Vertical integration and why the AI stack is collapsing into single owners

What has landed. OpenAI now designs its own inference silicon with Broadcom (Jalapeno, June 2026). Anthropic runs a heavily customised training and inference stack on cloud infrastructure partly optimised for its workload. Google runs the whole stack. Meta runs the whole stack. Microsoft acquires the model layer through capital and runs the compute layer through Azure. Amazon owns the compute layer and buys the model layer through partnership and investment.

The economic logic

Frontier AI is compute-bound. The training runs are large enough that a five per cent efficiency gain at the silicon layer is a first-order economic effect. Custom silicon, custom interconnect, custom cooling, custom scheduling, and model architectures tuned to the resulting hardware all compound. A firm that owns all five layers can extract the compounding. A firm that rents the stack cannot. Once one firm owns the stack, the others follow or exit the frontier.

What this changes for safety

Two things. First, the correction mechanism the programme cares about, whether beta is greater than k, is a property of the whole stack, not the model layer alone. A safety architecture that treats the model as separable from the substrate misses design choices that live at the silicon and the scheduler. Second, the audit surface collapses with the ownership. External evaluators no longer need cooperation from three companies to evaluate one system; they need it from one. That is either an efficiency gain or a governance failure, depending on who has the statutory authority to require the audit.

What the Chokepoint chapter argued

The book's Chokepoint argument identified the concentration of chip manufacturing capacity, and by extension the concentration of the full stack, as leverage. Leverage cuts both ways. It gives the small number of jurisdictions that host the concentration the ability to require safety architecture in exchange for market access. It also gives the small number of firms that own the concentration the ability to argue that competitive dynamics prevent them from adopting safety measures unilaterally.

Where policy has to catch up

The current export-control apparatus regulates the chip layer well and the model layer badly. It is not yet clear what regulating the scheduler layer, the interconnect layer, or the training-workflow layer even means as a legal instrument. The Chips Act 2.0 and the SAFE Chips work are chip-layer instruments. The Fable 5 order was a model-layer instrument. The intermediate layers are not yet named in the statute.

What the reader keeps

The collapse into single ownership is not a governance failure by itself. It changes the governance surface, and the surface has moved faster than the statute. The safety instruments the programme proposes, measured invariants of the whole stack under preregistered evaluation, are compatible with the collapse. Instruments that treat the model as separable from the substrate are not.