🤖 The week AI killed software

Last Monday, $285 billion of market cap evaporated from software, financial services, and asset management stocks. Thomson Reuters lost $8.2 billion. In a single day. LegalZoom dropped 20%. India's Nifty IT index posted its worst month since October 2008 — worse than the financial crisis.

A trader at Jefferies coined the term. "SaaSpocalypse." He described the mood on the desk as "very much 'get me out' style selling."

The trigger was Anthropic releasing Claude Cowork plugins for legal, financial, and sales workflows. The market's conclusion was instant: why pay for ten software licenses when one AI agent handles the workflow?

But the selloff isn't the story.

The story is why now. Why this week and not six months ago? What snapped.

In 2011, Marc Andreessen wrote the famous essay "software is eating the world." Back then the consensus investor believed Facebook and LinkedIn could never deliver value (seriously). 

Since them underneath the trillion-dollar hyperscalers emerged hundreds of public SaaS companies — Salesforce, Adobe, Intuit, ServiceNow — valued between $2bn and $100bn. The 2010s were the golden era: 75%+ margins, 100%+ customer retention, no physical inventory. Build a nice interface on a workflow, charge per seat, watch the NRR compound.

The SaaS boom happened because software disrupted spreadsheets. Every janky process that lived in Excel got replaced by a polished product with a login page and a per-seat price tag.

The per-seat model had zero marginal cost. Every incremental new user added no extra cost. So margins remained incredibly high (80%). This wonder drug is now existentially threatened.

Because AI agents are killing SaaS companies.

You're going from a department of 10 Bobs using SaaS tools and spreadsheets, to 5 Bobs and 50 AI agents — making custom workflows that fit the problem exactly. Not the problem the SaaS vendor decided to solve. AI agents cost pennies per task, work 24/7, and can read Bob's macro and the three Confluence pages that explain what it does and the Slack thread where Bob argued with finance about the formula.

The interface and its ability to integrate used to be the value. Now the outcome is the value, and the interface is optional.

That's the repricing. It happened in an afternoon.

The market considers UIs — CRM, project management — to be the most vulnerable. Companies that are data or security first (like Palantir, Datadog or Palo Alto Networks) are continuing to grow. As are vertical specialists (like Veeva for life sciences).

Industry-wide, IT budgets are up 8%. AI budgets are up 100%.

AI investment is a black hole — capturing every available dollar of IT spend and investor appetite from other stocks (and even crypto), because AI is delivering ROI.

We're well past death by proof-of-concept. But we have a new problem.

Most people use ChatGPT as better search. They're missing 99% of the value.

The AI companies call this a capability overhang. The models have been able to do far more than anyone uses them for, but nobody noticed because the tooling hadn't caught up. Like snow building up on a mountain ledge — the weight was there. It just needed a crack.

This week, four cracks appeared at once.

1. The models got better — and started improving themselves.

OpenAI released GPT-5.3-Codex on February 5th. 25% faster. Half the tokens. First model classified "high capability" for cybersecurity.

But the line that should stop you mid-scroll: it's the first model instrumental in creating itself. OpenAI used early Codex to debug its own training runs. The model helped build the model. Better models → better tools → better models. The flywheel is spinning.

2. Models escaped the chat box.

The biggest rate limiter to AI value hasn't been intelligence — it's been the interface. That's over. Claude is now working directly in Excel, Notion, Linear, Slack — not as a sidebar, but as the analyst. It can take a request via Slack, run a cashflow summary, update the task list, store it in Notion, and notify you when it's done.

OpenAI launched Frontier — a full governance platform for autonomous agents. Identity management, permissions, audit trails, a "Business Context" layer that plugs into data warehouses and CRMs. Early adopters are enterprises like HP, Uber, Oracle, State Farm, and Intuit. Governance is the unlock.

Perhaps most exciting for legacy enterprises: the ability to wrap their existing systems of record and transform them.

The agent goes to the tool now. Not the other way around.

Research from METR shows autonomous task horizons doubling every four months. At 30 minutes: code snippets. At 5 hours: module refactoring. At multi-day: full codebase audits. Each doubling unlocks exponentially more of the total automation pie.

Cursor built an agent harness that orchestrates thousands of agents simultaneously. Their FastRender experiment had 30,000 commits, 2,000 concurrent agents, and a million+ lines of Rust. With a planner/worker architecture with specialized roles.

This isn't "AI helping a developer." This is a development team made of software.

The capability overhang avalanched.

There used to be a reliable lag. Startups experiment, enterprises follow 12–18 months later. That breathing room gave incumbents time to watch, assess, and respond at committee speed.

• Goldman Sachs embedded Anthropic engineers directly in their tech teams for six months. AI agents for trade accounting, reconciliation, client onboarding. Goldman's CIO told CNBC they started with a coding assistant and quickly realized Claude's reasoning was strong enough for actual financial operations. CEO David Solomon announced a multi-year AI reorganization: "constrain headcount growth."

• Norges Bank — $1.7 trillion sovereign wealth fund, ~650 staff — reported ~20% productivity gains. 213,000 hours saved. 100+ FTE equivalent. At a fund that already ran lean.

"With Claude, we estimate that we have achieved ~20% productivity gains, equivalent to 213,000 hours." That's more than 100 full-time positions.

Instead of another PoC or Accenture-infested pilot — this is the baseline. This is how these institutions are choosing to operate.

This is ROI. And it’s all being unlocked from AI getting scarily good at coding and tool use.

One number. According to SemiAnalysis, 4% of all public GitHub commits are now authored by Claude Code. Not "assisted by" — authored by. A 42,896x increase in 13 months.

Dylan Patel projects 20% of daily commits by December. One model. One company.

This matters for the SaaSpocalypse because coding unlocks every other category of tool use. Code is how you build internal tools. Code is how you connect APIs. Code is how you replace a SaaS seat with a custom workflow. Coding is the wedge that makes agent-powered everything possible.

What's happening today in software will happen to all knowledge work in the next 12 months.

Imagine 4% of financial analysis done by AI agents. Or 4% of legal contract reviews. Or 4% of compliance checks. Then 20%.

That's the SaaSpocalypse.

If you want to see what every company looks like in 5 years, look at what Ramp is doing today.

They're going underneath their SaaS tools — to the APIs, the data layer — exposing those to developers who build custom internal UIs.

Not "add AI to our product." Not "use Copilot." Go to the API layer.

• Top tier: Unique data and IP. Bloomberg's decades of financial data. Palantir's ontology. Datasets that agents need to access, not replace.

• Mid tier: Great APIs and composability. Companies that make functionality easy for agents to consume. Stripe. Plaid. Twilio.

• Bottom tier: Commoditized UI. Pretty dashboards on data anyone can access. The CRM with a drag-and-drop interface and nothing unique underneath.

If you're still using ChatGPT as better Google, you're driving a Formula 1 car in first gear.

Some of these stocks will bounce. Software isn't going to zero. But the structure of how software creates and captures value is permanently changing. The seat is dying. The API is rising. The agent is the new user.

Zero marginal cost may be dead. Replaced by the API call and intelligence APIs, where margins at best are closer to 60%. This is the new normal that just got priced in.

The models just learned how to do your job. And your software's job.

Your job now is to learn how to use these tools and adjust your assumptions accordingly.