If every layer of the AI industry is making money, why are serious investors betting that it will fail?
Fast Track
This article evaluates a recent podcast debate, "Is the AI Boom a Bubble?", in which one presenter argued the bull case and the other the bear. The bull's figures hold up remarkably well: demand is outrunning supply, memory chips are sold out until 2029 and the big cloud companies hold $1.8 trillion of signed orders. The bear's case is not really about today's numbers but about who is paying for them: much of the profit in the industry is funded by investors and lenders betting on a productivity boom that has not yet appeared. Both can be right. The likeliest form of failure is not a dramatic crash but a long, grinding fall in share prices, and the first cracks would show among the heavily indebted firms at the edge of the industry.
Introduction
In early November 2026, two presenters sat down to settle a question that hangs over every pension fund in the Western world. In a podcast episode titled "Is the AI Boom a Bubble?", Piers took the bull case and Stephen the bear. They recorded it on Friday 6 November, the day after one of the sharpest market sell-offs in years, and three days after a Democratic landslide in the American midterm elections had put data centres at the centre of politics.
This article is an evaluation of that debate. It sets out each side's argument fairly, checks the bull's figures against the latest company results, and asks the question that matters most to investors: if the AI trade does fail, what would failure actually look like? Readers new to the subject may find it helpful to read our earlier guide, "The AI Trade: Who Builds, Who Borrows, Who Benefits", first.
WHY IT MATTERS
The biggest companies in the stock market are the ones building AI, so anyone with a pension or an index fund has a stake in this argument. Some professional investors are now betting heavily on the boom ending badly; others are betting on it running for years. Understanding the strongest version of each case is the best protection against being swept along by either.
CONTENTS COVER THE FOLLOWING
1. The Debate and Its Moment
The timing gave the debate its edge. On Tuesday 3 November, American voters handed the Democrats a sweeping victory in the midterm elections, which some called a "blue tsunami". Markets initially took the news well. Then, on Thursday, came what Piers described as "a proper old-fashioned sell everything that's not nailed down kind of session": the Nasdaq was down by more than 5% at one point, and companies that build or supply data centres were hit hardest.
Piers read the sell-off as a reaction to politics rather than proof of a bubble. Data centres are unpopular with American voters, who see them as noisy buildings that push up electricity bills and may take their jobs. The Democrats, who take control of Congress in January, have made clear they intend to act. Their main proposal, the GRID Act, would require new data centres to supply their own power rather than draw on the public grid, a rule that would cover more than 99% of planned capacity. Twenty House Democrats have backed a bill for an outright moratorium on new data centres, although the president would almost certainly veto it, and states such as Texas are extending their own temporary bans. The market drew the obvious conclusion: regulation would slow the build-out and constrain the supply of new computing capacity, and the companies selling the chips and equipment for it would sell less. He also pointed out that semiconductor shares had already been falling for months, with some down by half: a slow deflation rather than a sudden burst.
2. The Bull Case: A Bubble Without the Signature
Piers built his case on a simple observation about history. Bubbles, he argued, have a signature. From the railway mania of the 1840s to American housing in 2008 and the shale oil boom, crashes were preceded by the same three things at once: falling prices, rising supply and idle capacity. Supply overtook demand, and the collapse followed. In AI, he said, we see the exact opposite.
Demand is accelerating, and much of it is contracted. Nvidia's revenue more than doubled in a year, to $96bn in a single quarter. The big cloud companies are sitting on signed, legally binding orders stretching years into the future. And adoption has barely begun: today there is far less than one AI agent per person, and Piers expects that to become several each.
Supply is sold out. SK Hynix, the world's biggest maker of AI memory, has sold its entire output until 2029. Copper and tin, both vital to data centres, are at or near record prices. Anthropic, one of the leading AI labs, has been rationing how much customers can use.
