Why is Elon Musk "schadenfreude" about a power outage in San Francisco?
In the night of December 20, 2025, San Francisco suddenly "went dark".A fire broke out at a substation facility of Pacific Gas and Electric Company (PG&E) in the city, causing a large-scale power outage across the entire city. Official data shows that at the peak, about 130,000 households and businesses lost power, accounting for one-third of its total users. Traffic lights, street lamps, and public transportation systems failed simultaneously.
The power outage lasted for several hours. Amid the chaos, Waymo, which has been operating over 1,000 vehicles regularly in San Francisco, caused severe traffic jams. Multiple social media videos show more than a dozen Waymo vehicles stagnating at the same time in the same neighborhood, completely blocking some intersections. Towing companies had to recover multiple Robotaxis late at night.
Waymo stated in a statement that its vehicles are designed to treat malfunctioning signal lights as four-way stop signs. "The scale of this failure led to vehicles staying longer than usual to confirm the conditions at affected intersections, which exacerbated traffic friction during peak congestion periods."
Tesla CEO Elon Musk retweeted multiple tweets indicating that Tesla's Robotaxis were not affected by the power outage in San Francisco. The two Robotaxi technical routes represented by Waymo and Tesla were once again put to comparison.
Beyond the debate over the pros and cons of autonomous driving technology itself, what is easier to overlook is the real starting point of this turmoil—the power outage. It not only determines whether traffic lights work but also involves the present of electric vehicles and the future of AI.
I. The Resurgence of Technical Controversy: Tesla Gains Allies
Musk's statement has sparked another round of discussions on Robotaxi technology.
After the Waymo incident, Musk posted a video on platform X showing a Tesla Model Y driving with FSD (Supervised Full Self-Driving) activated on streets where street lamps and signal lights were all out. However, quite a few voices argue that such a comparison is not rigorous, because Tesla's Robotaxis were still equipped with human safety drivers at that time and had not yet officially launched large-scale commercial operations.
Musk also retweeted a post that read: "All of Waymo's Robotaxi fleets equipped with numerous lidars are out of service. Meanwhile, all Tesla Robotaxis continue to operate... Through simple software updates, Tesla's vision-based artificial intelligence can run anywhere, not just in pre-approved or carefully controlled environments like lidar-dependent systems."
In fact, prior to the power outage in San Francisco, Musk had publicly expressed similar views multiple times. On December 10, he bluntly stated: "Waymo never really had a chance to compete with Tesla. In hindsight, this is obvious."
Waymo's technical route is recognized in the industry as a "high-safety redundancy" solution—relying on multi-sensor fusion of lidar, millimeter-wave radar, and cameras; using high-precision maps and clear Operational Design Domains (ODD); and providing fallback assistance through remote technicians. The advantage of this approach is that multiple layers of redundancy improve safety and it is easier to pass regulatory approval, but compared with the pure vision solution, it has higher costs, more restrictions on operating areas, and greater difficulty in scaling up.
During this power outage incident, Waymo's system design itself does not rely on real-time communication or external power supply. Multiple industry insiders judge that the more likely scenario is: against the backdrop of a large-scale power outage, the response capacity of the remote support system was limited, and vehicles were forced to extend their stay in the "uncertainty confirmation" stage, ultimately forming local or chain congestion.
In contrast, Tesla emphasizes a pure vision solution: weak or no map dependence, learning human driving behavior through neural networks. This does not mean that its route is safer, but it is more flexible in emergency scenarios, less dependent on deployment areas, and closer to the cost structure required for large-scale commercialization.
Chinese Robotaxi companies have long adopted a compromise route. Represented by Apollo Go, Pony.ai, and WeRide, their technical systems retain both multi-sensor fusion and high-precision maps while continuously strengthening end-to-end capabilities to achieve human-like driving. However, this high-cost compromise is being broken.
For example, Apollo Go, which has the world's highest cumulative number of service orders, according to multiple media reports, Baidu founder Robin Li clearly stated at an internal meeting in the second quarter: based on cost and scalability considerations, Baidu's Robotaxis will fully switch from a multi-sensor fusion solution to a pure vision technical route.
XPeng Motors, which has begun to enter the Robotaxi field, has also chosen a direction highly consistent with Tesla. Yuan Tingting, senior director of XPeng Motors' autonomous driving products, shared her experience in North America on social platforms: when a section of road was temporarily closed, Tesla FSD could independently re-plan the route, while Waymo had to wait for the police to remove the roadblocks before passing. She said: "XPeng's thinking is highly consistent with Tesla's in judging the technical route of Robotaxis. We believe that the second-generation VLA (Vision Language Action) will drive the system from L2 to L4."
