Anthropic shipped Claude Cowork plugins (legal, finance, sales, marketing, all automated) and days later dropped Claude Opus 4.6, a model that spins up and coordinates entire AI agent teams.
Nearly a trillion dollars in software market cap disappeared. Salesforce down 26% year-to-date, the second-worst stock in the Dow. Intuit, Adobe, Workday, Autodesk, ServiceNow all got hit. LegalZoom dropped 20%. The S&P 500 Software & Services Index fell 20% from its October peak.
What makes this unusual is that analysts were actually raising earnings estimates for these companies while the sell-off was happening. The near-term numbers look fine. The market isn't pricing in a bad quarter. It's pricing in the possibility that these companies don't exist in five years.
Why This Is Structural
Normal bad news expires. Missed earnings, the market moves on. Regulatory risk gets priced in.
This sell-off is about something deeper. The entire SaaS model was built on one assumption. Building software is hard and expensive, so customers will pay recurring fees for pre-built tools. That assumption held for decades. AI is breaking it. Every new model makes software cheaper to produce, and foundation model companies are competing to accelerate that trend. One person can now vibe code a product in a few days that used to take a five-person team six months. In some cases, the tool for humans isn't even needed anymore because AI agents do the work directly. Fewer humans in the loop means seat-based pricing is becoming obsolete.
Newspapers Already Showed Us
This analogy comes from Ben Thompson's Microsoft and Software Survival. Mark Leonard at Constellation Software has been circling the same idea for years.
Old newspapers had expensive distribution. Printing presses, physical delivery, geographic monopolies. A paper in Seoul didn't compete with the New York Times. Margins were fat and moats were deep.
Then the internet dropped distribution cost to zero. At first this looked like a gift for publishers. Suddenly they could reach the entire world instead of just their delivery area. But free distribution wasn't exclusive. Every other publisher got it too, and so did bloggers and anyone else with something to say. The shift from scarcity to abundance destroyed the economics. Geographic monopolies vanished, competition went global, advertising moved to platforms. Most papers died.
Old software had the same structure. Coding was expensive, engineers were scarce, building enterprise software took massive teams and years, and switching costs locked customers in. The result was a neat, siloed SaaS ecosystem. Find a business function, write an app, hire a sales team, IPO.
Now AI is doing to coding what the internet did to distribution. And the same paradox applies. Cheap code looks great for any individual software company because they can build faster and more efficiently. But it's catastrophic for software companies collectively, because every competitor gets the same advantage, the value they can extract shrinks, and their products get commoditized fast.
What Survives
Not every newspaper died. The New York Times and the Wall Street Journal survived because they had genuinely irreplaceable content and found new models. But they couldn't survive on distribution alone. The ones that made it had to offer something the internet couldn't replicate.
Software companies face the same reckoning, and you can already see where the line falls. For decades, the dominant model was the system of record. Store the data, charge per seat, lock in the workflow. Salesforce doesn't close your deals. Workday doesn't run your payroll. They hold information while humans do the actual work. That was a fine business when building an alternative was expensive. It's not anymore.
The value is shifting from systems that record work to systems that do work. A tool that actually reviews your contracts, files your taxes, or runs your ad campaigns is fundamentally different from a database you pay to access. AI agents don't just organize, they execute. That changes what a moat looks like.
The surviving software companies will need a different kind of moat. The moat has to come from somewhere that cheap code can't replicate. The biggest opportunity is in replacing human labor directly. The labor arbitrage is massive. A task that costs a company $200/hour for a human specialist can be done by an AI agent for pennies. Any software company that can capture even a fraction of that spread has a real business. The ones still charging seat licenses for access to a database don't.
The common pushback is that today's agentic AI services look rough. That's the wrong thing to focus on. Disruptive technologies almost always look worse than incumbents at first because they enter through gaps the existing products don't serve well. What matters is whether they're cheap enough and whether they meet a real demand. If both are true, the polish comes later and the incumbents never catch up. Dismissing agentic AI because the first wave of products looks scrappy is the same mistake newspapers made about blogs in 2004. Markets don't wait for perfection. They flip at an inflection point.
References
- Ben Thompson, "Microsoft and Software Survival," Stratechery, February 3, 2026.
- Ben Thompson, "The AI Unbundling," Stratechery, September 15, 2022.
- Ben Thompson, "Economic Power in the Age of Abundance," Stratechery, 2014.
- Anthropic, Claude Cowork product page, 2026.
- Anthropic, Claude Opus 4.6 announcement, February 5, 2026.