The threat AI poses to software
The selloff in software stocks, called a "SaaS apocalypse," looks like an underreaction rather than an overreaction. Software has long been a great business, but AI autocoding tools change the math. They can't replace enterprise software today. Give them 5 years, or 10, and they will be able to replicate most of it. That pushes out a real terminal value for these businesses, 10 to 15 years out, a risk that never existed before.
Some firms still hold an edge over the next 5 years. Service Now builds tools for managing AI agents, which could bring in enough new revenue to offset whatever it loses to AI over time.
The key question: for which companies was the moat just the code? If your only advantage is the technical ability to write software, you are in trouble, because AI coding tools turn what used to be a moat into a cheap, common commodity. If your moat rests on something else - customer relationships, brand strength, network effects, or other intangible assets AI coding agents can't touch - you can survive. Many of those stocks are down hard yet remain positioned to make it through.
History backs this up. Study past disruptions - brick-and-mortar retail, the newspaper collapse - and you find heavy headwinds plus wider dispersion between winners and losers. Some companies survive and then thrive. The New York Times lived through the newspaper decimation. Walmart lived through the retail apocalypse. Stocks like Service Now, Adobe, Salesforce, and Intuit, once darlings of growth investors, are down 60 to 70% and now face real challenges.
A useful test when judging disruptors: ask what the incumbents can't do themselves that makes a new product better. Incumbents that can apply AI on top of the customer loyalty they already own hold an advantage.
Memory: priced like a cyclical, growing like AI
The chairman of SK Group, the company behind SK Hynix, warned that memory prices are abnormally high and that supply must rise to prevent "chipflation." Chipflation is real. Prices have jumped a lot.
Memory makers normally run 30 to 40% gross margins in a good year, 50% in a great one. They are now doing 80% plus and growing fast, yet the market still prices them as cyclical companies and bets on a downturn. This mirrors the semicap equipment cycle around 2010, when stalwarts like Lam Research and ASML broke out of that cyclicality and became very good stocks.
Demand comes from AI capex, which is eating huge amounts of memory. If large spatial models start training on video content instead of text, that is another 5x to 10x jump in memory needs. AI requires more compute, more network throughput, and more memory. Memory is the standout, especially with these stocks trading at six and seven times earnings, which makes them still very attractive.
Cheaper Chinese models and the capex question
Since the Kimi release and a wave of improved Chinese models that are open source and, more importantly, open weight, money has started rotating out of the big trade that drove the market over the past year: semiconductors, memory, hyperscalers, anything tied to AI capex.
The worry is simple. If you need less compute to run an equal model, the buildout may be overdone. The same overbuild hit the telecoms in the internet boom and the railroads a hundred years before that. The risk is real.
The counterargument is Jevons paradox. When the price of a commodity falls, demand can rise, and if demand is elastic enough, use cases for AI that were once too expensive open up. Total consumption then climbs. Chip stocks are less exposed here, because they benefit whether the model running is open source or closed.
It is more concerning for the closed-source labs, the OpenAIs of the world. They haven't gone public yet, but trillion-dollar valuations get thrown around. Competition is rising, not only from Chinese open source but from US firms like Meta and Google. It has become a crowded space.
There is a plausible future where AI revolutionizes how society works, yet the production of intelligence itself becomes a commoditized utility. In that case, value flows to other parts of the stack rather than the labs. One sign of how early this still is: only 2% of people currently pay for these models.
