When the Grid Fails: What San Francisco’s Blackout Revealed About Autonomous Vehicle AI

A citywide outage becomes an AI systems audit

San Francisco’s December 20, 2025 power outage was initially characterized as a conventional utility failure, with approximately 130,000 Pacific Gas and Electric customers losing electricity and city officials urging residents to limit travel as traffic signals across multiple neighborhoods went dark. Beyond its immediate impact on households and transit, however, the outage quickly emerged as a real-world stress test for autonomous mobility systems operating at urban scale. By the following morning, PG&E reported that power had been restored to roughly 110,000 customers, while about 21,000 remained without service as repair work continued. During the disruption, Waymo, which operates one of the most prominent commercial autonomous ride-hailing fleets in San Francisco, suspended service citywide. Video footage and eyewitness accounts showed driverless vehicles halted in intersections with hazard lights activated, and there were additional reports of passengers temporarily stranded inside vehicles as systems entered safety protocols.

The most important analytical point is not that autonomous vehicles paused, but why. In the popular imagination, “self-driving” implies independence from the surrounding city. In practice, autonomy is best understood as an engineered relationship between onboard perception and decision-making, and an external environment that remains partially structured by powered infrastructure, communications networks, mapping systems, and governance protocols. When electricity fails at scale, the autonomy claim is tested not at the level of sensors, but at the level of system coupling.

Autonomy is conditional, not absolute

Autonomous driving stacks are designed to perceive lane markings, vehicles, pedestrians, and roadway geometry through cameras, radar, and in many deployments, LiDAR. They fuse this perception with prediction and planning models that attempt to infer what other road users will do next, then select a safe path. The San Francisco outage exposed a critical boundary: these models operate most reliably when cities provide standardized signals that compress uncertainty, especially traffic lights that impose a shared, machine-legible rule set across thousands of daily interactions.

When traffic signals failed across significant parts of the city, intersections shifted from structured coordination to negotiated coordination. Human drivers rely on eye contact, subtle motion cues, and local driving norms to resolve ambiguity. Autonomous systems can approximate some of this through intent prediction, but the edge cases multiply quickly when whole neighborhoods simultaneously degrade into non-standard intersections. At that point, the dominant rational engineering choice is conservative behavior: stop, wait, and escalate to a fail-safe or remote support channel.

This is not a narrow Waymo story. It is a category insight: large-scale autonomy in dense cities remains dependent on infrastructure that reduces ambiguity. Removing that infrastructure increases uncertainty faster than model confidence can compensate, and at fleet scale, those conservative decisions become visible as service-wide disruption.

Vehicle Systems Exposed During the San Francisco Outage

Fail-safe behavior is safe, but it is not neutral for cities

Safety-first immobilization is often the correct choice for an individual vehicle. At the urban systems level, however, immobilization has externalities. A stopped vehicle can obstruct lanes, complicate emergency response routing, and contribute to congestion at the exact moment when traffic management capacity is already impaired. San Francisco’s emergency management guidance during the outage emphasized reduced travel and four-way-stop behavior at dark intersections, effectively shifting the city into a manual coordination mode. Autonomous vehicles that cannot confidently participate in that manual mode can become friction points.

The distinction that matters for policymakers is the difference between safe driving and safe system behavior. If a fleet defaults to stopping when infrastructure fails, the outcome may still be locally safe, but systemically costly. The question becomes whether approvals and operating permits should account for networked urban resilience requirements, not only collision avoidance metrics.

In other infrastructure-reliant sectors, resilience is governed explicitly. Connectivity disruptions, for example, are widely recognized as systemic risk events. Institute of Internet Economics research has argued that outages impose measurable macroeconomic costs and propagate unevenly across sectors depending on digital dependence. Autonomous mobility, as a hybrid of AI and critical transportation service, belongs in that same risk taxonomy.

Autonomous Vehicle Outage Readiness

The hidden dependency: connectivity, fleet orchestration, and remote support

Autonomous vehicles are frequently described as edge AI, and many real-time driving decisions do occur onboard. Yet commercial robotaxi systems also depend on centralized orchestration: dispatching, routing optimization, dynamic geofencing, incident triage, and remote assistance. Power failures stress these dependencies unevenly. Traffic lights go dark immediately, while cellular networks may degrade gradually due to backup power constraints and localized failures. Operationally, this is a worst-case mix: street conditions become less predictable while fleet-level coordination becomes less reliable.

This matters economically because fleet autonomy is not only a robotics problem; it is also an internet economics problem. Commercial robotaxi viability depends on high utilization and consistent service availability. When the system encounters a low-confidence city state, providers may choose to suspend service rather than operate with higher risk and higher support cost. SFGate reported Waymo suspended operations during the outage in part to maintain safety and allow emergency vehicles clear access.

From a business-strategy perspective, these decisions are rational. From a city perspective, they surface a governance question: if autonomous fleets are permitted as a component of urban mobility, what continuity expectations should apply during infrastructure disruptions? Cities may not require fleets to operate through blackouts, but they may reasonably require fleets to clear roadways, avoid blocking critical routes, or maintain degraded operations within defined parameters.

Measurable safety performance does not equal outage readiness

One reason the outage is analytically significant is that it intersects with a separate trend: the autonomous vehicle industry’s growing use of quantified safety comparisons. Waymo publishes comparative crash-rate statistics in incidents per million miles, broken down by severity and city. In its public safety impact reporting, Waymo shows that for serious injury or worse crashes, its rates are lower than benchmark rates across reported categories, and it provides city-specific comparisons including San Francisco.

