT equipment only, 2006: 2.1 %IT equipment only, 2007: 2.1 %IT equipment only, 2008: 2.1 %IT equipment only, 2009: 1.9 %IT equipment only, 2010: 2 %IT equipment only, 2011: 2 %IT equipment only, 2012: 2 %IT equipment only, 2013: 2 %IT equipment only, 2014: 2 %IT equipment only, 2015: 2 %IT equipment only, 2016: 1.9 %IT equipment only, 2017: 2 %IT equipment only, 2018: 2 %IT equipment only, 2019: 1.9 %IT equipment only, 2020: 1.9 %IT equipment only, 2021: 1.9 %IT equipment only, 2022: 1.9 %IT equipment only, 2023: 1.7 %IT equipment only, 2024: 1.8 %IT equipment only, 2025: 2.1 %IT equipment only, 2026: 2.6 %
Show the numbers
Values (%)IT equipment and softwareIT equipment only
19902.92.2
19912.92.1
199232.2
19933.12.2
19943.12.3
19953.32.5
19963.52.5
19973.72.6
19983.92.6
19994.12.7
20004.42.9
200142.5
20023.62.2
20033.52.1
20043.52.1
20053.42
20063.52.1
20073.62.1
20083.62.1
20093.51.9
20103.52
20113.62
20123.72
20133.72
20143.72
20153.72
20163.81.9
20173.92
201842
201941.9
20204.11.9
20214.11.9
20224.21.9
202341.7
20244.11.8
20254.52.1
20265.12.6
Private fixed investment in information processing equipment and in software, as a share of gross domestic product; current dollars, each year the average of its quarters. 2026: first 3 quarters only. Excludes data-centre buildings and power. US Bureau of Economic Analysis national accounts. Source: US Bureau of Economic Analysis, National Income and Product Accounts, Tables 1.1.5 and 5.3.5 (series Y034RC, B985RC, A191RC), checked .
  • In 2000, at the height of the dot-com boom, it was 4.4% of GDP. After the crash it fell back to about 3.5%.
  • In 2026 so far it is 5.1%, about $1.7 trillion a year.
  • Equipment alone is 2.6% of GDP, almost back to its 2000 level of 2.9%.

The buildings come on top. Spending on constructing data centres, which the US Census Bureau tracks separately, has gone from $1.8bn in 2014 to $49.7bn in 2025, and was running at an annual rate of $96.5bn in September 2026:

US spending on building data centres

US spending on building data centresUS spending on building data centres. Data-centre construction: 2014 1.8, 2015 2.7, 2016 4.1, 2017 4.7, 2018 6.9, 2019 8.5, 2020 9.2, 2021 9.9, 2022 12.6, 2023 20, 2024 34.7, 2025 49.7, 2026 70.9020406080$bn2014201520162017201820192020202120222023202420252026Data-centre construction, 2014: 1.8 $bnData-centre construction, 2015: 2.7 $bnData-centre construction, 2016: 4.1 $bnData-centre construction, 2017: 4.7 $bnData-centre construction, 2018: 6.9 $bnData-centre construction, 2019: 8.5 $bnData-centre construction, 2020: 9.2 $bnData-centre construction, 2021: 9.9 $bnData-centre construction, 2022: 12.6 $bnData-centre construction, 2023: 20 $bnData-centre construction, 2024: 34.7 $bnData-centre construction, 2025: 49.7 $bnData-centre construction, 2026: 70.9 $bn
Show the numbers
Values ($bn)Data-centre construction
20141.8
20152.7
20164.1
20174.7
20186.9
20198.5
20209.2
20219.9
202212.6
202320
202434.7
202549.7
202670.9
Value of private construction put in place, data centres, average of the monthly seasonally adjusted annual rates for each year; 2026 covers January to September only. In September 2026 the annual rate was $96.5bn (preliminary). Buildings only - the servers and chips inside are counted separately, as business investment in equipment. US Census Bureau, Construction Spending. Source: US Census Bureau, Construction Spending (Value of Construction Put in Place), private, seasonally adjusted annual rate, data centers, checked .

These are US figures in US dollars. Our article How big is the AI capex wave? shows the same build-out company by company.

