AI Is Oil, Not God

The prophets sell you the wellhead

The economic value of AI accrues to the people who refine and distribute it, not to the labs that pump it out of the ground. Treat AI as a commodity input rather than a sovereign intelligence, and you get the map right: value sits in the application, integration, and distribution layers, not at the model.

Packy McCormick made this case at Not Boring under the same title, and it’s the correct frame. The dominant story about AI runs the other way. Labs are prophets, models are oracles, scale is salvation. Every incremental parameter count arrives as revelation. The implied conclusion is that whoever builds the smartest model wins everything downstream, because intelligence is the scarce thing and the lab owns the source.

That conclusion is wrong, and it’s wrong in a way that costs operators real money when they build on it.

Intelligence is commoditizing on a schedule

Start with the price. The cost of a given level of model capability has been falling on the order of 10x per year across the frontier labs. What cost a dollar to generate in early 2023 costs pennies now, and the pennies buy a better answer. That is not the price curve of a scarce, defensible asset. It’s the price curve of a commodity in a competitive market.

The supply side tells the same story. At least four labs ship frontier-class models: OpenAI, Anthropic, Google, and a rotating cast including Meta and the strong open-weight releases out of China. When a customer can swap one model for another behind an API in an afternoon, the model is not a moat. It’s an input. The switching cost that was supposed to lock everything in never showed up, because the interface is a text prompt and the outputs are close enough substitutes that most applications route to whichever model is cheapest that week.

The counter you’ll hear is that the frontier keeps moving, so the lead compounds. It doesn’t compound the way the story needs it to. A six-month capability lead is real, but it’s a lead in a race where the prize keeps getting cheaper to reach. Being first to a capability that costs a tenth as much in two quarters is worth far less than being the company that owns the workflow the capability runs inside. The lead is temporary, and the workflow is what lasts.

Crude is worthless until somebody refines it

Oil is the right analogy and God is the wrong one. Crude out of the ground is nearly useless. You can’t put it in a car. The value chain that turns crude into a functioning economy (pipelines, refineries, distribution networks, the retail station on the corner) is where the margin lives. The wellhead is a commodity business with brutal economics and constant price wars. The refiner and the distributor capture the spread.

Raw model output is crude. It’s a probability distribution over tokens. It doesn’t know your customer, your data, your compliance requirements, or the specific shape of the job someone is trying to get done. Turning that raw capability into something a person or a business will pay for takes refinement: the data pipeline that grounds the model in your context, the workflow that sequences its outputs into an actual task, the distribution surface that puts it in front of the user at the moment of need, the trust layer that makes the answer safe to act on.

None of that ships from the lab. All of it is where defensibility accumulates.

Margins show up where the hype doesn’t

Look at where the durable businesses are forming. The application-layer companies that are compounding, in legal, in coding, in support, in finance, are the ones that own a proprietary workflow and the data exhaust that comes with it. Every task a user completes makes their system a little better at the next one. That’s a self-learning loop the model provider doesn’t have access to, because it lives in the application’s data, not in the weights.

Meanwhile the foundation-model layer runs a capital-intensive business with falling prices and undifferentiated output, the classic profile of a commodity producer. Enormous training runs, enormous inference bills, and customers who will leave for a two-cent price advantage. That’s not a knock on the labs. Refineries need crude, and the labs are extraordinary at producing it. It’s a statement about which layer keeps the margin.

The pattern held in the last platform shift too. In cloud, the money didn’t concentrate in raw compute, which became a price war between three hyperscalers. It concentrated in the software built on top: the CRMs, the data warehouses, the vertical applications that owned a workflow and the data inside it. Compute was the input. The application captured the customer. AI is running the same play, faster, because the input is commoditizing faster than compute did.

Build refineries, not shrines

For an operator, the instruction is direct: invest in the layers that refine and distribute, not in faith that your model choice is a moat.

Concretely, that means owning your data pipeline as a first-class asset, because grounding and context are what turn generic capability into a specific answer nobody else can produce. It means building the workflow so tightly around the real job that ripping it out costs the customer more than the switching-cost math on the model ever would. It means treating distribution as strategy, since the next surface where users meet AI is conversational and agentic, and the business that’s already there when they arrive owns the relationship. And it means designing trust as a constraint from the start, because the refined product a business will actually deploy is the one it can rely on, not the one with the highest benchmark score.

What it does not mean is betting the company on model supremacy. Pick the model that’s cheapest and good enough for the job, architect so you can swap it when the price drops again, and put your durable investment into the layers the lab can’t replicate. The model is the crude. Your refinery is the business.

The worship framing gets one thing right and everything downstream wrong. It’s right that the capability is real and improving fast. It’s wrong that the value flows to the source. Value flows to whoever stands between the raw capability and the person willing to pay for a finished answer. The labs pump, and the refiners get rich. Build accordingly.