The $470 billion question nobody wants to answer straight
The hyperscalers will spend north of $470 billion on AI infrastructure this year. Meta alone has guided for $115 to $135 billion in capital expenditure — the company just committed $21 billion to CoreWeave on top of a prior $14.2 billion arrangement, locking in compute through 2032, with early deployments on Nvidia's Vera Rubin platform. Microsoft is at $75 billion annually. The numbers are so large they have become aesthetically meaningless, like trying to visualize a light-year. But they are not meaningless in practice. Every dollar in those figures is a dollar that used to live somewhere else.
A significant portion of it lived in software budgets.
This is the uncomfortable arithmetic that has taken a chainsaw to the IGV ETF — down more than 27% year-to-date — and that no amount of bullish sell-side notes has been able to stop. ServiceNow is off 28% for the year. Salesforce down 26%. Intuit has lost more than a third of its market cap. The Stoxx Europe Software and Computer Services index shed over 5% in a single February session. RELX, the British data analytics giant, dropped more than 14% in one day. And these moves have been persistent, not a one-session panic — they have repriced, re-repriced, and repriced again across three months.
The bear case, stated bluntly, is this: if one AI agent can do the work of five human employees, you don't need five Salesforce seats. You need one. The seat-based subscription model, the elegant recurring revenue machine that made SaaS stocks the darlings of every fund manager from 2015 to 2022, is structurally threatened not because the software stopped working but because the unit of value it was attached to — the human worker — is being displaced faster than the industry modeled for. CIOs are consolidating vendors. Budget is rotating to API credits and model access. The "best of breed" era, which spawned Workday and ServiceNow and Snowflake and the rest, may be ending as enterprises decide they want fewer platforms and more agents.
Whether this is rational or hysterical depends, to an uncomfortable degree, on how quickly agentic AI actually rolls out at enterprise scale. The bulls — Dan Ives at Wedbush calling the selloff a "generational buy," JPMorgan arguing the disruption fears are "overblown," every incumbent CEO dutifully telling earnings calls that AI "depends on" enterprise software — all of them are essentially betting on friction. Switching costs. Data lock-in. The sheer organizational inertia of a Fortune 500 company that has spent fifteen years embedding Salesforce into its sales operation. They may be right. They usually are, until they're not.
But here is the thing that the disruption narrative conveniently skips past: the money going to Nvidia and CoreWeave and the hyperscalers is not appearing from nowhere. It was allocated. The capital expenditures that Meta is throwing at AI infrastructure — the company spent $72 billion on capex in 2025 alone — are coming from the same pools that once funded everything else: software licenses, headcount, vendor relationships. And those dollars are building something, even if nobody has yet demonstrated what that something's ROI looks like at scale.
Nvidia is the only institution that cannot lose this trade. Vera Rubin trains mixture-of-experts models with four times fewer GPUs than Blackwell, per Nvidia's own specifications, while delivering up to ten times lower cost per token. That is the kind of generational efficiency gain that makes Jensen Huang sound less like a CEO and more like a geologist who found a new mountain. AWS, Google Cloud, Microsoft Azure, CoreWeave, Lambda — all of them are scrambling to deploy Rubin clusters. The demand is so aggressive that CoreWeave, which had 62% of its 2024 revenue concentrated in Microsoft alone — a dependency that made institutional investors uncomfortable enough to be a drag on its IPO — now says no single customer will exceed 30% of revenue after the Meta expansion. That is what genuine diversification via a commodity war looks like.
Nvidia is the only institution that cannot lose this trade. Everyone else is running a very expensive bet.
The software incumbents are running a different and less comfortable race. They are betting they can bolt enough "agentic AI" features onto their existing platforms to justify continued seat counts — or better, to shift the pricing model entirely toward outcomes and tasks rather than users. Salesforce is building AI agents. ServiceNow CEO Bill McDermott keeps insisting that AI doesn't replace enterprise software but "depends on it." Maybe. ServiceNow's stock is still down 28% on the year.
What has happened in 2026 is not a sector correction in the conventional sense. It is an epistemological crisis. The entire analytical framework that made software stocks investable — net revenue retention, dollar-based expansion, the compounding logic of seat growth inside sticky enterprise accounts — has had its premises called into question simultaneously. Not refuted. Called into question. That is actually worse for price discovery than a clean refutation. Markets can price a known bad outcome. They cannot easily price a fundamental model uncertainty, and the result is exactly what we have seen: indiscriminate selling followed by fitful relief rallies followed by more selling, with every new AI model release functioning as a fresh catalyst for another leg down.
The historical parallel that nobody seems to be reaching for is the newspaper industry circa 2006. Revenues were still growing in nominal terms. Advertising was still flowing. The classified section still existed. The executives were making sensible arguments about brand equity and trusted relationships and the irreplaceable value of professional editorial judgment. They were not wrong about any of those things. They were wrong about the speed of the structural shift, and they were wrong about where the incremental dollar of advertiser attention would go when given an alternative. By the time the model broke visibly, it was already too late to adapt.
Whether enterprise software is the newspaper business of the 2020s is a genuine open question. The switching costs are real. The data moats are real. The incumbents are not passive. But the capex flowing into AI infrastructure is also real, and $470 billion a year tends to produce things.
Today JPMorgan reports. Citigroup reports. Dimon will say something careful and ominous about the macro environment — oil at $103 after a Hormuz blockade that reignited yesterday when ceasefire talks in Islamabad broke down, stagflation risk rising, the Fed paralyzed, Kevin Warsh not yet confirmed as FOMC chair. The banks will likely beat. They are living organisms optimized for disorder.
The software story will not be resolved by one earnings cycle. It will take several quarters of actual revenue data — not sentiment, not analogies, actual enterprise spending decisions — to know which side of this argument is correct.
In the meantime, the capital is flowing. Nvidia is building the next mountain. Meta is claiming the first acre. And ServiceNow is down 28% on the year while posting its ninth consecutive earnings beat.
Hold that last sentence in your head for a moment. A company that beats estimates nine quarters in a row loses more than a quarter of its value. That is not a market efficiently processing known information. That is a market pricing a future it cannot yet see, and doing it loudly, and doing it fast.
It might be right.