It is starting to be borrowed

For most of the boom, the largest tech companies paid for data centres out of the cash their businesses generate. In 2026 that changed. Alphabet and Amazon are spending more on property and equipment than their operations bring in:

Capital spending as a share of cash from operations

  • Microsoft
  • Alphabet
  • Amazon
  • Meta
Capital spending as a share of cash from operationsCapital spending as a share of cash from operations. Microsoft: 2021 26.9, 2022 26.8, 2023 32.1, 2024 37.5, 2025 47.4, This year so far 67.9; Alphabet: 2021 26.9, 2022 34.4, 2023 31.7, 2024 41.9, 2025 55.5, This year so far 103.8; Amazon: 2021 131.8, 2022 136.1, 2023 62.1, 2024 71.6, 2025 94.5, This year so far 154.3; Meta: 2021 32.4, 2022 61.8, 2023 38, 2024 40.8, 2025 60.2, This year so far 85.7050100150200%20212022202320242025Thisyear sofarMicrosoft, 2021: 26.9 %Microsoft, 2022: 26.8 %Microsoft, 2023: 32.1 %Microsoft, 2024: 37.5 %Microsoft, 2025: 47.4 %Microsoft, This year so far: 67.9 %Alphabet, 2021: 26.9 %Alphabet, 2022: 34.4 %Alphabet, 2023: 31.7 %Alphabet, 2024: 41.9 %Alphabet, 2025: 55.5 %Alphabet, This year so far: 103.8 %Amazon, 2021: 131.8 %Amazon, 2022: 136.1 %Amazon, 2023: 62.1 %Amazon, 2024: 71.6 %Amazon, 2025: 94.5 %Amazon, This year so far: 154.3 %Meta, 2021: 32.4 %Meta, 2022: 61.8 %Meta, 2023: 38 %Meta, 2024: 40.8 %Meta, 2025: 60.2 %Meta, This year so far: 85.7 %
Show the numbers
Values (%)MicrosoftAlphabetAmazonMeta
202126.926.9131.832.4
202226.834.4136.161.8
202332.131.762.138
202437.541.971.640.8
202547.455.594.560.2
This year so far67.9103.8154.385.7
Cash spent on property and equipment divided by net cash provided by operating activities, both as reported to the SEC. Above 100% means a company spent more on equipment than its operations brought in, before any share buybacks or dividends. "This year so far" is January to September 2026 for Alphabet, Amazon and Meta, and July to September 2026 for Microsoft, whose year ends 30 June. Source: Microsoft Form 10-K filed 29 July 2026 (SEC EDGAR), checked .
Show all 7 sources

The gap is being filled with debt:

Long-term debt: a year earlier and now

  • A year earlier
  • Latest quarter
Long-term debt: a year earlier and nowLong-term debt: a year earlier and now. Amazon: A year earlier 50.7, Latest quarter 173.1; Alphabet: A year earlier 21.6, Latest quarter 125.1; Meta: A year earlier 28.8, Latest quarter 80.9; Microsoft: A year earlier 35.4, Latest quarter 31.1; NVIDIA: A year earlier 8.5, Latest quarter 32.4$bn050100150200AmazonA year earlier, Amazon: 50.750.7Latest quarter, Amazon: 173.1173.1AlphabetA year earlier, Alphabet: 21.621.6Latest quarter, Alphabet: 125.1125.1MetaA year earlier, Meta: 28.828.8Latest quarter, Meta: 80.980.9MicrosoftA year earlier, Microsoft: 35.435.4Latest quarter, Microsoft: 31.131.1NVIDIAA year earlier, NVIDIA: 8.58.5Latest quarter, NVIDIA: 32.432.4
Show the numbers
Values ($bn)A year earlierLatest quarter
Amazon50.7173.1
Alphabet21.6125.1
Meta28.880.9
Microsoft35.431.1
NVIDIA8.532.4
Debt due after more than a year, as each company reports it on its balance sheet. Dates: Amazon Sep 2025 and Sep 2026; Alphabet Sep 2025 and Sep 2026; Meta Sep 2025 and Sep 2026; Microsoft Sep 2025 and Sep 2026; NVIDIA Jul 2025 and Jul 2026. From the companies' SEC filings. Source: Amazon 10-Q filed 2026-10-30 (SEC EDGAR), checked .
Show all 5 sources
  • Amazon's long-term debt rose from $50.7bn to $173.1bn in a year, Alphabet's from $21.6bn to $125.1bn, Meta's from $28.8bn to $80.9bn and NVIDIA's from $8.5bn to $32.4bn. Microsoft's fell.
  • The US Federal Reserve (Financial Stability Report, 13 November 2026): "Advanced computing-related debt issuance rose to more than $680 billion by October 2026, about a fivefold increase from $130 billion in all of 2023." Such borrowing went from less than 6% of non-financial investment-grade company debt issuance in 2023 to about 31% in 2026.
  • The Bank for International Settlements: "Gross bond issuance this year by the five largest hyperscalers reached $222 billion by September 2026."
  • The Fed compares the pace with the telecom boom: it "may surpass that of the 1990s telecom cycle relatively soon".

Borrowing is not a problem in itself, and these companies are highly profitable. But debt has to be repaid whether or not the AI revenue arrives, which is what makes a downturn more painful.

