On Thursday 5 November the Nasdaq fell 5.5 per cent and the S&P 500 fell 4.4 per cent, the worst day for American shares since the tariff shock of April 2025. Advanced chip makers, Nvidia lost 9.3 per cent and AMD 17.3 per cent. The Nasdaq closed 11.9 per cent below the record it set in October, which is a correction by the usual definition. On the same day the Bank of England raised its main rate for the first time in three years, and the FTSE 100 slipped 0.4 per cent. Two markets, two stories.

On the left the argument that followed has mostly taken one form. Either AI really raises productivity, in which case the firms are sound and this is a wobble, or it doesn’t, in which case it is a gold rush and the bubble is bursting. Both sides argue as if the answer to one question settles the other.

It doesn’t. Whether a technology works and whether the money invested in it is mispriced are separate questions, and the evidence on each points in a different direction. This article tries to set them apart, because the argument goes better once they are.

It is the same mistake that Red Mole Guide No. 16, The Fourth Forecast, traces through seventy years of the Fourth International’s press: treating automation as a forecast about a machine, and then deducing a politics from it. The movement made that forecast in 1954, 1962 and 1984, each time with good data, and each time the thing that decided what happened to workers was not in the forecast.

1. A bubble is about prices, not about whether the machine works

The great investment manias were nearly all about technologies that worked. The British railway mania of the 1840s ruined thousands of investors, and it left Britain with railways. The telecoms boom of the late 1990s ended in bankruptcies, and it left the fibre that the internet still runs on, most of it laid years before anyone needed it. Cisco was a profitable company in 2000, selling real equipment to real customers, and its shares lost most of their value over the next two years.

So “AI raises productivity” is no answer to “this is a bubble”, and “this is a bubble” is no answer to “AI raises productivity”. A bubble is a judgement on whether the price paid today for a claim on future profits can be met, and on who has borrowed to bet that it will be. Marx called share capital of this kind fictitious capital. The name doesn’t mean the factories are imaginary. It means the price of the paper can come away from them.

2. The productivity gains are real at the desk and hard to find in the economy

Each side in this argument has half of the evidence.

At the level of a single task the gains are large, and they have been replicated. In the best-known field study, call-centre agents using an AI assistant resolved 14 per cent more queries an hour, and new, less experienced agents 34 per cent more. Controlled experiments on professional writing found the work done about 40 per cent faster and to a higher standard. These are not vendor claims. They are randomised studies, published in the leading journals.

The same research also finds where the gains stop. In a 2025 trial, experienced open-source developers took 19 per cent longer with AI tools, and believed they had been faster. In a 2026 survey of about 6,000 executives in the US, the UK, Germany and Australia, 70 per cent of firms used AI but more than 80 per cent reported no effect on productivity or employment over three years. Danish administrative data on 25,000 workers found no measurable effect on earnings or hours.

The Productivity Institute’s report of 29 October, The AI Ambition Gap, sits across that divide, which is why it is quoted by both sides. It reports the task-level gains, the 14 per cent and the 40 per cent. It also reports that developers using AI write more than three times as much code but release only 30 per cent more software, that 89 per cent of organisations use AI somewhere, and that only about one in four has changed how work is actually done. Its summary is in its subheading: “Faster tasks do not automatically mean more productive businesses.”

In the economy as a whole, nothing has broken from trend yet. American labour productivity grew by about 2 per cent in 2025, and the third quarter of 2026 came in at 3.6 per cent annualised. That is better than the 2010s, but nobody at the Federal Reserve is yet willing to call it a new regime. British productivity is close to flat.

So the precise claim is this: AI raises the productivity of particular tasks, and has not yet shown up in the productivity of firms or of the economy. “AI boosts productivity” is too strong as a statement about the economy. “Studies show the opposite” is too strong as a statement about the research. Which way it goes is not something the machine decides.

3. How big is the investment? It depends what you measure

There are three figures here, and the argument keeps swapping one for another.

The first is the flow of investment. The Bureau of Economic Analysis has no line for AI. Its nearest category is information-processing equipment plus software, which counts every computer and every program. That was 32 per cent of American business investment in 2025 and about 36 per cent in the third quarter of 2026. That is where “a third of business investment” comes from, and on that definition it is about right. But the share was already 28.5 per cent in 2019, before the present boom. Counted narrowly, as the capital spending of the leading AI firms, a recent NBER paper puts AI at 8.5 per cent of all American private fixed investment in 2025, heading for about 15 per cent in 2026. That is very large for one technology in so short a time. It is not a third.

The second is the stock of capital, which is a different thing. American private non-residential fixed assets are worth about $36 trillion. A year’s spending on computers and software comes to about 4 per cent of that, and the AI share of the stock is smaller still, because chips wear out or go obsolete in a few years. Nobody claims that AI is a third of America’s capital, and nobody needs to.

