Big tech companies are spending hundreds of billions of dollars on artificial intelligence, and for almost four years people have asked whether the stock market is inflating a bubble.
Here are the numbers to judge it and what the main players have said, each with its date and source.
In late July the Nasdaq-100, the index of the hundred largest Nasdaq companies, fell 10% from its peak. On October 6 the S&P 500 hit a new record high. In the same days the yield on ten-year US Treasuries reached its highest level since 2002. Borrowing costs more, including for companies taking on debt to build data centres. NBC News, July 28, 2026 ↗ · Axios, October 7, 2026 ↗ · Forbes, October 7, 2026 ↗
Prices
41.6
The S&P 500 price divided by average inflation-adjusted earnings over the past ten years (Shiller CAPE). The record is 44.2, in December 1999; the long-run average is 17.4.
The share of S&P 500 returns, earnings growth and capital spending that has come since January 2024 from 42 AI-related stocks, less than 10% of the index.
What Amazon, Alphabet, Meta and Microsoft expect to invest in 2026, according to the guidance they gave in late July.
Fortune, W.Media, CFO Dive, Meta · July 2026 · details →
Cash
−$5.9bn
Between April and June 2026 Alphabet spent more on investment than its businesses generated in cash. Amazon is negative over twelve months, at −$7.6bn.
“I am seeing the first signs of what feels to me like unsustainable silliness of random new neoclouds popping up, people claiming that they’re going to build gigantic amounts of compute next year that I think they don’t have the revenue to support or a buyer.”
Neoclouds are companies that rent out AI servers. He says he is not worried about OpenAI’s own plans. In August 2025 he had said investors were “overexcited” about AI.
How much the companies investing in AI are earning
Microsoft, Alphabet, Meta and Amazon close every quarter with a profit and invest unprecedented sums. None publishes separate accounts for AI, so how much AI earns on its own remains unknown.
Revenue and operating profit, April-June 2026
Whole companies, all businesses included. Operating profit is what is left of revenue after running costs, before interest and taxes.
Microsoft
Revenue
$90 bn
− Operating costs
$49.4 bn
= Operating profit
$40.6 bn
Out of every $100 of revenue, 45 remain as operating profit.
Profit and cash are different measures: a data centre is paid for up front, while its cost is spread over many years in the accounts. Free cash flow is the money left after paying for investment. When it is negative, the company spent more than its businesses brought in.
Microsoft
$19.6 bn
Apr–Jun 2026 · quarter
Whole company. Operating cash flow less cash paid for property, plant and equipment.
Microsoft went from $190bn to about $175bn because of an accounting change; the spending plan stayed the same.
Who sells the chips: Nvidia
Nvidia sells the chips other companies use to fill their AI data centres. In the quarter ended July 26, 2026 its revenue more than doubled year on year (+106%).
Sources: company quarterly releases and financial press, checked October 9, 2026. Values rounded.
03 · Indicators
The signs of a bubble, and the signs against
Prices, concentration and debt show how exposed the market is to a fall. Earnings and data-centre use show how real demand is. None of these numbers says when a fall might come.
Signs of a bubble
60-80%
42 AI-related stocks account for most of the US market’s growth
Since January 2024 they have produced 60-80% of S&P 500 returns, earnings growth and capital spending, while making up less than 10% of the index. Meanwhile about 60% of S&P 500 stocks are more than 20% below their all-time high; normally the share is 20-40%.
The Shiller CAPE divides the S&P 500 price by average inflation-adjusted earnings over the past ten years. On October 8, 2026 it stood at 41.6. The record is 44.2, in December 1999, a few months before the dot-com crash; the long-run average is 17.4.
Morgan Stanley estimates about $250bn of AI-related debt in the first half of 2026 and $500bn for the full year. The bank rates the issuers as solid: the risk, in its view, is the sheer amount of debt the market has to absorb. Its base case is 1997-1998, two years before the dot-com peak.
According to Goldman Sachs, US share prices have grown together with company earnings, as well as with what investors pay for each dollar of profit. That is why it sees a lower risk than in 1999 of a bubble made of prices alone. The remaining risk is overestimating how long such high earnings will last.