Who funds whom

A feature of this boom is how much of the money circles between suppliers and customers:

  • The IMF (12 October 2026): "Our review of 14 such transactions since 2023 found about $300 billion in equity investments and $650 billion in debt, in exchange for $1.4 trillion in future revenue commitments." If the companies receiving the money struggle, the IMF warns, the value of those deals and the debt in them falls together.
  • NVIDIA, which sells the chips, has agreed to provide credit support "capped at a total of $105 billion" for a data-centre build-out for an affiliate of OpenAI, one of its customers.
  • Anthropic, an AI lab now listing its shares, says in its registration statement that it has entered into commitments "with an aggregate committed spend of approximately $900.9 billion, substantially all of which is related to compute and infrastructure". Alongside its listing, $26.0bn of its shares are being bought by companies that make or sell the chips and computing power AI runs on: Google, NVIDIA, AMD, Broadcom, Micron, Microsoft and TSMC.

When suppliers fund their own customers, sales can grow faster than the end demand behind them. That does not make the sales false, but it means a problem at one company can spread to several.

The revenue so far

The revenue is real and growing quickly:

  • Google Cloud revenue was $30.7bn in July to September 2026, up 103% on a year earlier.
  • Anthropic made $33.2bn of revenue in January to September 2026, up from $4.6bn in all of 2025.
  • Microsoft recorded $10.0bn of revenue from its arrangements with OpenAI in July to September 2026.

But three things keep the question open:

  • Almost no company reports "AI revenue" as a number. Cloud growth, backlogs and usage statistics are given; how much of that is AI, and how profitable it is, mostly is not.
  • The commitments dwarf the revenue. Anthropic's $900.9bn of commitments is more than 25 times its revenue for the first nine months of 2026. Anthropic itself says: "We do not expect our revenue will continue to grow at comparable rates."
  • The payoff in the wider economy is small so far. About a quarter of US firms use AI, and the Bank of England estimates AI has added about 0.1 to 0.2 percentage points a year to UK productivity growth. Our article Better models or better use? has the detail.

What the watchdogs say

  • US Federal Reserve (November 2026): in its survey of market contacts - whose views, the report notes, are not the Fed's own - "AI was the top-cited risk". Most respondents pointed to "the possibility of a sharp correction in technology-related asset valuations" and "complex financing arrangements for data center investment".
  • IMF (October 2026): stress at key AI companies "could lower revenue forecasts and investment valuations", which "would weaken the value of the accumulated debt in those deals".
  • Bank for International Settlements: AI investment is "increasingly built on credit".

None of them says a crash is coming. All of them say the risks have grown.

The verdict

This is our reading of the evidence, not a forecast and not a view on any company's shares.

The AI build-out has the size and the financing pattern of past booms that ended badly, but not yet the proof that this one will. It is historic in scale, increasingly paid for with debt and with money that circles between suppliers and customers, and the revenue that would justify it is growing fast but is still small and mostly undisclosed.

  • What looks like a bubble: spending above the dot-com peak as a share of the economy; debt rising quickly; commitments many times current revenue; suppliers financing customers.
  • What does not: the biggest spenders are highly profitable companies, not start-ups; AI revenue is real and growing very fast (Anthropic's rose 1,206% on a year earlier, by its own account); and the official bodies describe risks, not a collapse.

The honest answer to "is AI a bubble?" is: it has the ingredients of one, and the next year or two of revenue will decide whether it becomes one.

What would change this reading:

  • Towards "bubble": capital spending plans cut sharply, AI revenue growth slowing while debt keeps rising, or a large AI company failing to meet its commitments.
  • Away from it: companies starting to report AI revenue and profits that match the spending, and free cash flow recovering as the new data centres fill up.

What to watch

  • 2027 spending plans from Microsoft, Alphabet, Amazon and Meta with their next results (late January and early February 2027).
  • Free cash flow and new debt in the same results; the Fed's next Financial Stability Report (usually published in spring).
  • Anthropic's first results as a listed company, the first time an AI lab's accounts appear quarter by quarter.
  • NVIDIA's results on 17 November 2026.
  • Data-centre construction, published monthly by the US Census Bureau.

What these numbers do and do not show

  • Investment and construction figures are for the US only, in current US dollars. The investment share counts IT equipment and software across the whole US economy, not only AI; the construction figure counts buildings, not the chips inside them.
  • Debt is long-term debt as each company reports it. Leases and other commitments not on the balance sheet are extra, and the Fed notes they are not directly comparable with debt.
  • The IMF and Fed figures are their own estimates, from their own data sources.
  • No share prices or valuations appear in this article on purpose: they are a separate question, and we do not publish market calls.

Disclosure: this article was drafted with help from Claude, an AI model made by Anthropic, which is one of the companies discussed above. Anthropic has no part in what we publish; every fact here is taken from the sources listed under Sources.

Drafted with help from an AI tool, then checked and edited by a person. Sources are listed below.

Sources (21)

Show all 21 sources