Every layer is making money. This was Piers's most striking claim. He walked up the industry layer by layer: Nvidia, the chipmaker, earns gross margins of 75%; data-centre landlords are raising rents; the cloud divisions of Amazon, Google and Microsoft run operating margins of 36% to 39%; and Anthropic's IPO filing shows an operating profit.
Real adoption is only starting. Piers gave an example from his own family. His brother runs an electronics manufacturer with about 300 staff; today 3 of them use AI tools seriously, and next year he expects around 30. Multiply that across the economy and the demand still to come is enormous. Beyond that lies an "agentic" phase, in which AI agents work on people's behalf and use far more computing power than individuals do, and a robotics phase, in which every robot needs AI to run.
This time, overbuilding is unlikely. Piers's most original point was that past bubbles burst because infrastructure was built far ahead of demand. This time, something outside the industry's control acts as a brake: governments, and above all the supply of electricity. Grid connections are scarce, voters are hostile to new data centres and the midterms have made both constraints tighter. That, he argued, makes the overbuild that ended previous booms much harder to achieve. Here the bull turns the midterms to his advantage. Regulation that constrains supply cuts both ways: it limits how much the builders and their suppliers can sell, which is why chipmakers fell on Thursday, but it also prevents the glut that ruins prices, protecting the profits of those whose capacity is already built. Scarcity, in other words, may be bad for growth but good for margins.
3. The Bear Case: A Bet on a Productivity Boom That Has Not Arrived
Stephen's case is that the whole boom is a bet on a productivity boom across the economy, and that this boom has not arrived. If it does not come, much of the investment will be wiped out. He cited a Morgan Stanley estimate of $570bn of AI-related borrowing, before counting the far larger sums of shareholders' money, and argued that the productivity gains are "nowhere to be seen" outside the largest technology companies.
Instead of money flowing out into the wider economy, he sees a circular washing machine: money spinning round a small circle of AI firms through equity stakes, loan guarantees and computing contracts. This is his reply to Piers's claim that every layer is profitable. The profits are real, he argued, but they are funded by investors and lenders betting on the future, not by customers. As the Financial Times columnist Katie Martin put it, in a crisis circular financing becomes a circular firing squad.
He also doubted that more AI means more value. A 2025 study by the research group METR found that experienced software developers believed AI made them about 20% faster, when in fact they took about 19% longer. App releases doubled in the first half of 2026, he said, while downloads rose only 2%.
Finally, if the boom stalls there is no safety net. Governments from Europe to the United States have little room to spend their way out of a downturn, with inflation keeping interest rates high and the American deficit at about $2 trillion. He ended with the investor Paul Kedrowski's warning that AI has reached its "Minsky moment".
IN PLAIN TERMS: WHAT IS A MINSKY MOMENT?
The American economist Hyman Minsky argued that long periods of stability breed their own instability. When times are good, lenders relax, borrowers take bigger risks, and eventually debts are taken on that can only be repaid if prices keep rising. A Minsky moment is the point at which that stops being true: lenders lose their nerve, borrowers are forced to sell, and prices fall faster than anyone expected.
The bear's strongest card, which he did not play. The AI build-out now runs on borrowed money: the big technology companies have issued well over $200bn of bonds this year, and Amazon has spent $42bn more cash than it generated over the past twelve months. That borrowing competes for the same pool of savings as the American government's deficit and the Covid-era debt wall, as thousands of companies refinance cheap loans at higher rates. The liquidity analyst Michael Howell argues that this pool is not growing but draining away, as maturing debt soaks it up (see our summary of his analysis).
So are long-term interest rates high because of inflation, as the Federal Reserve says, or because three enormous borrowers are competing for a shrinking supply of money? The two overlap, since the AI boom itself pushes up the price of electricity, copper and memory chips. But if inflation is the cause, rates should fall as it cools; if competition for capital is the cause, they may stay high regardless, and so will the cost of financing the build-out. That is an interpretation rather than an established fact, but it is the strongest form of the bear case, and it does not require AI to fail as a technology.