II. Robotaxis, Electric Vehicles, and Even AI Face Infrastructure Challenges
On the overseas forum Reddit, one comment read: "The infrastructure is not ready for these things."
Another was more straightforward: "We can't even keep the cars themselves running, let alone autonomous driving cars."
The direct reason Waymo stopped was the power outage. After the substation fire, the power company's slow response prolonged the traffic paralysis. In first-tier and core second-tier cities in China, large-scale, long-duration power outages have become very rare. According to operational data from the State Grid Corporation of China, there have been almost no city-level power outages affecting millions of people in China in recent years. But in the United States and Europe, sudden power outages are becoming normalized.
Data from the U.S. Energy Information Administration (EIA) shows that in 2024, the average power outage duration for U.S. users reached 662.6 minutes (about 11 hours), a year-on-year increase of 80.7%; Texas, where data centers are concentrated, even reached 1614 minutes, a year-on-year increase of more than 170%. In 2025, Spain and Portugal also experienced large-scale power outages affecting more than 10 million people.
Behind the power outages is a growing power gap. The North American Electric Reliability Corporation (NERC) warned in a report that most parts of the United States face the risk of power shortages from 2025 to 2029. This risk is highly correlated with the development of AI. The Institute for Energy Research (IER) estimates that a single training of OpenAI's "Orion" model consumes electricity equivalent to that used by about 1 million U.S. households in a whole year; and the computing power centers that OpenAI plans to deploy by 2033 will add a load exceeding a quarter of the current maximum national power load in the United States.
NVIDIA CEO Jensen Huang has also vividly described this gap: "By 2027, the GPU clusters supplied by NVIDIA alone will consume 150-200GW of electricity globally, equivalent to the total electricity consumption of 1.5-2 Frances."
The problem is that the speed of grid expansion is far from keeping up with the growth in demand. Grid construction has high costs and long payback periods, leading to insufficient investment motivation from governments and private power companies and low construction efficiency. For example, since the major power outage in Texas in 2021, the revision of power equipment reliability standards has still not made substantial progress.
Voltage instability further amplifies the problem. Data shows that the U.S. power grid is seriously aging: 70% of transformers have exceeded their 25-year design life, the average service life of transmission lines is 40 years, the grid load reserve rate is only 20%, and its shock resistance capacity is obviously insufficient. The instantaneous load brought by AI computing is more likely to trigger voltage fluctuations. This directly limits the installation of charging piles, which in turn affects the purchasing decisions of electric vehicles.
Many American netizens mentioned on Reddit that Tesla charging piles, including installation costs, are about $2,000, but not all distribution boxes can withstand the corresponding current. Some apartments have therefore banned the installation of charging piles, leaving pure electric vehicle owners to rely on public charging stations that are often queued.
Insufficient power supply capacity has also driven up electricity prices. Data from the U.S. Energy Information Administration shows that electricity prices in many parts of the United States have soared in 2025, reaching 20 cents (about 1.4 RMB) per kilowatt-hour in some areas. During the major European power outage in April this year, electricity prices in Germany once exceeded 3.3 RMB per kilowatt-hour. In contrast, residential electricity prices in China generally range from 0.52 to 0.62 RMB per kilowatt-hour.
These structural differences determine that overseas markets are difficult to replicate the speed of electrification in China. Data from the China Passenger Car Association (CPCA) shows that the global penetration rate of new energy vehicles in the fourth quarter of 2025 was 25.2%, including 49% in China, 30% in Germany, and only 7% in the United States; Japan, where public electricity prices are higher than oil prices, has a penetration rate of only 1.7%.
Unstable infrastructure not only slows down the electrification process but also continuously amplifies public doubts about intelligence. A Reddit user wrote: "Robotaxis and AI are packaged futures. They do not solve immediate problems but make them more complicated."
Network carrying capacity has also become a bottleneck. Residents in many places have reported that during crowded periods with network congestion, Waymo has repeatedly failed to continue driving due to insufficient network connectivity. This not only affects Robotaxis but also restricts vehicle-road coordination, intelligent signal systems, and even the entire "smart city" framework.
What Waymo's shutdown in San Francisco really exposed is not a technical flaw of a single company, but a crueler fact: in an environment with unstable infrastructure, the ceiling of technology is not determined by technology itself.
页:
[1]