From the public website’s September 2026 source edition. This is architectural analysis and commentary. For current status and source documents, use the publication record.
What distributed systems spend to stay internally consistent while they run, and why that spend is not on anyone's bill.
Inference cost is measured. Transport cost is measured. Storage cost is measured. The work a system performs to remain coherent across its own boundaries is not measured, because no component owns it.
Every time an interaction crosses a service boundary, the system re-establishes identity, re-evaluates policy, re-synchronizes state, and reconstructs context that fragmented at the previous crossing. That work runs concurrently with inference. It produces nothing a user sees. It consumes capacity that could have produced output.
It does not scale with model capability. It scales with the structure of the system the model runs inside. And it does not add. It multiplies.
The pages below approach that spend from five directions: the physics of it, the energy accounting it distorts, the billing it breaks, the ledgers it leaves unreconciled, and the revenue it makes unrecoverable.
Read the set
- The Coordination Limit
The physics. Coordination overhead compounds across boundaries rather than summing across them, which is why capacity gains keep arriving without the throughput that should follow.
- Session and Energy
Avoided-emissions claims for AI require a boundary at which the carbon ledger closes. Without one, the accounting is an estimate rather than a record.
- Token and Authority
A meter counts consumption. It cannot price an outcome, because the outcome is not the unit the meter observes. This is why AI spend overruns surprise the organizations running them.
- The Token Toll Reconciliation
Four ledgers run during a live interaction and none of them closes against the others. The session is the only boundary at which they can.
- Session-Bound Revenue Surfaces
Intent, wait states, and inbound communication already carry commercial value. Without a governed boundary there is no auditable unit to bill against.
Why this is an authority problem
Coordination cost is usually read as an efficiency problem, which leads to efficiency answers: better caching, faster interconnect, fewer hops. Those help at the margin and do not touch the cause.
The cause is that no boundary in the system is authoritative for the interaction as a whole. Each component re-derives what it needs because nothing holds it on the component's behalf. Reduce the number of crossings and you reduce the multiplier. Give the crossings a single authority to reference and the multiplier collapses into a sum.
That is the same constraint described in Authority Surfaces under a different vocabulary, and one of the recurring patterns catalogued in The Seven Convergent Failure Domains.