Separately, peer-reviewed research has examined Waymo rider-only operations at scale. A 2025 paper in Traffic Injury Prevention reports analysis over 56.7 million rider-only miles through the end of January 2025, comparing crash rates to human benchmarks and finding statistically significant reductions for certain crash categories. These are meaningful indicators of performance under normal operating conditions.

But outage readiness is a different dimension of safety and reliability. Traditional crash metrics measure whether the system avoids collisions. Outage readiness asks whether the system can function when the environment’s coordination signals fail, and whether it fails in a way that is compatible with emergency response and city continuity. A fleet can score well on collision metrics and still be operationally fragile under infrastructure stress.

The infrastructure baseline is not theoretical: reliability metrics quantify exposure

The outage also re-centers a practical point: grid reliability is measurable, and those measurements define how frequently autonomy systems will face degraded urban states. PG&E publishes annual reliability indices commonly used across the power sector. For 2024, PG&E reported a SAIDI of about 276.4 minutes per customer, a SAIFI of about 1.832 interruptions per customer, and a CAIDI of about 150.9 minutes average restoration time per sustained outage. These are system-level averages, not San Francisco-only values, but they indicate that sustained outages are not rare events and that restoration times can be material.

For autonomous mobility operators, these metrics describe a non-trivial operational envelope: outages and partial outages will occur, and cities will periodically shift into degraded coordination. If robotaxi services scale into essential mobility, then infrastructure volatility becomes a cost driver, shaping staffing for remote support, operational playbooks, service-level guarantees, and insurance and liability risk.

The strategic implication is that autonomy is constrained not only by AI model capability but by the reliability of the city’s physical layer. That physical layer is not under the operator’s control, but it defines the operator’s service promise.

What regulation may look like after a blackout stress test

Events like San Francisco’s blackout tend to accelerate a particular regulatory evolution: from evaluating the vehicle to evaluating the system. Early autonomous vehicle oversight emphasized the driving task itself: perception accuracy, collision avoidance, and rule compliance. The next phase is likely to focus more on resilience planning, including infrastructure failures, emergency response coordination, and service continuity obligations.

A plausible regulatory direction is to require documented degraded mode behavior that is explicitly city-compatible. This could include rules for clearing intersections, routing away from critical corridors during signal failures, and minimum capabilities for self-extraction to safe pull-over zones when conditions degrade. Regulators may also push for tighter integration with municipal emergency operations centers, ensuring fleets can be instructed quickly to pause, reroute, or clear priority routes.

These requirements are not merely bureaucratic. They reflect a shift in how cities value autonomy: not as novelty transportation, but as infrastructure-adjacent service. When the city’s infrastructure fails, the autonomous fleet’s behavior becomes part of the outage’s total cost.

Reliability, not novelty, is the real commercialization threshold

The commercialization narrative for self-driving cars often centers on technical milestones: more miles, more cities, fewer disengagements. The San Francisco outage suggests a parallel commercialization threshold: the ability to operate reliably under infrastructure stress, or at minimum, to fail gracefully in a way that does not burden the city.

This reframes the competitive landscape. Differentiation may come less from marginal perception gains and more from systems engineering for resilience: stronger offline intersection negotiation, clearer interaction protocols with human drivers, improved onboard autonomy when connectivity is degraded, and city-level coordination agreements. In internet economics terms, the binding constraint is shifting from algorithmic intelligence to infrastructure dependence, and from novelty to robustness.

If autonomous mobility is to become a durable component of urban transportation, it will be judged not only by how it performs on clear days with powered signals, but by how it behaves on the days when the city is dark.

Key Takeaways

• San Francisco’s December 2025 blackout demonstrated that self-driving remains operationally dependent on powered city infrastructure, especially traffic signaling.
• Safety-first immobilization is rational for individual vehicles, but at fleet scale it can create urban externalities during emergencies.
• Quantified crash-rate improvements and peer-reviewed safety studies do not directly measure outage readiness or city-level continuity performance.
• Grid reliability metrics indicate sustained outages are not exceptional events, meaning degraded-mode operation should be treated as a core design and governance requirement.
• The next regulatory phase is likely to evaluate autonomous mobility as a system with resilience obligations, not only as a vehicle with safety metrics.

Sources

• Associated Press; Power restored to most in San Francisco after massive outage; – Link
• Reuters; Power restored for about 110,000 customers in San Francisco after outage; – Link
• San Francisco Chronicle; PG&E outages in S.F. mostly resolved, but more than 20,000 customers still without power; – Link
• SFGate; Waymo temporarily suspends service in SF amid power outage; – Link
• Waymo; Safety Impact (Crash-rate comparisons by location); – Link
• Traffic Injury Prevention; Comparison of Waymo rider-only crash rates by crash type to human benchmarks at 56.7 million miles (2025); – Link
• PG&E; Electric Reliability Reports (SAIDI, SAIFI, CAIDI summary); – Link
• U.S. Energy Information Administration; Definitions of SAIDI, SAIFI, CAIDI; – Link
• Institute of Internet Economics; Connectivity Is Becoming a Geopolitical Asset, Not Just Infrastructure; – Link