The market is simultaneously pricing AI capex failure and AI destroying all software. Both cannot be true.

JP Morgan research note on the February 2026 software sell-off by Mark R Murphy titled Software Collapse Broadens with Nowhere to Hide, questioning the leap from Claude Cowork Plugins to full enterprise software disruption

Anthropic released 11 open-source plugins for Claude Cowork on January 30. The plugins are Apache-2.0 licensed and file-based, and they run in a macOS-only research preview. Within a week, the IGV software ETF had fallen 32% from its September peak to a 52-week low of $79.65, market cap had dropped by about $2 trillion, and hedge funds had made $24 billion shorting the sector. The RSI hit 18, the most oversold reading since 1990. JP Morgan titled their note “Software Collapse Broadens with Nowhere to Hide.” Jefferies coined the term SaaSpocalypse. It was the worst software stock crash since the dot-com bust.

Bank of America’s Vivek Arya identified the paradox: investors are punishing hyperscaler stocks because AI capex might generate weak returns, and at the same time destroying software stocks because AI adoption will be so pervasive it renders all existing software obsolete. If AI tools aren’t generating meaningful ROI, they’re not replacing enterprise software at scale. If they are replacing enterprise software at scale, the hyperscalers are earning extraordinary returns on their infrastructure investment.

The BofA AI paradox in the 2026 SaaSpocalypse showing two mutually exclusive narratives: AI capex generating weak returns with $670B spend and 4 percent coverage ratio, versus AI destroying all software with 32 percent IGV drawdown and $2 trillion lost despite 17 percent sector earnings growth

This paradox can only resolve in one of three ways: AI adoption is real and hyperscaler capex is justified, AI adoption stalls and software incumbents are fine, or the truth is somewhere in between and the market has mispriced both sides. The first two are internally consistent, and the market is pricing neither.

The bear case for enterprise software

Enterprise software monetizes through per-seat licensing. The SaaS business model depends on a stable correlation between headcount and license count. AI agents break that correlation. If 10 agents do the work of 100 people, a company can keep its software and still cut the headcount that justifies the seats, and CRM seat revenue drops with it. AlixPartners estimates up to $500 billion in enterprise software revenue could be at risk over time. IDC predicts pure seat-based pricing will be obsolete by 2028.

Enterprise software’s traditional defense was the trained-user-interface moat: the years of institutional muscle memory that make switching costs prohibitive. Databricks CEO Ali Ghodsi told TechCrunch that this moat collapses when the interface becomes natural language. If the value of Salesforce or ServiceNow lived in their UI rather than their data, and the UI can now be replicated by a general-purpose model, then the moat was shallower than anyone thought. VC has fled traditional SaaS entirely; as one investor noted, “an entrepreneur approaching a VC fund today with a SaaS startup won’t even reach the pitch stage.”

The build-versus-buy equation is tipping toward build. Klarna ditched Salesforce and Workday, consolidated onto its own AI-augmented stack, and used an OpenAI-powered bot to handle work that previously required 700 employees. SaaStr’s analysis of Gartner’s $1.43 trillion 2026 software spending forecast finds that roughly 9 percentage points of the 14.7% headline growth are price increases on existing software, so most of that growth is not net new demand. AI is eating SaaS budgets. It pulls IT spend toward infrastructure and reduces the headcount that generates software seats. The IGV at $80 prices in this bear case.

The bull case for software stocks

The bear case assumes a shrinking total addressable market (TAM). Goldman Sachs Research argues the opposite: the application software market grows to $780 billion by 2030 at a 13% CAGR, with agents accounting for over 60% of the total. The profit pool shifts from SaaS seats to agentic workloads, but the overall market gets larger. a16z’s Alex Rampell takes it further: if AI enables software to complete work, the addressable market expands from roughly $350 billion in enterprise software spend (about 1% of GDP) to the ~$6 trillion white-collar services market (~20% of GDP), a 20x expansion into work that was never software-addressable before.

David Friedberg made the sharpest version of this argument on the All-In Podcast: software transitions from helping people do work, to completing work, to doing work humans cannot do. At that point, the SaaS pricing model transitions from per-seat to value-based, and “SaaS basically takes over the services economy.” He estimates that the combined market cap of software companies could be 4x to 10x higher in five years, but “not evenly distributed.”

TAM expansion analysis from $350B enterprise software at 1 percent of GDP to Goldman Sachs $780B projection by 2030 with over 60 percent AI agent share, to the a16z thesis of $6 trillion in white-collar services at 20 percent of GDP, a 20x expansion

The software vs semiconductor valuation gap supports this case. The software sector is delivering 17% aggregate earnings growth in 2026 while trading at November 2022 EV/Sales multiples, back when the Fed was aggressively hiking into recession fears. The Russell 1000 Software subsector now trades at 32.4x forward earnings versus 43.6x for semiconductors. Recurring-revenue businesses with 90%+ gross margins and 95%+ renewal rates trade at a lower multiple than cyclical chipmakers with 40-60% margins and concentrated customer bases. Historically that’s an inversion that has not persisted.