The question is not whether AI works, but whether there is enough money in the world to pay for it on the current timetable.
4. Checking the Bull's Numbers
Debates are easy to win with numbers nobody checks. So we checked Piers's against the latest company results. They hold up remarkably well.
| Claim | What the latest results show | Verdict |
|---|---|---|
| Nvidia: 75% gross margin, $96bn a quarter | $96.2bn of revenue and a 75% gross margin in the quarter to July 2026 | Confirmed |
| Memory sold out until 2029 | SK Hynix said on 28 October that its DRAM and HBM output is fully sold out through 2029 | Confirmed |
| Copper and tin near records | Copper set repeated records in 2026, above $14,700 a tonne; tin also broke its record | Confirmed |
| Data-centre rents rising | US vacancy fell to a record low of 1.4% in the first half of 2026, and rents are rising | Confirmed (no figure published) |
| Cloud margins of 36% to 39% | Amazon Web Services 38.7% and Google Cloud 36.2% in the third quarter; Microsoft does not report a comparable figure | Confirmed |
| Over $1 trillion of signed orders | About $1.8 trillion: Microsoft $677bn, Amazon $578bn, Google Cloud $561bn | Understated |
| Anthropic makes an operating profit | Adjusted operating profit of $1.62bn on $16.3bn of revenue in the first half of 2026 | True, with an asterisk |
| Anthropic is rationing usage | Several reports of tighter usage limits in 2026 because of compute shortages | Confirmed |
On demand and supply, then, the bull is right. But each of his headline claims carries an asterisk, and the asterisks are where the bear's case lives.
Anthropic's profit is an adjusted figure. Under standard accounting it reported a net loss of $80.7bn, mostly from non-cash revaluations of its convertible debt. More important is what sits behind it: $498bn of computing commitments it cannot cancel, which must be paid whether or not the capacity is used. That is roughly $70bn a year, against $16bn of revenue in the first half. In its favour, the revenue Anthropic earns for each dollar it spends on computing tripled between late 2025 and mid-2026.
Profitable is not the same as cash-generating. Amazon Web Services earns a margin of nearly 39%, yet Amazon's free cash flow over the past twelve months was minus $42bn, because it spent $211bn on building.
Signed orders are promises, not cash. A large share of the $1.8 trillion is owed by a handful of AI labs, which are themselves funded by investors. The orders are only as strong as the labs' ability to keep growing.
"Every layer is profitable" skips the weakest one. Piers folded the specialist AI cloud firms, the so-called neoclouds, into the same layer as Amazon, Google and Microsoft. Shown separately, they tell a different story: CoreWeave, the largest, lost $626m in the second quarter of 2026 and paid $640m in interest.
On today's numbers the bull wins. The bear's case is about who is paying for those numbers, and for how long.
5. What Failure Would Look Like
Some professional investors are now betting on the AI trade failing. It is worth being precise about what that means, because the trade can fail without the technology failing. The internet transformed the world after 2000, yet most investors in the fibre-optic networks that carried it lost their money. Failure could take three broad shapes, set out here from most to least likely on current evidence.
A long, grinding fall in share prices. Revenues keep rising, but investors stop paying for the future and demand cash today. Share prices drift down 20% to 40% over many months. The warning sign is good results greeted by falling prices, and it has already begun: in late October Alphabet's shares fell 13% and Meta's 7% after results that were, by most measures, strong.
A credit accident at the edge. One of the indebted neoclouds fails to refinance its debts. Lenders seize the chips pledged as collateral and sell them, rental prices for AI computing collapse, and the losses spread to lenders and suppliers. The warning sign would be neocloud bond yields, already above 15% at CoreWeave, climbing towards 20% or 25%, or a bond sale that fails.
