The screens within a trading desk in Midtown Manhattan flickered red simultaneously on the morning of February 3, 2026. Salesforce went down hard. Thomson Reuters plummets 16% before lunch as Adobe slides. The one-day decline in the Goldman Sachs software basket was the most severe since the tariff shock of the previous year. The software industry lost almost $300 billion in market value by the end of it. A recession wasn’t the cause. Rates weren’t involved. It was a plug-in.
The announcement of Anthropic’s new “Cowork” tool, an agentic AI system that can automate financial analysis, compliance assessments, and CRM operations, came as a quiet thunderclap. Next was Claude Opus 4.6, which managed “agent teams.” Frontier was released by OpenAI. Investors were suddenly faced with an awkward question: what if the conventional per-seat SaaS model is already out of date? It seems as though the panic was unavoidable.
| ategory | Details |
|---|---|
| Event | “SaaSpocalypse” Market Selloff |
| Date | February 3–4, 2026 |
| Estimated Market Loss | ~$300 Billion |
| Trigger | Anthropic’s “Cowork” AI agent plugins |
| Key Impacted Sector | SaaS (Software-as-a-Service) |
| Related Index | S&P 500 Software & Services Index |
| Industry Context | AI agent-driven automation |
| Reference | https://www.cnbc.com |
Subscription licenses offered per human user were the lifeblood of software companies for twenty years. Every seat meant steady income. Growth required department expansion or the appointment of additional staff. It was a straightforward model. predictable. profitable. AI agents, meanwhile, do not require chairs.
The revenue logic breaks down if a single autonomous system is able to do functions that were previously done by costly software suites, such as drafting contracts, reconciling finances, updating CRM fields, and analyzing data. Anything that appears to be replaceable by an AI worker is turning investors away.
The S&P 500 Software & Services Index is currently down around 20% for the year after dropping more than 4% in a single session. Once-defensive names including FactSet, Moody’s, and S&P Global witnessed steep drops. Since January, HubSpot and Adobe have experienced double-digit percentage declines. It seemed more like a change in the story than a correction. However, this is where the narrative gets trickier.
As this was happening, it was difficult to ignore the fact that infrastructure names were subtly gaining and software equities were falling. According to Jensen Huang of Nvidia, next-generation reasoning models can use up to 100 times as much processing power per task as previous AI systems. That figure persists.
Because inference is the foundation of every AI agent, including automated compliance reviews and synthetic financial analysts. Inference is a continuous learning process. It never stops. Every task, query, and document that is produced uses computing power. Somewhere that computer must reside.
Recent estimates indicate that Amazon, Google, Microsoft, and Meta will collectively spend close to $700 billion on capital projects in 2026. Google alone has $175 billion to $185 billion planned. Analysts were taken aback by Amazon’s $200 billion capital expenditure forecast, which exceeded projections by about $50 billion. Up to $135 billion is being guided by Meta. Based on its quarterly pace, Microsoft makes roughly $145 billion a year. These investments are not incremental. They are on an industrial scale.
Conversations with investors frequently bring up the late 1990s fiber buildout as an analogy. The bubble burst, and many dot-com enterprises went out of business. However, the fiber lines installed during that flurry served as the foundation for the contemporary internet. Infrastructure survived. The similar situation can be developing right now.
Data centers are being built faster than ever before, while SaaS companies argue about pricing schemes and seat compression. Tens of thousands of GPUs are being ordered. To manage the load, power grids are being modernized. Which AI agent prevails is irrelevant to the physical layer, which includes chips, networking, and cooling systems.
The worry is genuine. There are real existential risks for certain software companies. Subscription economics will shift if an AI agent can perform all of a suite’s functions at a lower marginal cost. It’s possible for margins to compress. Models of pricing will change. Workforce composition may become smaller. However, markets frequently ignore demand generation in favor of disruption.
The volume of inference rises with each automated process. Each AI “coworker” increases the amount of computing power. Whether per-seat SaaS will completely vanish or change into subscription bundles that combine AI and humans is still up in the air. The underlying infrastructure need is increasing more quickly than expected, which is more certain.
By late afternoon, the terror had subsided but not completely disappeared from that Midtown trade floor. Instead of talking about churn rates, traders leaned across desks to discuss inference loads. The change in discourse may be more significant than the actual selloff. There was more to the $300 billion wipeout than a simple reset in valuation. It was a sign.
The market for software isn’t broken. Autonomous agents perched atop a massive, energy-hungry computational stack is the new center of gravity around which it is being re-priced.
It seems likely that February 3–4 will be viewed as a pivot rather than a crash as we watch this play out. Overnight, the SaaSpocalypse might have destroyed billions.
