The Body Electric · Sociogencia Case 07 · The Next Hotspot

2036

Six machines work a shift.

Each one is three computers: the body it carries, the edge site it reaches in milliseconds, and the campus that rebuilds it overnight.

None of this is science fiction. The record underneath is verified.

Only the year has not arrived yet.

The question was never whether the machines could work.

It is whether the ground beneath them can carry the load.

2036. Six machines, three computers each, and the infrastructure underneath them. The record is verified; the year has not arrived.

The cast · six machines

Meet the machines the decade is being built for.

Six characters carry this case: a humanoid, a robotaxi, a quadruped, a warehouse runner, a production arm and a domestic unit. Different bodies, identical anatomy: onboard reflexes that can never wait for a network, an edge site one hop away, and a training campus that turns the fleet’s day into a better model overnight.

Drag to orbitScroll to zoomTap markers for componentsStress test to break the infrastructureKeys 1–6 switch cast

Infrastructure event

Drag to orbit · scroll to zoom · keys 1–6 switch cast

Prototype · procedural holograms, no sourced imageryOpen the lab full screen →

Dossiers · the six on the call sheet

Six files. Three computers each.

The budget · three computers

Every task is priced in milliseconds first.

The price decides which of the three computers is allowed to do the thinking. The budget is not a design preference; it is physics with an invoice attached.

Fig. 05-1 · The three-tier latency budget shared by every machine in the cast

Onboard is sovereign. A humanoid fuses fifteen camera feeds, tactile arrays and inertial sensors into action across forty-one degrees of freedom, with its motor controllers synchronised over TSN or EtherCAT at sub-millisecond determinism.Estimated Its brain is a robot computer in the 2,070 FP4 TFLOPS class, the figure Nvidia’s Jetson Thor shipped at in 2025, drawing 40 to 130 watts.Verified A Level 4 vehicle carries the same logic at road scale: around twenty cameras, five radar and up to three lidar, with 200-plus TFLOPS of automotive compute, because a 100 millisecond delay at speed is the difference between a stop and an impact.Estimated

Source · the 2,070 FP4 TFLOPS figure, 128 GB memory and 40 to 130 watt envelope are NVIDIA’s published Jetson Thor specification, which is why that clause alone is graded Verified. Sub-millisecond determinism is per the IEEE 802.1 TSN and EtherCAT Technology Group specifications. The degrees-of-freedom count, sensor counts and automotive-compute figure are a desk composite across published humanoid and Level 4 platform disclosures, not a single vendor’s number. The stopping-distance point is arithmetic, not a source.

The edge is the social layer. Nothing safety-critical crosses a network, but almost everything economic does: floor traffic control for a thousand warehouse runners, corridor choreography for a robotaxi fleet, vision QC shared across a line of arms. One hop, under fifty milliseconds, or the fleet decays into a crowd of strangers.

The campus is the species memory. A single Level 4 vehicle generates 1.8 to 4.0 terabytes a day; roughly 30 per cent, compressed, is worth uploading, feeding pipelines that run up to three billion simulated miles a month.Estimated The robots’ equivalent is the episode schema and the dataset passport: standardised records of body, scene, action and outcome that let one machine’s Tuesday train every machine’s Wednesday.Estimated

Source · per-vehicle data generation and the share worth uploading are a desk range drawn from published autonomous-fleet disclosures; operators report these inconsistently and the range is wide for that reason. Simulated-mileage figures are as publicly reported by the largest Level 4 operators. The fifty-millisecond edge budget is the desk’s, derived from the choreography tasks listed rather than from a standard. Episode schema and dataset passport describe emerging practice, not a ratified standard, and the framing is ours.

Source · the table is a desk synthesis of the three tiers described above. Budgets are the operating envelopes the tasks imply, not vendor guarantees. The failure column is the desk’s reading and is the argument Act V then tests.

Layer	What lives there	Budget	If it fails
Body network	41 DoF actuator sync over TSN / EtherCAT	Sub-1 ms, deterministic	The machine falls over
Onboard brain	VLA model, perception fusion, reflexes · ~2,070 FP4 TFLOPS class	<10 ms	The machine stops being safe
Edge site	Fleet choreography, QC, map deltas, task queues	<50 ms, one hop	The machine works alone, slowly
Training campus	Nightly fleet learning, simulation, skill releases	Overnight	The machine stops improving

Where the three computers actually sit.

The budget is a map before it is a number. Follow one machine on one shift: the body works a site, the edge site answers in a heartbeat from the nearest metro, and the campus that retrains it overnight can be a continent away. Distance is the whole argument.

One shift · body 0 m → edge metro → training campus

Onboard · 0 m · <10 ms Edge site · metro · <50 ms Training campus · continental · overnight

Six machines, priced in gigawatts.

Scale the cast up from a lab to an economy and they become the numbers this report exists to track: capacity, consumption, interconnection queues and the distance between demand and the grid. Everything in this act carries the Estimated chip: modelled futures, not observed facts.

Global DC IT capacityEstimated

24.4 GW2025

A sixfold rise to 147.1 GW by 2035 · 27% CAGR

US DC demandEstimated

31 GW2024

More than triple to 106 GW · ~8.6% of US generation

Global DC electricityEstimated

447 TWh2025

Up to 1,700 TWh by 2035 · ~4.4% of world demand

Modelled trajectories, not observed facts, and modelled by this desk rather than lifted from a published forecast. Each line takes the cast of six and scales it from a lab to an economy on the assumptions set out above: three computers per machine, a fleet that trains nightly, and a campus tier that has to be built before any of it compounds. Treat the endpoints as the shape of an argument, not a number to plan against.

