Cheap intelligence makes the incumbents richer

Part two of three on what happens to the market as intelligence commoditises. Back to the bet against itself; on to the internet goes headless.

The loudest prediction in software right now is its own funeral. If anyone can point a frontier model at a problem and get a working tool back, the argument goes, then every SaaS company is a dead man walking, its product about to be rebuilt over a weekend by a competitor who pays nothing for the intelligence. The apocalypse is real. It is just aimed at the wrong layer.

The mistake is treating cheap intelligence as something that destroys value. It doesn’t destroy value; it relocates it. When the price of a capability falls fifty times a year, the money that used to sit in that capability does not evaporate. It moves to whatever the now-cheap thing plugs into. Follow it down the stack and you can see exactly where it lands, and who is standing there.

The bottom of the stack is a utility

At the floor sits energy and compute: the datacentres, the silicon, the power to run them. This is xAI’s whole bet, and it is a real moat. Whoever owns the cheapest electrons and the most chips has a durable, physical advantage nobody can distil away.

It is also a brutal business. A capital moat with utility returns. You need a balance sheet rather than a product, and the economics come to look like a power company’s, because that is roughly what this layer is becoming. Margin does not pool at the bottom of the stack; volume does. The marginal cost of intelligence really is trending toward the cost of electricity, and the people who own that floor will earn what electricity utilities earn: a steady, capital-hungry, unglamorous return on enormous scale. Worth owning. Not where the money is.

The model layer melts

One layer up is the frontier model itself. Margin here is real and temporary, because the lead is. A model that is best today rents that position for about two quarters before the open-weight pack and the other labs close the gap. Pricing power lasts exactly as long as the lead does.

This is why a pure-model company is a hard thing to hold and a great thing to take public before the next release. The margin passes through this layer; it does not settle here. Hold the best weights for six months and you have a remarkable quarter and no business. Everything in the first part of this was a story about labs trying to escape this exact layer before it closes over them.

Margin settles where the model plugs in

Keep following the money and it stops moving at the layers a falling token price cannot reach.

The consumer half of this is already decided, and I’ve written it: distribution and context don’t commoditise. Whoever owns the surface a person speaks into, and the data about that person, captures the value of a commodity model dropped onto it. That is Google’s game, and it barely needs restating.

The enterprise half is the newer ground. A business does not run on a model. It runs on a system of record: the customer data nobody else holds, the workflow the whole company has wired itself around, the integrations into a dozen other tools, the audit trail, the someone-to-sue when it goes wrong. None of that commoditises when the model does. It just gets a cheaper, better engine bolted into it, and the margin on the whole bundle goes up.

That is the part the apocalypse misses. Cheap intelligence is an input. If you sell a product whose value is the intelligence, a falling input price is a falling output price, and you are in trouble. If you sell a product whose value is the data, the lock-in, the distribution, and the accountability, with intelligence as one input among many, then a falling input price is pure margin expansion. The same trend, opposite sign, depending on which layer you sit on.

Who actually dies

Two kinds of company are genuinely in the blast radius.

The first is the thin wrapper: a product that is a prompt and a logo, whose only asset is access to a model everyone else can also rent. No data of its own, no workflow nobody can replicate, no distribution. It was always renting its entire value from the layer below it. When that layer commoditises, there is nothing left to charge for.

The second is the pure-model lab that fails to climb. Owning the best weights for six months is not a business, and the labs that survive the decade are the ones that became enterprise and application companies before the model underneath them went free. Anthropic’s thirty-billion enterprise run-rate is precisely that escape in progress.

The incumbents inherit it

Which leaves the SaaS incumbent, the company everyone is writing the obituary for, as one of the best-positioned players in the entire shift. It already owns the two things that don’t commoditise: the proprietary data and the workflow the customer can’t leave. All cheap intelligence does is hand it a better engine, at a lower price, to drop into a moat it already dug. Microsoft, Salesforce, ServiceNow, Adobe, the whole system-of-record layer, don’t have to win the model race. They have to not fumble the integration, and they get the model itself for almost nothing.

I want to be honest about the condition this rests on, because it is a bet on a specific speed. It holds only where the incumbent can bolt AI into its moat faster than a challenger can rebuild the moat with AI. The case depends on workflow lock-in and proprietary data being genuinely slow to replicate. Where they aren’t, where the “system of record” was really just a database and a form that anyone can now generate, the incumbent has no moat and earns its apocalypse on schedule. The prediction is not “incumbents are safe.” It is sharper than that: incumbents with a real data-and-workflow moat get richer, and incumbents whose moat was only ever inertia get eaten, on the same timeline, by the same force.

Cheap intelligence is the best thing that ever happened to a company with a moat the intelligence can’t reach, and the worst thing that ever happened to a company that was selling the intelligence itself. The apocalypse is not coming for software. It is coming for the thin layer of software that was only ever a wrapper on a model, and for the labs that forgot to stop being one.