Siting the computation.
A machine moving through the world produces sensor data and consumes decisions, and the computation that turns one into the other has to happen somewhere. The plausible locations span four orders of magnitude in distance, from the machine itself to a hyperscale campus hundreds of kilometres away. This review surveys the published evidence for each, grades every figure it uses, and derives the physical limits that bound the most distributed end of the range.
Grant Markhart, Industrial Research Fellow · The Charlot Lab & The Hiner Lab
176 TWh
US data centre electricity, 2023
4.4% of national consumption, up from about 76 TWh in 2018. A bottom-up model built from equipment shipments, not a metered total.
verified (LBNL)
~7 ms
radio term that does not move
Measured 5G round trips run 17.27 ms to a node inside the operator's network against 44.08 ms to a cloud region, but the radio access term is invariant to where the server sits.
verified (measured)
1.6-2.0×
marginal against average grid CO2
The environmental comparison between topologies turns on which factor is used, and on utilisation, more than on cooling.
verified (EPA)
0.12-1.52%
duty a harvested 100 cm² aperture sustains
For a 7 W embedded inference module. The same aperture sustains tinyML-class inference continuously.
modelled
What the evidence supports.
The siting question is open for reasons specific to embodied systems, and the literature that bears on it is scattered across four disciplines that rarely cite one another.
Latency is the argument usually made for distribution, and it is weaker than it is usually stated. Measured 5G round trips are 17.27 ms to a node inside the operator's network against 44.08 ms to a cloud region, but the radio access term of about 7 ms does not move when the server moves. Published vision-language-action policies run inference every 0.5 to 0.8 seconds behind chunked action sequences, so a regional round trip leaves substantial headroom for policy inference. The workloads that genuinely cannot be offloaded are bounded by certifiability rather than by milliseconds, and no reduction in round-trip time relaxes that.
The environmental comparison changes character with scale. For a centralised accelerator the dominant term is operational electricity; for a device it is the carbon embodied in manufacture, and the crossover depends on utilisation and on grid intensity rather than on any property of the hardware. Marginal and average emission factors differ nationally by a factor of 1.6 to 2.0, which is large enough that the choice of factor decides several published comparisons on its own.
The review's central finding is that the siting question is not currently settleable from published evidence, and the reason is specific: the two quantities that would decide it, measured accelerator utilisation by facility class and joules per delivered inference on a named embodied task, are not published by anybody. The review recommends no siting decision, forecasts no market, and states what would have to be measured to close each open question.
18 located topologies.
Every distributed topology the review located, with its defining property and the deployment record found for it. The withdrawn programmes are included because a survey that records only what launched would misstate the category's history.
Every distributed topology the review located, with its defining property, the deployment record found for it, and the grade of that evidence. Self-published entries are the operator's or vendor's own account of its own deployment.
Network-sited and metro (5)
| System | Who | Evidence | What it is | Status |
|---|---|---|---|---|
| Telco edge, hyperscaler A | self-published | Compute inside one operator's network, reachable only by that operator's mobile subscribersSource [29] in TR-2026-33. | 33 zone entries across 7 carriers, 20 of them one carrier's US zones; no capacity, utilisation or customer count published | |
| Telco edge, hyperscaler B | self-published | Same class, single carrier partnerSource [31] in TR-2026-33. | 3 live sites; documentation archived, last content update August 2024, no formal retirement notice located | |
| Telco edge, operator consortium platform | reported | Operator-founded platform intended to span member networksSource [32] in TR-2026-33. | Founded 2018, sold April 2022 after its website and social presence were taken down; code open-sourced | |
| Carrier-neutral metro | self-published | Metro facility reachable over any networkSource [30] in TR-2026-33. | More than 30 metros available, 7 announced; no capacity or utilisation published | |