In December 2025 only 1.6% of data-centre capacity was vacant, and three quarters of the capacity under construction was already leased. In a mature bubble, capacity usually sits empty. For J.P. Morgan a bubble was more likely still forming than already at its peak.
How much AI models are improving in independent tests
Investors are betting that models will keep improving. Two independent organisations, METR and ARC Prize, measure them with public tests. Neither measures intelligence in general.
METR · Time Horizon 1.1
How long the tasks an AI completes on its own can be
METR gives AI agents programming, machine learning and cybersecurity tasks, and measures them by the time a human expert would need. A one-hour bar means a task worth one hour of human work.
Tasks the agent is predicted to complete 5 times out of 10. Longer bars mean harder tasks.
Human expert time · linear scale
08 h16 h
GPT-44 min
Claude 3.5 Sonnet (Jun 2024)11 min
Claude 3.7 Sonnet1 h 0 min
o32 h 0 min
Claude Opus 4.54 h 53 min
Claude Opus 4.611 h 59 min
GPT-5.3 Codex5 h 50 min
Gemini 3.1 Pro6 h 24 min
GPT-5.45 h 42 min
Claude Mythos Preview (early)17 h 25 min *
* Estimates above 16 hours are unreliable; bars stop at 16 h. Selection of 10 models, ordered by release date. Same TH 1.1 benchmark.
Uncertainty and exact values
Estimates have uncertainty: the ranges below are the intervals supplied by METR. Model and agent tools are evaluated together.
GPT-4 · 2023-03-14 2 min – 8 min
Claude 3.5 Sonnet (Jun 2024) · 2024-06-20 5 min – 22 min
Claude 3.7 Sonnet · 2025-02-24 33 min – 1 h 44 min
o3 · 2025-04-16 1 h 15 min – 3 h 11 min
Claude Opus 4.5 · 2025-11-24 2 h 42 min – 10 h 24 min
Claude Opus 4.6 · 2026-02-05 5 h 17 min – 60 h 34 min
GPT-5.3 Codex · 2026-02-05 3 h 15 min – 13 h 36 min
Gemini 3.1 Pro · 2026-02-19 3 h 54 min – 11 h 35 min
GPT-5.4 · 2026-03-05 3 h 7 min – 12 h 49 min
Claude Mythos Preview (early) · 2026-04-07 8 h 29 min – 55 h 4 min
METR data updated May 8, 2026 · checked October 9, 2026
ARC Prize
How many never-seen puzzles it can solve
In ARC-AGI-1 and 2 the model sees coloured grids transformed by a hidden rule: it has to work out the rule and apply it to a new grid. In ARC-AGI-3 it has to learn on its own how to win video games it has never seen.
Percentage of grid puzzles solved correctly. 100% means every puzzle in this test, not every possible reasoning problem.
Puzzles solved · cost per puzzle
0%50%100%
GPT-6 Astra (Max)95%
$1.12 / puzzle
GPT-6.1 Sol (Max)94.17%
$0.25 / puzzle
Claude Opus 5.5 (High)93.33%
$0.41 / puzzle
GPT-5.6 Sol (Max)92.5%
$1.44 / puzzle
Claude Opus 5 (Max)90.42%
$2.06 / puzzle
Claude Fable 5.1 (XHigh)90%
$3.12 / puzzle
GPT-6 Sol (Max)89.58%
$0.44 / puzzle
Gemini 3.8 Flash (High)89.17%
$0.4 / puzzle
Top 8 model groups: highest-scoring configuration for each, with its cost. Semi-private evaluation set.
The three versions are different tests, to be read separately. On ARC-AGI-1 the best models reach 98.5%, just above the 98% scored by a panel of people.
The curves are schematic and the markets differ: the percentages describe the past and do not give the probability of an AI crash. Sources for each episode are above.
How this page is made
The newsroom updates this page by hand. Every number carries its source and the period it refers to, and a quote is added only after being read word for word at the linked source. Translations are ours.
