The Autonomy Build: What Self-Driving Asks of the Grid, the Compute Stack, and the City
Autonomous driving has quietly crossed from perpetual demo into early commercial reality — hundreds of thousands of paid driverless rides a week, and freight moving on public highways with no one in the cab. The interesting question for this series is not who wins the robotaxi race. It is what autonomy physically builds and permanently re-prices: a compute-and-energy footprint the size of a data-center industry, a capital bet on whether the car or the road gets smart, and a reshaping of the city and the corridor.
The road that thinks for the traveller must first be taught the road; and the teaching is a labour far greater than the journey, and it is never wholly done.
Autonomy Is a Build, Not Just a Model
For a decade, self-driving was a technology perpetually five years away. In 2026 that framing has broken. Waymo is running on the order of 500,000 paid, fully driverless rides a week across roughly eleven US cities with a fleet near 3,000 vehicles, and is targeting a million rides a week by year end; Aurora is hauling commercial freight on Sun Belt highways with an empty driver’s seat. The demonstration phase is over. What matters now, for an infrastructure analyst, is that autonomy is not a piece of software that ships — it is a physical build that draws power, fills data centers, and re-prices the road it runs on.
This note treats autonomy the way this series treats any build: by asking what it durably requires and what it permanently re-prices, and by separating the durable build from the passing hype. The robotaxi leaderboard is a rent — it will churn, cities will be won and lost, and today’s leader may not be tomorrow’s. The compute-and-energy footprint, the charging and mapping infrastructure, and the reshaped city and corridor are the build. Own the build; rent the rent.
The Compute-and-Energy Build
Every autonomous vehicle is, in the industry’s own phrase, a supercomputer rolling down the highway. A single car generates somewhere between four and forty terabytes of sensor data a day — cameras, radar, lidar, sonar — which must be transmitted, stored, and fed back into training. The training itself runs in data centers. Multiply by a fleet, then by a global rollout that Uber’s own management sizes at hundreds of thousands to a few million robotaxis by 2035, and autonomy becomes one of the largest new sources of data-center, storage and wireless demand on the horizon — the same demand curve this series tracks in its compute and net-zero work, arriving now from a second direction.
The energy shows up on board as well as off. The sensor-and-compute stack on a current robotaxi draws on the order of a kilowatt continuously, and it is not free: one production robotaxi platform saw its driving range fall from about 303 miles to 168 — a 46% penalty — purely to power its autonomy. Newer designs are pushing that down, but the direction is set: a robotaxi is an electric vehicle carrying a data center, and it needs both the charging infrastructure of a fleet and the grid capacity behind it.
An MIT study modelled that a billion autonomous vehicles, each driving an hour a day with an 840-watt computer, would consume enough energy to rival the emissions of every data center on Earth as of 2023. We are decades and orders of magnitude away — but the vector is clear: autonomy is a new, mobile, and largely unbudgeted claim on electricity and compute.
The Capital Question: Does the Car Get Smart, or the Road?
Underneath the compute build sits the decision that actually allocates the capital, and the two market leaders have taken opposite sides of it. Waymo runs an infrastructure-heavy model: lidar, radar and cameras in redundant layers, centimetre-accurate high-definition maps, and a slow, city-by-city process of mapping and geofencing before a single paid ride. It is expensive per vehicle and per city, but it has produced the only fleet with an independently verified safety record — a July 2026 IIHS study found Waymo vehicles involved in roughly 68% fewer police-reportable crashes per mile than human drivers across three of the four cities studied. Tesla runs the opposite bet: cameras only, no lidar, no pre-built maps, wagering that a general vision model trained on millions of consumer cars will generalise to any road cheaply. Its robotaxi service remains confined to Austin, small, and still ironing out basic reliability.
This is the build’s central capital question, and it is not settled. The heavy model front-loads infrastructure — maps, sensors, geofences — and buys proven safety and public trust; the light model front-loads nothing and bets on software generalisation to scale cheaply, at the cost of maturity so far. The tell is that the infrastructure-heavy approach is the one with paying customers at scale and an independent safety benchmark, while the infrastructure-light approach is the one still promising. For now, the road has to be taught the road — and the teaching is capital. That may change; it has not yet.