A spending cliff. The giant cloud companies cut their building plans, chip orders collapse, and the chipmakers, who were paid first, are hit last but hardest. This is what happened to Cisco, the star of the dot-com era: in 2001 its telecom customers stopped buying, it wrote off about $2.25bn of unsold stock, and its shares fell roughly 86% from their peak. The warning sign would be the first big cloud company guiding its spending lower, or memory makers cutting prices instead of reporting sell-outs.
What could set these off? Four triggers stand out. AI lab revenue could stall while their computing commitments stay fixed. A price war, led by Chinese models such as DeepSeek that cost a fraction of American ones, could cut what customers will pay. A financing shock, from higher long-term interest rates or a bond market tired of technology debt, could choke off the borrowing. Or politics could intervene, although here Piers has a point: limits on power and planning may cap the overbuild as much as they threaten the trade.
The hardest part of betting against the boom is timing. Today, almost every measure of demand is running ahead of supply. A bear may be right about where the story ends and still lose money waiting for it to get there.
6. The Verdict: Two Kinds of Bubble
The debate ended on common ground, and it is the most useful idea in it. Stephen drew on the book Boom by Byrne Hobart and Tobias Huber, which separates two kinds of bubble.
A mean-reversion bubble is destructive: a financial miracle that promises returns it cannot deliver, like the South Sea Bubble of 1720 or the American subprime mortgage boom. When it bursts, the value simply disappears. An inflection bubble is productive: a genuine breakthrough, like railways, electricity or the internet, around which investors get carried away. It can still bring years of pain, but it leaves behind infrastructure that changes the world, and a new generation of companies that learn to use it.
Both presenters agreed that AI belongs to the second kind. Piers summed it up as "bad in the short, good in the long", while arguing that November 2026 does not look like the eve of a dot-com crash.
Our evaluation is that both are partly right, about different things. The bull wins on the evidence available today. Every figure he gave checks out, and several understate his case. The bear wins on the question that matters next. Too much of the industry's profit rests on money supplied by investors and lenders rather than by customers, and that money will keep flowing only as long as faith in a future productivity boom holds. It must also be borrowed in a crowded market, competing with government deficits and the Covid-era debt wall for a pool of liquidity that is no longer growing. Piers's point about power is the most original in the debate, and it cuts both ways: scarce electricity may prevent a classic overbuild, but it also means the industry's growth is now partly in the hands of voters and governments.
If the boom does deflate, history suggests it will start at the edge, among the indebted neoclouds, and work inwards. That is the place to watch.
Bottom Line
Is the AI boom a bubble? On the evidence of this debate, it is the productive kind, an inflection bubble rather than a swindle, and the industry's current numbers are stronger than many bears admit. Demand is outrunning supply, signed orders run to $1.8 trillion and almost every layer of the industry is profitable on paper.
What the debate reveals is that the real question is not whether AI works, but who is paying for it while we wait to find out. For now, much of the answer is investors and lenders, and their patience rests on a productivity boom that has not yet shown up in the wider economy. If that boom arrives in time, the bulls will have been right all along. If it is late, the money will run out before the demand catches up, and the losses will start at the edge of the industry and work inwards.
The railways and the internet both changed the world while ruining many who financed them. The question worth asking of AI is not whether it will change the world, but whether it will do so on a timetable its lenders can wait for.
DEEP DIVE: A DASHBOARD FOR BULLS AND BEARS
Seven signals will show which way the argument is going. Each is listed with its reading in late October and early November 2026, and what failure would look like.
| Signal | Reading now | What failure would look like |
|---|---|---|
| Rent for AI chips | Cheapest reliable H100 rate about $2 an hour, flat in October after summer cuts | A sustained fall below $1.50 an hour, with newer chips following |
| Data-centre vacancy | 1.4%, a record low | Rising vacancy, falling rents, cancelled leases |
| Signed cloud orders | About $1.8 trillion and growing | Growth stalling, or contracts renegotiated |
| Cloud company spending | About $730bn planned for 2026, unchanged after October's results | The first company to guide 2027 spending lower |
| Neocloud borrowing costs | Above 15% on some CoreWeave bonds | Above 20%, or a failed refinancing |
| Memory chips | Sold out until 2029 | Price cuts or delayed orders |
| AI lab revenue | Fast growth, with some reports of a slowdown | Flat quarters against fixed computing commitments |
Related Articles
- The AI Trade: Who Builds, Who Borrows, Who Benefits: the guide to the six layers of the AI industry, who takes the profits, who takes the risks and who takes the benefits.
