Skip to the report

Where should the computer sit?

It looks like an economics question and it has a physics floor underneath it. The review measured the round trip to every tier a robot could reach, so instead of a recommendation you get a calculator: say how fast your loop has to close, and see which tiers can serve it.

How fast does your loop have to close?

Every tier below carries a measured round trip. The closed-loop ceiling is what that round trip permits if the network were the only cost, which makes it a generous upper bound rather than a promise. Set a frequency and the tiers sort themselves.

Logarithmic, 1 Hz to 200 Hz, because the interesting range spans two decades.

The thing that is actually in the way

Two published system figures, from the same review, which between them reframe the whole question:

MeasurementAgainstGrade
An off-board policy inference call, 86 ms total of which 13 ms is networkself-published
A published humanoid model, one forward pass: about 160 ms a 33 ms control periodreported

In the first, the network is 15% of the delay and the model is the rest. In the second there is no network at all and the model still misses its deadline by five times. Moving the computer closer cannot fix either one. That is the review's central point about siting: for policy inference the question is already economic rather than one of feasibility, because a 44.08 ms round trip to a cloud region comfortably meets a published re-inference cadence of 0.5 to 0.8 seconds.

And the part that no siting decision can move

The radio access network component alone measures 7.02 to 7.60 ms, and is invariant to where the server sits. Whatever you do about siting, that is already spent. The entire spread between the nearest off-board tier and the furthest is 26.81 ms, a 60.8% reduction that buys you the difference between 22.7 Hz and 57.9 Hz, and nothing at all if the loop you need is faster than either.

What binds the eleven hypotheses

The review resolves eleven priced hypotheses rather than issuing one recommendation. Reading only the binding-constraint column: energy is the modal answer, and nine of the eleven bind outside the computer itself. Only two are held by the computation: one by accelerator utilisation and one by the output-to-input token ratio, which is a property of the model rather than of where you put it.

Round trips on this page are the review's verified measurements; the closed-loop ceilings are modelled from them and labelled as such in the source. Read from Siting the computation (TR-2026-33). Constraint vocabulary: the eight.