Q4 2025 earnings vs stock performance disconnect in the 2026 software sell-off: Palantir plus 70.5 percent revenue growth but minus 11.6 percent stock, ServiceNow plus 21 percent but minus 28 percent, Oracle plus 10 percent but minus 53 percent from peak, sector aggregate plus 17 percent earnings growth versus minus 32 percent IGV drawdown
Software vs semiconductor valuation inversion in 2026: Russell 1000 Software at 32.4x forward PE trades below Russell 1000 Semiconductors at 43.6x, an 11.2x multiple gap, with IGV at $79.65 and S&P 500 software weight compressed from 12 percent to 8.4 percent

The hyperscaler AI capex question that connects both sides

Combined 2026 capex guidance from Microsoft, Alphabet, Amazon, Meta, and Oracle now approaches $700 billion, more than doubling from $256 billion in 2024. Bank of America calculates this consumes 94% of operating cash flows after capital returns. The Big Five raised $108 billion in bonds in 2025. AI-related services generate roughly $25 billion in direct revenue against $400+ billion in annual infrastructure spending, a coverage ratio of about 4%.

FY2026 hyperscaler AI capex vs cash flow: MSFT META GOOGL AMZN and ORCL estimated cash from operations less dividends and buybacks versus guided capital expenditure, with only Microsoft generating a $5B surplus while Meta shows minus $23B, Google minus $20B, Amazon minus $18B, and Oracle minus $30B

If the bear case is right and AI agents are replacing enterprise software at scale, this capex should already be generating enormous returns, and it isn’t. If the bull case is right and AI is expanding the TAM into the services economy, this capex is early-stage infrastructure investment that will compound over a decade. In that reading, $700 billion in annual spend is the foundation of a $6 trillion market. The two readings give the same capex figure opposite meanings, and the market hasn’t decided which one to believe.

Microsoft shows this tension most clearly. Its quarterly capex went from $1 billion in early 2015 to a record $37.5 billion in Q2 FY2026, with roughly two-thirds going to short-lived GPU/CPU assets. Yet Microsoft is the only hyperscaler that can fund this buildout from operating cash flow. Azure grew 39% in Q2 FY2026, crossing $50 billion in quarterly cloud revenue for the first time. The company is the biggest AI capex spender, the one best positioned to generate returns on that spend, and the company whose products (365, Dynamics, Azure) are supposedly being disrupted by Claude plugins. The market is punishing it for all three at once.

Microsoft quarterly AI capex from FY2015 to FY2026 showing growth from $1 billion to $37.5 billion per quarter, a 2048 percent increase, with recent quarters showing AI infrastructure acceleration

Bifurcation, not extinction: the SaaSpocalypse resolved

A 60% recession probability, a partial government shutdown, elevated tariffs, and a structural pricing transition are being sold as a single story, but only the pricing transition is structural. To separate it from the macro, ask which software categories are genuinely at risk and which are being sold by association.

Janus Henderson makes a useful distinction between “systems of record” and “systems of engagement.” Systems of record are deeply embedded in business processes, require regulatory compliance, and carry enormous switching costs: ERP, core finance, cybersecurity, observability. PitchBook described replacing one as “effectively open-heart surgery for an enterprise.” Systems of engagement are user-facing workflow tools where the interface is the product: content creation, tier-1 support, basic analytics. When the interface becomes natural language, that moat collapses.

Software bifurcation map by AI disruption risk: ERP cybersecurity and observability at low risk, core CRM and dev tools at medium risk, content creation tier-1 support and basic analytics at high risk, showing the market is pricing every category as if it faces equal threat

The bear case is correct about systems of engagement and the bull case is correct about systems of record, so the market is wrong to price them identically. Selling both at the same multiple compression implies that switching costs, regulatory requirements, data gravity, and enterprise procurement cycles have all vanished at once. Gartner predicts over 40% of agentic AI projects will be cancelled by 2027. Salesforce’s Agentforce reached 18,500 customers in its first year, the fastest-adopted organic product in company history. Incumbents like Salesforce are absorbing AI into their existing businesses.

The bear case is a zero-sum repricing where AI agents compress existing software revenue by eliminating seats and commoditizing interfaces. The bull case is a positive-sum expansion where the surviving software companies capture the white-collar services market. The cost of intelligence has fallen 99.7% in two years, according to a Stanford study cited by a16z. Cumulative AI infrastructure investment is expected to exceed $3 trillion by 2030. With that much capital going in, I expect the definition of “software” to expand until it includes most of the services economy.

I wrote recently about how passive flows create mechanical, price-insensitive selling that overwhelms fundamental buyers. This software sell-off is a textbook case. JP Morgan’s Murphy described index arbitrage basket selling, programmatic de-grossing, and passive flow liquidity vacuums. The IGV recorded its highest single-day trading volume in 25 years. JP Morgan’s follow-up argued the sell-off has gone far enough. BofA called it a paradox that “doesn’t make any sense.” History suggests that extremes like these (the 2016 LinkedIn panic, the 2022 rate-shock drawdown, the January 2025 DeepSeek crash) tend to mark the point where a sell-off turns.

I think the IGV at $80, with the most oversold RSI since 1990, prices in an extinction event that operating results don’t remotely support.