(12) The Real Moat Is Fleet Operations, Not the Stack

August 6, 2026/5 min read

Depots Beat Demos

Commercial robotaxis now stall on running enough clean, charged, insured, remotely assisted vehicles, dense enough in a city, to fill demand without burning the balance sheet. Waymo already runs paid driverless service across roughly ten to eleven U.S. metros, with about 500,000 paid rides a week by late March 2026 and more than 200 million cumulative autonomous miles. The harder number sits in the fleet file. A June 2026 NHTSA recall filing listed 3,871 vehicles in the regulated fleet count (secondary filing analysis of the primary disclosure). Third-party analysis of that build-out pace suggests the company's target of more than one million weekly paid rides by year-end 2026 may need roughly 7,000 vehicles, while current monthly additions point closer to 6,000. If the target is missed, the story will be manufacturing, depots, and ops. Autonomy demos still dominate the narrative. Residual value sits with whoever keeps cars productive after the safety driver leaves.

The Cost Stack After the Driver

McKinsey's shared-autonomy cost model puts combined operating costs near $8.20 per vehicle-mile for a typical 1,000-vehicle U.S. city fleet, with local fleet operations about half. The figure is a model, not audited Waymo COGS, and it still shows that removing the driver relocates cost into fixed capital, insurance, remote assistance, cleaning, charging, and municipal throughput. BCG's 2026 robotaxi work has made a similar point from another angle: current operating costs remain high in many markets, with full-scale cost compression and multi-city fleet thresholds still years out. Stack quality matters for safety and for how often a car needs help. It does not erase the physical work of turning a vehicle overnight.

Waymo told Congress it runs roughly 70 remote-assistance agents on duty worldwide against a fleet near 3,000 vehicles, about one agent per forty cars on a total-fleet basis. Concurrent active vehicles imply a tighter ratio, closer to the mid-teens in community recalculations. Either framing is still a labor system. Cruise's pre-suspension San Francisco fleet was reported, in investigative coverage, at roughly 1.5 support workers per vehicle, with human help every few miles. GM later exited robotaxi funding after more than $10 billion invested, citing the incremental time and capital required to operate the business. That exit was an ops and capital story as much as a software story.

Empty Miles Are the Tax

California CPUC filings analyzed in Transport Findings cover Waymo's first thousand-odd commercial days: about 13.8 million trips and 86 million vehicle miles. Passenger-onboard share later plateaued near 55 to 57 percent, leaving roughly 43 to 45 percent of miles empty. Independent U.S. studies of human ride-hail often put empty miles in a similar band, so robotaxis are not uniquely wasteful. They still pay the empty-mile tax. Utilization and city density decide whether a mile is a product or a cost center.

Swiss Re's December 2024 study of 25.3 million Waymo autonomous miles found large reductions in property-damage and bodily-injury claim frequency versus human baselines. That is real actuarial evidence that a better stack can lower insurance and incident load. The counterargument follows: if the stack keeps improving, remote-assist rates and downtime fall until ops becomes a thin, commoditized layer. The counter is only half right. Safer driving improves the ops math. It does not remove depots, cleaning cycles, charging queues, municipal permission, or the need to put thousands of cars through a city without stranding demand.

Scale Is a Throughput Problem

Zoox and other entrants are already building dedicated command hubs and contract manufacturing ramps because the scarce input is not another demo video. It is concurrent vehicles that stay available. Waymo's own ride target is a throughput equation: vehicles in service, trips per vehicle per day, and deadhead share. The stack can be good enough for paid driverless service in ten-plus metros and still leave the operator short of a million weekly rides if the fleet cannot grow fast enough.

Markets still rank robotaxi winners by autonomy prestige and city-count scoreboards. The scarcer asset is the operating layer around the car, from remote assist under load and falling deadhead miles to fast depot turns and municipal permission that survives the first ugly incident. Whoever owns that layer decides how fast paid robotaxis scale once the car can already drive itself.

Sources