When the lights went out, the Waymos weren't confused. They were queuing.
A substation fire blacked out a third of San Francisco and left driverless cars stopped at dead intersections. Waymo's own explanation is more interesting than the pictures: the cars knew exactly what to do, and asked permission anyway.
On Saturday 21 December 2025, a fire at a PG&E substation cut power to around 120,000 customers in San Francisco — roughly a third of the city. Traffic signals went dark. Muni was disrupted. Mayor Daniel Lurie told residents to stay off the roads unless they had to travel. And a number of Waymo robotaxis stopped at intersections and stayed there, with traffic, buses and in at least one case a fire truck behind them.
Waymo suspended its San Francisco service on the Saturday evening and resumed late on Sunday afternoon. The pictures did the rounds, the jokes wrote themselves, and the story was filed under “driverless cars can't cope”. That filing is wrong, and the real explanation is considerably more uncomfortable.
The cars knew the rule
Waymo published its account three days later. Its software already treats a dead traffic signal the way the law does: as a four-way stop. That is the correct behaviour, and it is what the cars were doing.
What the cars then did was ask for confirmation. Waymo's vehicles can request a check from its fleet response team — the humans who advise a car that has encountered something ambiguous. In an ordinary outage, a handful of cars ask, humans answer, traffic moves. On 21 December, a third of the city's signals went dark at once, and the requests arrived as what Waymo called a “concentrated spike”.
The fix Waymo shipped is a fleet-wide update giving the software “specific power outage context, allowing it to navigate more decisively” — which is to say: stop asking, you already know the answer.
Dark signals Waymo says its cars traversed successfully that Saturday. The city counted nearly 1,600 stoppages of two minutes or more. Both numbers are true.
The bottleneck was a queue, not a brain
This is the part worth sitting with, because it is not a Waymo problem. Every driverless fleet on the road today has a human layer behind it: fleet response, remote assistance, teleguidance, whatever the company calls it. That layer is a shared, finite resource, and it is staffed against expected demand.
Expected demand assumes that the events which make a car ask for help are independent — an unusual construction zone here, a double-parked van there, scattered across a city and a day. A blackout is the opposite. It is a single event that makes thousands of intersections ambiguous simultaneously, and therefore makes a large fraction of the fleet raise its hand at the same moment.
No amount of per-car intelligence fixes that, because the constraint is not in the car. You cannot compute your way out of a staffing ratio. And the industry's favourite safety metric — miles per intervention — conceals the problem completely, because it averages interventions over time as though they arrive politely spaced out.
The number nobody publishes is the one this outage asked for: what fraction of the fleet can need a human at the same time before the queue collapses? Until an operator answers that, “we have remote assistance” is a description of a feature, not of a capacity.
Two true numbers
Waymo says its cars successfully traversed more than 7,000 dark signals that day, which is a genuinely impressive number and exactly the kind of context that usually goes missing.
At a San Francisco Land Use and Transportation Committee hearing on 3 March 2026, the city produced a different one: nearly 1,600 vehicle stoppages lasting two minutes or more, and more than 60 vehicles that had to be removed manually. The fire department's deputy chief of operations, Pat Rabbitt, described a firefighter having to leave an incident to deal with a Waymo that had encroached on the scene. Supervisor Bilal Mahmood reached for Cinderella: the carriages, he said, “can turn into pumpkins at the drop of a hat”.
Both figures are accurate and they measure different things. 7,000 is a competence statistic. 1,600 is a congestion statistic. What makes the second one matter more than it looks is that none of it shows up in a collision record. Nobody was hurt. No car crashed. A stopped vehicle is, in the narrow safety sense, the safest thing a confused robot can be — and it is also what put a fire truck behind sixty of them.
What we take from it
Waymo behaved well after the fact: it suspended service rather than pressing on, published a technical explanation within three days rather than a statement about its commitment to safety, and shipped a change. That is a better response than most of this industry manages, and we have said so about Waymo before.
But the event exposed something the fleet numbers hide. Waymo now runs more than 4,000 vehicles and over half a million paid rides a week. The human layer behind them does not scale linearly with that, because it is not supposed to — the entire economic argument for autonomy is that it does not. Correlated demand is where that argument gets tested, and a blackout is the gentlest possible test of it. An earthquake would not be gentle, and San Francisco is the wrong city to find that out in.
The right lesson is not that driverless cars failed. It is that the part of a driverless fleet that most resembles a call centre is load-bearing, invisible, and unmeasured — and that when it is overwhelmed, the cars do not crash. They stop, politely and in unison, exactly where you least want them.
