The software-mageddon isn't about software

Posted on: 24 February 2026

Two trillion dollars in market capitalisation incinerated in under a month, and almost nobody is looking in the right direction. The sell-off that has gutted the software sector since late January 2026, turbo-charged on Monday by the convergence of a Citrini Research scenario report, Anthropic's Claude Code announcement and Nassim Taleb's warning about software bankruptcies, is not a market correction. It is the first visible signal of a far deeper repricing: that of the economic value of human cognitive intermediation.

The mechanism operating beneath the surface is simpler and more brutal than the financial narrative suggests. This is not about whether artificial intelligence will "replace" software, as though it were a product-versus-product contest. It is about the entire layer of professionals, consultants, developers and analysts who served as the bridge between a problem and its solution being compressed simultaneously, with equity markets beginning to price that compression with all the subtlety of a wrecking ball.

The facts, in their proper sequence. In late January, Anthropic's Cowork plugins triggered what Jefferies called the worst sentiment ever recorded in the software sector: "radioactive," according to Bloomberg Intelligence. The iShares Expanded Tech-Software ETF (IGV) erased every gain accumulated since ChatGPT's launch in November 2022. Palantir down 22 per cent year-to-date, Salesforce and ServiceNow off between 25 and 30 per cent, Microsoft down 18 per cent despite solid quarterly results. Thomson Reuters, whose business rests on legal databases built over decades, suffered its worst single-day plunge on record when the market grasped that an AI plugin could replicate in minutes what Westlaw sells for thousands of dollars an hour.

Then came Monday 23 February, the perfect storm. Citrini Research, a thematic equity firm virtually unknown until that moment, published a five-thousand-word Substack piece titled "The 2028 Global Intelligence Crisis": a fictional scenario set in June 2028 in which accelerated AI adoption triggers mass white-collar unemployment, consumption contraction and a 38 per cent collapse in the S&P 500 from its peak. The same day, Anthropic announced that Claude Code can modernise COBOL, the programming language underpinning much of the world's banking infrastructure and the core of IBM's mainframe business. IBM lost 13 per cent in a single session, its worst day since 2000. And as a bitter garnish, Taleb declared from a Universa Investments event in Miami that markets are underpricing structural risk and that bankruptcies in software are probable: "Someone will make a lot of money in AI. It's not guaranteed to be the companies that make up the AI trade today."

The most revealing aspect of Monday's rout was not the panic itself but its radius. It was not merely software stocks that cratered. DoorDash, American Express, Visa, Mastercard, Blackstone, Apollo, KKR: all down between 4 and 6 per cent. The contagion reached Indian IT services, with the Nifty IT Index falling for a fifth consecutive session. Private equity firms including Arcmont Asset Management and Hayfin Capital Management are reportedly hiring consultants to assess portfolio exposure to AI disruption risk, according to Bloomberg. When private equity starts checking for damage, the phase of academic speculation is over.

Citrini's report, for all its explicit disclaimers that it is a scenario rather than a prediction, identified a mechanism that deserves serious attention. It calls it the "human intelligence displacement spiral" and it works like this: AI capabilities improve, companies need fewer workers, white-collar lay-offs increase, displaced workers spend less, margin pressure pushes firms to invest more in AI, AI capabilities improve. A negative feedback loop with no natural brake. The sharpest section of the scenario concerns the software paradox itself: ServiceNow, whose entire business consists of workflow automation, is rendered obsolete by better workflow automation. The company that sold efficiency is devoured by greater efficiency. As Citrini writes: what were they supposed to do, sit still and die slower?

This is the most rationally elegant suicide in the history of capitalism. These companies spent a decade and hundreds of billions building the artificial intelligence that is now consuming them. They funded their own execution and sent the invoice to shareholders. The truly cruel part is that they cannot stop. If ServiceNow ceases investing in AI, a competitor will do so and kill it anyway; if it continues investing, it accelerates the technology that renders its own product obsolete. Every dollar spent on survival brings extinction closer. The hyperscalers, Microsoft, Meta and Amazon, have announced aggregate 2026 capital expenditure increases of roughly 60 per cent year-on-year, according to Morningstar. They are raising the stakes at a poker table where the possible winnings include their own irrelevance. The market, which until yesterday applauded every AI investment announcement, is now looking at those same figures and wondering whether they are suicide notes dressed up as strategic plans.

