ComputerWeekly

The big cyber short, or dot AI bubble?


The AI build-out will not cause a repeat of the 2008 financial crisis. It is far more likely to cause a repeat of 2000, and that is not the reassurance it sounds like. A dotcom-style bust would still wipe trillions from markets, take a decade or more to recover, and arrive at a moment when the world has less slack to absorb it than it had in either 2000 or 2008.

I watched The Big Short again this week. What struck me was not the collapse. It was the 18 months before it, when a handful of investors read the loan tapes, saw that the collateral could not support the paper, and bought insurance against it while the sellers kept selling.

Boards ask me a simple question. Could the AI bubble produce a crisis on the scale of the subprime collapse, one that freezes credit, takes down counterparties and reaches every balance sheet on the planet? My answer has three parts. A sharp correction is now more likely than not within two to three years. The shape of that correction will look more like the dot-com bust than Lehman. And a full 2008-scale systemic event is a minority outcome, perhaps one chance in five. One in five is not a tail risk. It is a planning assumption.

The reasons the systemic probability is lower than 2008 are structural. So are the reasons the correction probability is high. Both deserve a plain explanation, because the damage will not land where most executives think it will.

2008 vs. 2000

Subprime was not a housing problem. It was a leverage and disclosure problem that used housing as raw material. Mortgages became bonds, bonds became collateralised debt obligations (CDOs), which became synthetic CDOs referencing other CDOs. The paper outstanding was a multiple of the loans beneath it, and the demand for it was partly manufactured. Goldman Sachs sold more than $40bn of mortgage securities in 2006 and 2007 while its own desk, on the chief financial officer’s instruction, shorted the market. It did not tell the buyers – pension funds, insurers, unions and foreign banks.

Three features made it systemic. The banks that ran the payment system were the holders, levered 30 to one. The instruments were opaque, so nobody could price a counterparty. And the collateral was a single asset class that fell everywhere at once. Remove any one of those and 2008 becomes a bad year rather than a near-death experience for the global financial system.

However, the dotcom bust was a valuation collapse, not a credit collapse, and that distinction is the whole argument. The Nasdaq peaked at 5,048 on 10 March 2000 and fell 78 per cent to 1,114 by October 2002, erasing roughly five trillion dollars of market value. It did not regain its high until April 2015.

The companies that died deserved to. Only about 14% of Nasdaq-100 IPOs were profitable at the top. Pets.com booked $5.8m of revenue against a $61.8m loss and was delisted within a year. Cisco, the picks-and-shovels supplier to the internet, traded at 201 times earnings and 20 times sales.

Yet the banks held. The payment system never froze. The US economy had a short, shallow recession, and the technology turned out to be everything its promoters claimed. The losses fell on equity holders, on employees paid in options, and on the pension funds that had drifted into growth stocks. That is a very different casualty list from 2008, and it is why most investors today reach for 2000 as the comforting precedent.

The AI build-out by the numbers

AI stopped being an equity story in 2025 and became a debt story. American AI companies raised $217bn of debt last year, had passed $445bn by mid-August 2026, and are expected by Morgan Stanley to reach close to $600bn for the year. Hyperscaler leverage doubled from 0.9 to 1.8 times in roughly six months.

The cash that used to pay for this has turned. Google’s free cash flow went negative in the second quarter of 2026 for the first time since its listing. The paper that does not appear on any balance sheet has grown: Goldman counts $1.5tn of lease commitments, a trillion of which have not started, and another $1.5tn of purchase commitments for chips and power.

The second tier is where 2000 and 2008 meet. Neocloud providers such as CoreWeave carry debt approaching $50bn, mostly floating-rate private credit from Blackstone, Apollo, BlackRock and Magnetar. The collateral is GPUs with a three-to-five-year life. CoreWeave’s five-year credit default swap touched 1,195 basis points on 15 October, pricing a default probability above sixty per cent.

And the revenue to carry all of it is not there yet. Bain estimates the industry needs $2tn of annual revenue by 2030 to pay for its compute and will fall about $800bn short. MIT research found 95% of enterprises getting no measurable return on generative AI. OpenAI, valued at $852bn, does not expect a profit before 2030. This is the Pets.com problem at hyperscale: the spending is real, the demand is assumed.