The cautionary data point sits between the two. Cruise pursued a Waymo-style build, spent about $10 billion, and was shut down by its parent in December 2024 after a 2023 pedestrian-dragging incident destroyed its regulatory standing. The lesson is not that heavy infrastructure fails; it is that in autonomy the safety-and-trust ledger is as much a part of the build as the sensors — and it can be lost in a single event. That ledger belongs to this series’ failure-mode work as much as to its build work.
What It Re-Prices: The City and the Corridor
Autonomy’s second-order build lands on two pieces of infrastructure this series already tracks. In the city, a fleet that never parks in prime real estate, that concentrates demand at the kerb rather than in garages, and that competes with and complements transit, re-prices parking, kerb space and land use — the Cities thread’s territory. The geofenced, HD-mapped service zone — Waymo’s Bay Area zone alone exceeds 260 square miles — is a new kind of mapped, instrumented urban layer that has to be built and maintained road by road.
On the corridor, the economics are arguably cleaner than in the city, and moving faster. Aurora now runs driverless heavy trucks across roughly ten Sun Belt lanes, with more than 250,000 driverless miles and, it reports, no system-attributed collisions; it has validated a 1,000-mile Fort Worth-to-Phoenix run that no single human driver may legally complete in one shift. That last point is the whole freight case: an autonomous truck is not bound by hours-of-service limits, so it re-prices the long-haul corridor around continuous running, hub-to-hub transfer, and the mapping, rest-stop and charging infrastructure the lanes require. The autonomous long-haul market, on industry estimates, grows from a few billion dollars today toward the tens of billions within a decade. Freight is the quieter, and possibly the first durable, autonomy build.
The Positioning Read: Own the Build, Rent the Rent
The investable distinction is the same one this series draws for every chokepoint and every crisis. The durable build is the compute, energy and mapping layer autonomy requires no matter who wins: data-center and grid demand, fleet charging and depots, sensor and high-definition-map supply chains, and the freight lanes that become franchises once proven. The rent is the robotaxi leaderboard itself — the city-by-city land grab, the single-operator winner bets, the hype cycle that will churn. Autonomy will make some operators and unmake others; the build beneath them earns regardless.
Compute, energy & charging
Data-center and grid demand, fleet-scale charging and depots. Autonomy is a mobile claim on electricity and compute that grows whoever wins the app.
Maps, sensors & freight lanes
The infrastructure-heavy layer — HD maps, sensor supply chains, and proven driverless corridors — is franchise-like: costly to build, durable once built, hard to displace.
The robotaxi leaderboard
City-by-city share and single-operator winner bets are the churny, mean-reverting layer. Size them as a rent, not a franchise; today’s leader is not guaranteed tomorrow’s.
Safety & the geofence
Service is still geofenced, mapped and limited, and the safety-and-trust ledger can be lost in one event, as Cruise showed. The build is real but bounded — and the bound is regulatory and social, not technical alone.
This build note extends the compute-and-energy demand tracked in the series’ compute-anchor and net-zero-arithmetic work, arriving now from the road. Its city re-pricing belongs to the Cities thread and its freight re-pricing to Corridors; the capital-allocation and cost reads connect to The Cost-of-Capital Gap and Energy Security. The autonomy safety-and-trust ledger — the Cruise shutdown, the Tesla probes, the first pedestrian fatalities — is a natural subject for the Fault Lines failure-mode thread.
Autonomy has crossed from demo to commerce, and the right question is no longer who wins the robotaxi race but what the race is quietly building underneath itself. A supercomputer on every axle re-prices the grid and fills the data centers; the capital fight over whether the car or the road gets smart front-loads either software or infrastructure; and the reshaped kerb and corridor land on infrastructure this series already tracks. Own the durable build — compute, energy, charging, maps, freight lanes. Rent the leaderboard. And watch the safety ledger, because in autonomy it is part of the build, and it can be lost in a day.
It was never the wheels that were the marvel, but the unseen mind that fed the eyes; and a mind must be housed, and cooled, and fed with power, somewhere far out of sight of the road it drives.
Leave a Reply