This paradox is the key to understanding why the sell-off runs deeper than it appears. We are not witnessing the replacement of one product by a superior product, as happened when the smartphone displaced the flip phone. We are witnessing the compression of the entire cognitive intermediation layer that justified premium pricing on qualified human labour. The McKinsey consultant who spent months synthesising data, the full-stack developer who built applications from scratch, the legal analyst who combed through case law, the data scientist who cleaned datasets: all operated in the same economic stratum, the bridge between a complex problem and a structured solution. That stratum is thinning.

The numbers confirm it from multiple directions. McKinsey has announced a 10 per cent reduction in non-client-facing headcount, roughly four thousand positions over 18 to 24 months, after tripling its workforce to 45,000 over the previous decade. As one analyst observed, the consulting pyramid, that model in which a vast base of junior analysts produced work sold at premium rates by partners, is crumbling from the bottom. AI does not assist the junior consultant; it replaces them in their primary function of synthesis. Vibe coding, the term coined by Andrej Karpathy in February 2025 and named Collins Dictionary's Word of the Year for 2025, has reached a market size of $4.7 billion. Ninety-two per cent of American developers use AI-assisted coding tools daily, according to Index.dev, and 41 per cent of global code is already AI-generated. Y Combinator reported that 25 per cent of its Winter 2025 batch had codebases that were 95 per cent AI-generated.

The full-stack developer, that figure who seemed immune because they "can do everything," now finds themselves in a peculiar position. They can do everything, but everything they can do is precisely what vibe coding is learning to replicate. Not tomorrow, not perfectly, not without holes: 45 per cent of AI-generated code fails security tests, according to CodeRabbit. Simon Willison, co-creator of Django, predicts a "Challenger-type disaster" caused by an AI-written component that nobody truly understood. But the direction is unambiguous and the velocity is accelerating. An MVP that in 2021 cost $50,000 and three months of work now gets built over a weekend on the back of an API subscription.

Goldman Sachs offered the most unsettling comparison. Its strategist Ben Snider likened the future of software to that of newspapers in the internet age, warning that the current sell-off may be "the end of the beginning, not the beginning of the end." Anyone who lived through the digital transition of publishing recognises the pattern: first denial ("the print newspaper will never die"), then cosmetic adaptation ("let's build a website too"), then the revenue collapse when the market realises that the value never resided in the medium but in the intermediation between information and reader. That intermediation layer was compressed until it vanished. The same mechanism, applied to every form of structured cognitive work, is what the market is beginning to price.

There is one aspect that almost nobody is discussing, most likely because it is the most uncomfortable. Every digital transformation guru, innovation consultant, leadership coach, LinkedIn thought leader and corporate trainer operates in precisely the same cognitive intermediation layer that AI is compressing. They sold the ability to translate complexity into actionable recommendations, but that translation is exactly what a language model does at marginal cost approaching zero. The irony could not cut deeper: the industry that spent twenty years selling "disruption" as a service is discovering what it means to be disrupted. And there is no pivot available when your product is the very cognitive capability the machine is replicating.

Anyone seeking certainty about how this ends will not find it, and anyone claiming to have it is selling something. Taleb is right when he says the profits will be enormous but not necessarily for the companies that currently dominate the trade: the history of every technological transition confirms it, from railways to the internet. The most dangerous error at this juncture is the one the market is committing in both directions simultaneously: repricing all software indiscriminately as though it were destined for extinction, while assuming that today's AI leaders will be tomorrow's winners. The truth is likely more complex and more selective. Some software companies with proprietary databases built over decades, Thomson Reuters and RELX among them, possess moats that take time to erode. Others, those that were essentially selling packaged cognitive labour as product, never had any.

What is certain is that the cost of producing a unit of structured cognitive work is plummeting, and this repricing will not confine itself to software. Every sector that relies on selling hours of qualified human thought, from strategic consulting to programming, from legal analysis to advanced accountancy, is about to undergo the same compression. Citrini's report may be a scenario rather than a forecast, but the mechanism it describes, that reflexive loop in which automation generates redundancies that generate reduced consumption that generates more automation, is as real as gravity. The question is not whether it will happen. The question is how quickly, and who, in the meantime, manages to grasp that value no longer resides in knowing how to bridge a problem to its solution, but in knowing which questions to ask a machine that already finds the solutions on its own.