Industry Layers

Washington processes AI as one industry with one set of asks. It isn't one industry. There are at least four — the frontier that sells capability, the apps that sell workflow and buy inference, the hyperscalers that sell capacity, and the chips that sell the silicon underneath all of it. Their interests are diverging fast, and every policy on the table picks among them whether or not anyone says so.

The method is to place a company by where its revenue comes from, not by what it says it is building. The gap between the two is often large. xAI presents as a frontier lab and is understood in Washington as a peer to OpenAI and Anthropic, but its disclosed revenue comes substantially from renting Colossus capacity to Anthropic — a competitor's frontier lab as its anchor tenant. Google's capital is visibly moving from capability toward capacity, with third-party AI cloud revenue and TPU sales projected far above Gemini's. When these companies lobby, the income statement is talking, not the mission statement.

Once the layers are visible, positions that read as philosophy resolve into positioning. An argument that closed frontier leads are dangerous tends to come from a company without one. An argument that copying frontier models is dangerous tends to come from a company with one. Both risks are real; each happens to be the risk that would damage the arguer's employer. The same pattern runs through open weights, distillation, export controls, and datacenter siting, where identical patriotic vocabulary is used by layers in direct opposition — the chip layer's China argument opposes the export controls that the frontier layer's China argument demands.

Taxation is where the layer differences stop being rhetorical. Nearly every economic proposal on the map agrees on the spending side: money should reach displaced workers. The disagreement is almost entirely about the tax base, and the base is the whole policy. A tax on tokens consumed sounds like a tax on AI companies but is a tax on AI buyers, landing on the app layer and its enterprise customers while leaving frontier training compute untouched — a tax on diffusion, which sits awkwardly against the argument that diffusion is where the economic gains come from. A change to capital treatment lands on whoever owns the machines, which means the hyperscalers. Claims on enterprise value are the only family of instruments aimed squarely at the frontier. Industry assessments like Kelly's Horizon Fund name a payer without naming a base at all, and each possible answer sends the bill to a different layer.

Two absences are as informative as anything on the matrix. No proposal taxes training compute specifically, which is the one instrument that would fall on the frontier layer alone. And nothing reaches the chip layer except export controls, which are framed as national security rather than as revenue. The largest single concentration of AI wealth is the least exposed to every economic proposal in circulation.

The sharpest coalition line is not open versus closed. It is whether a company sells capability or sells capacity, because those two want opposite things from time. Capability wants speed. Capacity wants predictability and a contract long enough to underwrite the build. Chips want volume and are agnostic about who supplies it, which makes them a free agent in every fight.