AI Is Already Splitting the Software Industry in Two

Inside the U.S. software sector, the gap between this year's winners and losers has stretched to roughly 50 percentage points, the widest such split in over a decade. The market is not killing software. It is sorting it, and the line it drew this year is one any reader can learn to read.

The sorting shows up most clearly in the iShares Expanded Tech-Software ETF (ticker IGV), the largest U.S. software-sector fund. The top quartile of its holdings is still posting positive returns in 2026 while the bottom quartile has given back years of gains. The split is not random. Cybersecurity, observability, and vertical software (the kind sold to a single industry, like restaurants or hospitals) are being priced as AI-resilient. Horizontal platforms and general-purpose tools (the kind meant for every team in every company) are being priced as exposed.

That map is the part the headlines have been missing. The "SaaSpocalypse" narrative, the finance-influencer shorthand for a software-sector crash, treats the sell-off as one story. The chart says otherwise. Investors are not abandoning the sector. They are re-ranking it by moat, the structural advantage that keeps customers from switching to a competitor.

The price of inference, meaning the cost of running an AI model to produce an answer, is falling roughly five times faster than the price of a personal computer fell during the 1990s buildout, according to the Oztalking analysis of the Jevons paradox in AI. The economic term for cheaper supply unlocking more demand is the Jevons paradox, and applying it to AI means the bottleneck for the next decade is not the model's cost but what the model can do inside a customer's workflow. That tilts the moat toward software that owns a hard-to-replace data feed, a compliance audit trail, or a regulated industry. It tilts the moat away from software that mostly wraps a general capability a model now provides for free.

A 50-percentage-point spread inside a single sector is not a normal event. The closest historical analog, flagged by the same analysis, is what happened to U.S. newspaper stocks between 2002 and 2007: classified advertising revenue was collapsing, but the market sold newspaper shares for years before the earnings actually fell apart, because markets tend to price structural shifts before the data catches up. The 2026 software dispersion may carry the same warning.

The productivity lag is the second leg of the story. A Goldman Sachs economist, reported in Fortune and CryptoBriefing, has argued that AI's measurable effect on output may not show up in the macro data until around 2034, roughly fifteen years after the technology started landing in workplaces. The personal-computer buildout took about that long to show up in productivity statistics after the machines shipped. If the same lag holds for AI, the 2026 sell-off is pricing a future the data will not confirm for years.

A counterweight to the bearish read is visible on the demand side. Consumer spending on paid generative-AI subscriptions has grown across cohorts in 2025 and 2026, per CBS News reporting, and the academic literature is starting to treat large language models as a market for intelligence with its own supply and demand curves, per the NBER working paper w34608. Money is moving into AI-adjacent software at the same time money is leaving general-purpose tools.

Two things to watch. The a16z charts-of-the-week read treats the dispersion as a sorting event inside the sector, not a uniform crash. If the spread closes in 2026 because the bottom quartile rallies, the map held and the winners were simply oversold. If the spread widens further, horizontal platforms face a longer re-rating. The data series is weekly. A reader can pull it up on Monday and decide which side of the map the latest print falls on.