- Summary of Michael Howell's Liquidity Analysis, November 2026: why liquidity is past its peak, and how AI borrowing and the debt wall are draining it.
- The AI Boom Just Stopped Paying for Itself
- Is the AI Boom a Bubble?
Glossary
Adjusted operating profit - A company's profit from its business after removing items it considers one-off or non-cash; not the same as profit under standard accounting.
Agentic AI - AI that carries out tasks on a person's behalf, such as booking, buying or comparing, rather than simply answering questions.
Backlog - Revenue a company has been promised under signed contracts but has not yet earned.
Collateral - An asset pledged to a lender, which the lender can seize if the loan is not repaid.
Debt wall - A cluster of loans and bonds falling due at around the same time, which must be repaid or refinanced, often at higher rates.
Free cash flow - The cash a company has left after paying its running costs and its investment spending.
Gross margin - The share of revenue left after the direct cost of making the product.
HBM (high-bandwidth memory) - The very fast memory chips that sit alongside AI processors.
Inflection bubble - A speculative boom around a genuine technological breakthrough, which leaves useful infrastructure behind when it ends.
Liquidity - The total pool of money and credit available to finance spending and investment across the financial system.
Mean-reversion bubble - A speculative boom built on returns that cannot be delivered, which destroys value when it ends.
Minsky moment - The point at which a long period of easy borrowing turns into forced selling and falling prices.
Neocloud - A specialist company that rents out AI chips, usually financed heavily with debt.
Productivity - How much output an economy or a worker produces for each hour of work.
Yield - The annual return a bond pays at its current price; a rising yield means lenders see more risk.
References
- Piers and Stephen, "Is the AI Boom a Bubble?", podcast episode, recorded 6 November 2026.
- Monetary Matters, Michael Howell on the Liquidity Cycle, interview, November 2026.
- Nvidia, results for the second quarter of fiscal 2027, 26 August 2026.
- Newsquawk, SK Hynix says DRAM and HBM production sold out through 2029, 28 October 2026.
- SMM, LME copper hits fresh record above $14,700 a tonne, 2026; Mining.com, tin price breaks record, 2026.
- CIO Dive, data-centre vacancy hits record low, 3 September 2026, citing CBRE.
- SEC-API, Amazon third-quarter 2026 filing analysis, 29 October 2026.
- Long Yield, the state of the AI build-out, October 2026 (Google Cloud margin, cloud backlogs).
- Fortune, Anthropic's S-1 financials, 27 October 2026.
- PYMNTS, AI usage limits become the new reality, 2026.
- CoreWeave, second-quarter 2026 results, 11 August 2026; Fortune, CoreWeave and neocloud debt, 16 October 2026.
- Dreaming Press, what it costs to rent an H100, H200 or B200 in October 2026.
- METR, the impact of early-2025 AI on experienced open-source developers, July 2025.
- Yahoo News, Democrats eye tougher data-centre regulations as they regain control of Congress, 4 November 2026.
Further Reading
- Byrne Hobart and Tobias Huber, Boom: Bubbles and the End of Stagnation (2024): the case that some bubbles are productive, and how to tell them apart.
- Hyman Minsky, Stabilizing an Unstable Economy (1986): the source of the "Minsky moment" and its theory of how stability breeds instability.
- Carlota Perez, Technological Revolutions and Financial Capital (2002): why great technologies arrive with financial bubbles, and what follows.