Which crash is the better model

On the evidence, 2000 fits better than 2008, and the fit is uncomfortably close. Valuation is back at the 1999 level: the Shiller CAPE stands at 41 against a dotcom peak of 44 and a long-run median of 17. Concentration is worse than 1999: the top five companies are 30% of the S&P 500 against 17% at the dotcom peak, and the Magnificent Seven are a third of the index. Investor sentiment has already turned: 54% of fund managers surveyed by Bank of America call AI stocks a bubble and name AI the top tail risk. The single-day lesson has been given once, when DeepSeek’s cheap model wiped $589bn dollars from Nvidia on 27 January 2025. And the revenue gap, Bain’s $800bn, is the same gap that killed the dotcoms: infrastructure built for demand that had not arrived.

The differences from 2000 cut both ways. Nvidia trades at roughly 30 times earnings with a 53% net margin, not Cisco’s 201 times. The hyperscalers are profitable incumbents, not two-year-old IPOs. That is the case for a softer landing. But in 2000 the infrastructure was financed by equity that could simply evaporate. This time it is financed by half a trillion dollars of debt.

That is where 2008 re-enters. The analogy with subprime holds on opacity, manufactured demand and who ends up holding the paper. Hyperscalers guarantee the leases of smaller partners, creating contingent liabilities that surface only when called. The ecosystem is, in Robin Wigglesworth’s words, incestuous: the chip vendor invests in the model company, which signs a compute contract with the neocloud, which borrows against that contract to buy the vendor’s chips. It is the 2006 CDO daisy chain with better branding. And the holders are familiar: pension funds and life insurers anchor the bonds.

The analogy with 2008 breaks on the two things that made it systemic. The originators have cash: Microsoft, Alphabet and Amazon can write down a trillion dollars of datacentres and still make payroll. And the regulated banks are mostly originating to distribute, not holding.

So the honest model is a hybrid. The hyperscalers get the dotcom outcome: a long, grinding equity drawdown that the companies survive and their shareholders do not enjoy. The neoclouds and their lenders get the Hynix outcome: debt converted to equity, shareholders wiped, assets sold at a fraction of cost. The private credit funds, insurers and retail vehicles in between get the 2008 outcome in miniature, without a taxpayer behind them. A decade-long hangover is the base case.

The cyber short

When the correction arrives, the thing that gets shorted is not a stock. It is the operational resilience of every business that has quietly re-platformed onto these providers.

Your security tooling, your identity provider, your fraud models and your customer-facing assistants increasingly run on inference bought from a neocloud, which bought the GPUs with floating-rate private credit, secured on a contract with a model company that is itself years from profit. In 2008 the question was whether your bank would open on Monday. In this cycle the question is whether the provider holding your data and running your controls is in receivership by Friday, and what the administrator does with the hardware.

Three consequences follow. Distressed providers cut security first, because security is a cost centre and the lender wants cash. Distressed providers get acquired fast, and hostile actors treat a rushed integration as an open door. And distressed providers sell compute cheap. When the 1873 railway bust ended, the tracks were still there and freight became cheap for everyone. When the datacentres are repossessed, inference becomes cheap for everyone, including the people running credential stuffing, deepfake fraud and automated vulnerability discovery against your estate. The attacker’s cost curve falls at exactly the moment your budget does.

The verdict

A 2008-scale crisis needs a single collapsing asset class, leverage inside the institutions that run the plumbing, and paper nobody can price. The AI build-out has the first and the third. It has only partly got the second, because the banks have pushed the risk to private credit, insurers and pensions. A 2000-scale crisis needs stretched valuations, extreme concentration, infrastructure built ahead of revenue and a central bank tightening into it. The AI build-out has all four. My estimate: a correction that impairs a trillion dollars or more of infrastructure and takes the leading indices down by a third to a half is more likely than not by 2029. A Lehman-scale systemic event is roughly one chance in five. Neither number is a reason to relax. The dotcom recovery took 15 years in a far more forgiving world.

The investors who profited in 2008 did one thing differently. They read the collateral before the market did. For a board, the collateral is your supplier chain. The one thing that can be done right now is to commission a financed-supplier concentration review of every AI and cloud provider in your critical path before the next budget cycle, mapped to ISO 27001 supplier and continuity controls, with a costed answer to two questions: what happens if the provider fails, and what happens if the price triples.

Ian Thornton-Trump is fractional CISO at Inversion6.



Source link