Capacity, not capability, is the ceiling.

The failure mode of 2036 is not a robot uprising or a robot collapse. It is a brownout of intelligence. Machines built as three computers degrade one tier at a time, in reverse order of latency, and each loss is survivable, which is exactly why it is dangerous: the lights stay on while the learning stops.

Rung one · nominal

The world as designed

Reflexes onboard, choreography at the edge, learning at the campus. Fleets compound: every shift makes the next one better. This is the state every projection in Act IV quietly assumes.

Last night the campus trained on 4,100 of our shifts. This morning I hold boxes I have never held before as if I had. – VESPER

Rung two · edge lost

The fleet becomes strangers

A congested metro, an unhardened MEC node, an interconnection queue that never cleared. Machines stay safe and drop to solo autonomy: the warehouse treats every junction as contested, the robotaxi ballet becomes a queue, QC goes blind. Throughput falls by double digits while every dashboard stays green.

Four hundred of us drop to solo SLAM. The maths is humiliating. – TALLY

Rung three · campus starved

The species stops learning

This is the capacity crunch made visible: PJM projects 31 GW of data centre demand by 2030 against 28.7 GW of new generation, and SMR economics stay two to three times the cost of utility renewables until well into the 2040s. Source · PJM Interconnection load forecast and Base Residual Auction results; Monitoring Analytics, PJM independent market monitor. The same series is set out in full in Equilibrium.Estimated When training capacity is rationed, nothing switches off. Tonight’s four thousand shifts simply teach nobody anything, and the fleet you bought appreciates in rust instead of skill.

I still work. I just stop getting better. – VESPER

In the embodied decade, capability is a solved problem that ships in a box. Capacity is an unsolved problem that ships in a decade.

Case 07 · the thesis in one line

Run the stress test in the lab above and watch it happen per character. BRACKEN’s anomalies queue on its back with no triage to receive them. ODA places parts it can no longer inspect. PLUM’s house goes quiet and forgetful. MERIDIAN drives today on yesterday’s city. None of them stop. All of them shrink.

The box costs 16,000 dollars and falls in price every quarter. The decade is what the other nine chapters of this report are about: siting, interconnection, generation, glass and the politics of who gets the megawatts. The characters can only be as intelligent as the infrastructure they are allowed to reach.

The robot is the demand forecast with a face.

Every machine in this cast is a walking, driving or folding argument for the numbers in Act IV. A humanoid is a 130 watt data centre client that never logs off. A robotaxi is a terabyte-a-day subscriber to the edge tier. A fleet of either is a nightly training job with a gigawatt campus at the far end. When analysts ask where the demand for 147 gigawatts could possibly come from, the honest answer is: from characters like these, multiplied by every warehouse, street and kitchen that hires one.

The verified record says the hiring has started. First commercial deployment in 2024; a hundred thousand totes and sixty-five thousand paid hours by 2026; a public listing priced on hours worked. The three-computer anatomy this piece describes has since stopped being one vendor’s architecture and become the industry’s: Foxconn unveiled industrial humanoids built on Isaac GR00T and Jetson Thor at GTC in March, Nvidia extended humanoid partnerships across the United States, Europe and South Korea in June, adding secure boot and confidential computing to the stack, and LG has a bipedal machine on the same foundation targeted for early 2027. One deployment closes the loop entirely: Omnipresent Robotics began building the first thirty of a planned 143 humanoids to work inside a Michigan AI data centre campus, paired with the Nvidia infrastructure that trains them. The robot and the building it depends on are now the same project.

And the dependency runs the way this piece argued. On Nvidia’s own most recent numbers, edge computing, robotics and automotive together are 7.8 per cent of revenue against 92.1 per cent for the data centre, with robotics not yet broken out at all. The third computer is still overwhelmingly a hyperscale one. The projections say the payroll grows sixfold by 2035. Between those two statements sits every siting decision, interconnection queue and megawatt this report covers.

And the failure mode is already legible. Not collapse: dimming. Fleets that stay safe but stop coordinating; machines that keep working but stop learning; a bifurcated grid where the campus enjoys private nuclear baseload while the city around it queues for capacity. The stress test in this case takes twenty seconds. The real one takes a decade, and it is running now.

The Next Hotspot tracks where the capacity lands. Case 07 is the reminder of who is waiting for it: six characters with call signs, latency budgets and, by 2036, opinions about your cutlery drawer. The room only stays bright if the infrastructure underneath it is built. That is not a robotics story. It never was.

Written by

Adam Roberts

Entelligencia · Sociogencia, Case 07

Case 07 is a fiction built on a verified record. The characters are invented. The deployment figures, the latency budgets and the megawatts are not, and every one of them is graded in the dossier above.

The payroll · Case 07

Six on the books. None of them have asked about the holiday policy.

Staff photographs from the dossiers in Act I. Headcount is illustrative, as is the cutlery drawer. Everything with a megawatt attached to it is not.

The Next Hotspot · dropping

One case down. The full map is next.

The Body Electric is a Sociogencia case inside The Next Hotspot, an interactive read on where the world actually builds AI infrastructure. The full edition is live, and the remaining drops are dated below.

Live now · SociogenciaOriginal analysis is dropping already. Read the studies.Read the studies →

Sociogencia · Case 08The Compute Vote

Sociogencia · Case 09The First Data Centre Election