The industry meters tokens.
Tokens are linear. The cost is not.

Attention is quadratic. The invoice is not.

The unit of account in commercial AI is the token, and tokens are counted linearly. But the dominant cost of a transformer holding a long context is not linear — the work of attending scales with the square of what the model must hold in view. That term appears on no invoice, no dashboard and no pricing page. It is paid in silicon time, memory bandwidth and electricity.

It is measurable. Charge every call for the context it actually carried, apply the identical formula to both systems, and the difference between two ways of operating the same model stops being a matter of opinion.

Input consumed: 289,843,671.0 prompt tokens against 131,484,378.0. Output produced: 803,740.0 against 1,076,583.0 — the system that ate more tokens produced less. MagLev’s output total includes user-facing conversational streaming, a category largely absent from a protocol-constrained coding agent, so every ratio quoted here is conservative in MagLev’s disfavour.

One system’s cost of work compounds as it works. The other’s does not.

Both systems began the day at a comparable price per unit of output. The chart below indexes each system against its own opening hour, so neither is measured against the other’s scale — only against where it started.

What it costs to remember what you were doing.

Both systems were shut down cold overnight and resumed the next morning — a symmetric, controlled test of resumption. MagLev was additionally subjected to two deliberate process kills inside a thirteen-minute window while under load; its counterpart was never killed.

Calls of that shape — the whole working state rebuilt from nothing before any new work can begin — account for 79.1% of its total attention compute across the benchmark and 89% of every token it processed fresh. The single worst of them is 11.23% of its own run in one call. MagLev’s worst single call is under one percent of its own.

Every unit of attention is state moved through memory.

Quadratic attention compute is not an abstraction: it is memory traffic. Each unit corresponds to a fixed quantum of key-value state moved through the serving hardware. To produce a single token of output, the machine must read the entire context that token is conditioned on — so the cost of one more word rises with everything already said.

The compute ratio converts directly to a serving-energy ratio, because every architecture constant cancels between two systems running the identical model on the identical fleet. Absolute energy figures carry reference-architecture constants and are therefore estimates derived from measured token surfaces, not measured device or datacenter electricity. Against a serving-fleet reference architecture, the same measured delta for this one benchmark — one user, one day — falls in the band of a single household’s full day of electricity, against roughly a dishwasher cycle for the winning run.

One measured delta, scaled one honest layer at a time.

Everything below this line is extrapolation, and is labelled as such. It takes the one measured per-user, per-day delta and multiplies it by stated user counts and duty cycles, with no other assumptions. It is offered because the measured number is small and the industry is not.

Compounded, the four layers put the waste removed by operating the same models this way on the order of tens of gigawatts continuous — the scale at which new generating capacity is planned, permitted and built over years. These are illustrative extrapolations of a single measured run pair, presented as arithmetic rather than forecast. They also assume zero growth in token consumption, which no industry forecast does: every figure here is the smallest version of itself that will ever be true.