| Micro and modular, metro operator | self-published | Metro micro data centres across US marketsSource [34] in TR-2026-33. | 8 markets live of 37 listed; 26 described as customer ready within 90 days |
Micro and modular (1)
| System | Who | Evidence | What it is | Status |
|---|---|---|---|---|
| Micro and modular, tower-sited | reported | Six-rack containerised module, 48 kW IT capacitySource [33] in TR-2026-33. | About 3 sites deployed against 400 planned; liquidation after about $11M raised |
On-device (2)
| System | Who | Evidence | What it is | Status |
|---|---|---|---|---|
| On-device, current physical-AI module | self-published | Compute carried by the machineSource [58] in TR-2026-33. | 2,070 FP4 TFLOPS sparse (1,035 dense) in 40-130 W; no price and no joules per policy inference published | |
| On-device, lower-power module | self-published | Same class at a smaller envelopeSource [59] in TR-2026-33. | Configurable 7 W and 15 W module power modes; vendor envelope, no measured workload figure |
Colocation at generation (4)
| System | Who | Evidence | What it is | Status |
|---|---|---|---|---|
| Colocation at generation, behind the meter | verified | Load inside the plant fenceSource [43] in TR-2026-33. | 300 to 480 MW expansion rejected by FERC 2-1, November 2024, on inadequate justification of non-standard provisions | |
| Colocation at generation, restructured | reported | Same adjacency, different regulatory structureSource [44] in TR-2026-33. | Restructured June 2025 as a front-of-the-meter power purchase agreement of up to 1,920 MW to 2042 | |
| Colocation at generation, tariff status | verified | Rules for co-located load in the largest US RTOSource [45] in TR-2026-33. | FERC found the tariff unjust and unreasonable, December 2025, directing three service options; rates left to a paper hearing | |
| Colocation at generation, wellhead | reported | Modular units on otherwise flared gasSource [46] in TR-2026-33. | About 425 modular units and 270 MW sold March 2025; operator redirected to one 1.2 GW campus |
Heat-reuse siting (3)
| System | Who | Evidence | What it is | Status |
|---|---|---|---|---|
| Heat-reuse siting, district heating market | self-published | Facility sited to sell heat into a networkSource [51] in TR-2026-33. | More than 30 data centres supplying about 3.5% of one capital city's heat after roughly a decade, against a 10% target | |
| Heat-reuse siting, regulatory driver | verified | Siting made partly a heat-offtake decision by statuteSource [49] in TR-2026-33. | Germany: PUE at or below 1.2 and reuse of 10% (2026), 15% (2027), 20% (2028), above 300 kW | |
| Heat-reuse siting, building-integrated | modelled | Single rack inside an occupied buildingSource [56] in TR-2026-33. | Simulated 12 kW rack, 35-45 degrees C air outlet, below district heating supply temperature; no kWh, CO2 or reuse factor reported |
Other topologies (1)
| System | Who | Evidence | What it is | Status |
|---|---|---|---|---|
| Orbital | self-published | Compute in dawn-dusk sun-synchronous orbitSource [57] in TR-2026-33. | Feasibility study; two prototype satellites planned by early 2027; cost parity conditional on a launch price that does not exist |
Withdrawn programmes (2)
| System | Who | Evidence | What it is | Status |
|---|---|---|---|---|
| Withdrawn: short-range vehicular radio estate | verified | Distributed roadside radios under a designated national standardSource [82] in TR-2026-33. | Regulator found it had not been meaningfully deployed, reallocated the spectrum, set a sunset date; causes stated as spectrum use, technology transition and deployment economics | |
| Withdrawn: federal connected-vehicle pilots | verified | Grid- and pole-powered roadside units at three sitesSource [83] in TR-2026-33. | $42 million across three sites from 2015; documented difficulties were procurement, interoperability and installation |
What a harvested aperture can compute.
The most distributed end of the range is bounded by physics rather than by economics, and the bound is computable from measured quantities.
At a 100 cm² pavement-embedded aperture, measured trafficked-surface photovoltaic yields give 8.61 to 106.2 mW. Against measured per-inference energy from MLPerf Tiny, that sustains tinyML-class inference continuously, at 39 to 478 inferences per second after a harsh system derate and 13 to 160 at an assumed winter floor. It does not sustain continuous operation of a 7 to 15 W embedded inference module, which it admits only in bursts against a buffer at a duty fraction between roughly one part in a thousand and one part in seventy. Every step of that derivation is shown in the paper with its inputs and their grades, and seven of the inputs are the author's own assumptions, each labelled so a reader can replace it with a single multiplication.