The third is the share of growth, and here the numbers become dramatic. Jason Furman calculated that information-processing investment accounted for 92 per cent of American GDP growth in the first half of 2025; revised data roughly halve that. ING’s estimate for mid-2026, after taking out the imported chips and servers, is 36 per cent. Investment that is small next to the capital stock can still be most of the change in the economy, and that matters most for what happens next. If the spending stops, a third of American growth stops with it.

4. Profitable firms, borrowed money

It is true that the companies at the centre of the boom are profitable. Microsoft, Alphabet, Amazon and Meta are among the most profitable firms in history. But they are no longer paying for the build-out from profits alone. Goldman Sachs counts $296 billion of bonds issued by the hyperscalers so far this year, $67 billion of it since 29 October. The firms they are building for are a different matter: OpenAI’s own projections showed a loss of $14 billion for 2026, and CoreWeave, which rents AI chips to everyone else, now borrows at more than 16 per cent on its 2032 bonds. In the month to Thursday, the cost of insuring Nvidia’s debt against default rose by 76 per cent.

That is how a boom paid for by profits becomes one paid for by credit, which is the kind that ends badly. The Jefferies strategist Christopher Wood put it most simply: the arms race will end when “the credit card will be taken away.” The profitability of the leading firms is a reason why a fall would not wipe them out. It is not a reason why the prices cannot fall.

5. What moved the market was not a forecast about the machine

On Thursday the main reported causes were not new evidence about what AI can do. They were the cost of borrowing, and the results of Tuesday’s midterms. Data centres were on 34 local ballots this autumn, a third of them in Ohio. At least sixteen states had a candidate for governor calling for a moratorium. New York imposed the first statewide moratorium on hyperscale sites in July, and Texas has stopped new grid connections pending an audit. Voters facing higher electricity bills and water restrictions have started to decide where the machine gets built, and investors have noticed.

That is the point the Guide makes about every earlier wave of automation. In 1962 the same computers met three American unions and produced three outcomes. The Mine Workers and the Steelworkers were hollowed out. The Transport Workers and the Longshoremen traded working rules for job guarantees. The New York electricians of IBEW Local 3 struck and won a 25-hour week at the same pay. In January 1984 International Viewpoint reported a German Volkswagen engineer’s prediction of factories “empty of people”. That May, IG Metall struck for the 35-hour week. The robots arrived as forecast, and the week came down anyway, to 38.5 hours in 1985 and 35 by 1995.

So we can stop arguing about whether the machine works, because it does, in places, and we can stop arguing about whether there’s a bubble, because the credit markets will answer that whatever we say. What isn’t settled is who gets the hours that the tasks now save, and who pays when the borrowing is called in. Those questions don’t have forecasts. They have sides.

Sources

Markets, 5–6 November 2026. Index closes, stock moves and the Bank of England decision: Stan Choe, Associated Press, via NBC Bay Area, 5 November 2026. CoreWeave bond yield, CDS spreads, hyperscaler debt issuance (Goldman Sachs) and the Christopher Wood quotation: Deena Zaidi, “The AI ‘credit card’ is getting costlier as CoreWeave bond yield tops 16%”, CNBC, 6 November 2026.

Productivity. Brynjolfsson, Li and Raymond, “Generative AI at Work” (NBER w31161; Quarterly Journal of Economics, 2025). Noy and Zhang, Science, July 2023. METR, “Measuring the impact of early-2025 AI on experienced open-source developer productivity”, 10 July 2025. Yotzov, Barrero, Bloom et al., “Firm Data on AI” (NBER/SIEPR, February 2026). Humlum and Vestergaard, “Large Language Models, Small Labor Market Effects” (NBER w33777, 2025). The Productivity Institute (van Ark and Karountzos), The AI Ambition Gap, commissioned by Bytes Technology Group, 29 October 2026, and TPI news release of the same date. BLS, Productivity and Costs, Q3 2026 preliminary, 5 November 2026. St Louis Fed President Musalem, speech of 28 May 2026.

Investment. BEA NIPA via FRED: private nonresidential fixed investment (PNFI) and information-processing equipment plus software (A679RC); current-cost net stock of private nonresidential fixed assets (K1NTOTL1ES000). Shares computed from those series. Wachter and Wachter, “What Investment Data Implies about the AI Transition” (NBER w35290, revised October 2026). Jason Furman, as reported in Fortune, 7 October 2025. James Knightley, “How much is AI contributing to US economic growth?”, ING Think, 20 August 2026.

Data centres and the midterms. S&P Global Market Intelligence, “Data center backlash shapes midterm ballots, campaigns across US”, 1 October 2026; MultiState, 18 and 24 September 2026.

The archive. Red Mole Guide No. 16, The Fourth Forecast: Automation, Class and the Mode of Production in the Fourth International’s press, 1936–2026: Ed Beecher, International Socialist Review, 1962; Winfried Wolf, International Viewpoint, January 1984; Eurofound on the IG Metall working-time settlements.