The AI infrastructure build-out is, by announced numbers, the largest coordinated capital deployment in the history of technology. The memory is contracted, the financing is structured, the board mandates are in place and running underneath all of it, through every layer of a stack that has never been more interdependent, is a set of load-bearing assumptions that are quietly giving way at the fab, in the financing architecture, and at the enterprise paying the bill.
A Lapse in Memory starts at the fab. Three semiconductor companies with different boards, different cost structures, and different exposure to the AI build-out are printing identical record margins on flat shipment volumes. The mechanism is a wafer that has not been started: an optionality game that benefits both sides of the contract while the commodity customer pays the spread.
The Keystone follows the financing of the build out. The reciprocal commitments between NVIDIA, OpenAI, Oracle, CoreWeave and SoftBank assembled an arch around a single load-bearing outcome. The constraint has since moved, from chips to the electrons that run them, before the arch was complete. The piece asks what happens to the structure when the keystone is slightly off.
The Token Sink arrives at the enterprise. The board mandate is in place but the vendor has restructured the contract and now charges per token. The engineering team has no control surface for what the model decides to consume, and the substrate forecloses the mitigations they would normally reach for. A blank cheque, drawn on something that has been optimised to pass its own evaluation tests.
The argument runs from the fab to the financing to the enterprise boardroom, and the unsettling part is that each layer assumed the one below it was solid.