AI's Toll-Road Turn Has Started

AI didn't get worse this week. The subsidy did.

Google trimmed model capabilities. Anthropic shrunk free-tier access. TheStreet ran a piece asking whether businesses will still pay once the productivity demo can't keep up with the productivity promise. The framing is demand-side: capability falls, ROI falls, churn arrives. Read on its own, it's a coherent panic.

Here's the steelman. Enterprise contracts signed in the last 18 months were priced against benchmarks that were quietly subsidized — frontier outputs sold at sub-cost to lock in seats, free tiers wide enough to manufacture organic demand inside target accounts. Pull the subsidy and the math collapses. Renewals get re-papered downward. The pipeline that priced AI as productivity-positive runs into AI as a line item somebody has to justify. If you're a CFO who approved a seven-figure spend last October because the free tier was already loose inside your engineering org, you have a real problem. That's the bear case and it isn't stupid.

It's also not what the smart money is reacting to.

In the same 24 hours TheStreet was asking whether businesses will pay, Berkshire under Greg Abel was reported to have rotated more than a third of the portfolio into AI-adjacent positions — building a top-five Alphabet stake while exiting Visa, Mastercard, and UnitedHealth. Note which company that is. Alphabet is one of the vendors degrading its own model. Abel didn't buy in spite of that. He bought because of it.

This is the toll-road turn. The give-away phase ends. The infrastructure phase begins. The first phase is when frontier vendors compete for adoption by eating their own margin. The toll-road turn is when they stop. Capability stops improving on the free tier not because the labs forgot how to ship. It stops improving because the unit economics finally got audited by someone above the CTO. Yahoo's framing of Berkshire's rebalance as "a bigger AI bet" obscures what's actually being bet on. Abel didn't buy a model. He bought the road the model has to drive on — compute, distribution, identity, search inventory. The product layer is allowed to disappoint because the toll booth runs anyway.

Here's what the demand-side framing misses. "Evidence is mounting" that businesses won't pay, but TheStreet doesn't name a cohort, doesn't cite a churn number, doesn't separate the free-tier hobbyist from the enterprise seat. The actual signal isn't churn. It's repricing. Enterprises will pay. They'll pay more. They'll pay it to fewer vendors. And the vendors collecting it will be the ones who own a piece of the road, not the ones who own a chatbot.

The one thing I genuinely don't know is whether the labs have the discipline to hold the line through a single earnings cycle of complaint. They might fold and re-extend a discount. That's the only thing that delays this, and it doesn't change the destination.

If you signed a 2025 AI contract with capability assumptions and discount tiers baked in, you are about to learn what your vendor's actual cost-of-goods looks like. The renewals are not going to be friendly.

Here's the dare. By Q3 2027, at least two frontier vendors will have raised enterprise prices by more than 40% on renewal while quietly capping the capability ceiling of the tier the original contract was sold against. Watch for it. The toll-road turn doesn't announce itself in a press release. It shows up in the invoice.