On Tuesday morning, passengers in Wuhan, Beijing, and Shenzhen climbed into driverless taxis expecting to be whisked to their destinations. Instead, they sat in vehicles that wouldn’t move. Some cars stopped mid-route. Others never started at all. For roughly three hours, Baidu’s Apollo Go robotaxi service — the largest commercial autonomous vehicle operation in the world — was paralyzed by a system-wide failure that stranded riders, blocked traffic lanes, and reignited hard questions about the infrastructure underpinning China’s ambitious push into driverless transportation.
The outage, first reported by TechCrunch, began around 8:15 a.m. local time and affected an estimated 1,200 vehicles across multiple Chinese cities simultaneously. Baidu confirmed the incident in a terse statement posted to Weibo, attributing the disruption to a “cloud service anomaly” that severed communication between its centralized dispatch platform and individual vehicles. The company said service was fully restored by 11:30 a.m. No injuries were reported.
But the damage — reputational, regulatory, and strategic — may take far longer to repair.
A Fleet Without a Brain
Baidu’s Apollo Go robotaxis don’t operate in isolation. Each vehicle is connected to a cloud-based command layer that handles route optimization, fleet coordination, real-time traffic data processing, and remote human oversight. When that layer goes dark, the vehicles are designed to execute a “minimal risk condition” — essentially pulling over safely and stopping. That’s exactly what happened Tuesday morning, according to multiple social media posts from passengers and bystanders that circulated on Weibo and Douyin before Baidu’s official acknowledgment.
The problem is what “pulling over safely” looks like at scale. Videos shared on Chinese social media showed Apollo Go vehicles stopped in active traffic lanes, at intersections, and in one case blocking the entrance to a hospital in Wuhan’s Hanyang district. Traffic police in at least two cities had to manually direct vehicles around the frozen robotaxis. One Weibo user described the scene as “a parking lot of ghost cars.”
Baidu has not disclosed the root cause of the cloud service anomaly. The company’s statement said only that a “technical issue in our backend dispatch system” was identified and resolved, and that it is conducting an internal review. Industry analysts, however, are already speculating about whether the failure originated in Baidu’s own infrastructure or in third-party cloud services. Baidu operates its own cloud computing division, Baidu Cloud, which serves as the backbone for Apollo Go’s operations. A failure at that level would suggest a single point of failure in a system that carries paying passengers.
And that’s precisely what concerns regulators.
China’s Ministry of Industry and Information Technology issued a brief statement Tuesday afternoon noting that it was “paying close attention” to the incident and had requested a detailed report from Baidu. The ministry oversees the regulatory framework for intelligent connected vehicles in China, and it has been progressively expanding the zones where companies like Baidu, Pony.ai, and WeRide can operate without safety drivers behind the wheel. That expansion has been predicated on trust — trust that the technology is mature enough, that the companies are transparent enough, and that the failure modes are understood well enough.
Tuesday’s events tested all three assumptions at once.
The timing is particularly awkward. Just last month, Baidu CEO Robin Li told attendees at the company’s annual technology conference that Apollo Go had completed over 8 million cumulative rides and was on track to achieve profitability in select cities by the end of 2026. Li described the service as entering a phase of “scaled commercial maturity.” Investors responded positively. Baidu’s Hong Kong-listed shares climbed 4% in the days following the speech.
Those shares dropped 6.2% on Tuesday.
The broader autonomous vehicle sector in China felt the tremor. Pony.ai, which went public on Nasdaq in late 2025, saw its stock dip 3.1%. WeRide fell 2.8%. Neither company experienced outages of their own, but the market’s reaction reflected a collective reassessment of the sector’s operational risk profile. When the leading operator stumbles, everyone gets scrutinized.
The Centralization Problem
What makes this incident particularly instructive is what it reveals about the architecture of large-scale autonomous vehicle deployments. Most robotaxi operators, including Waymo in the United States, rely on a combination of onboard computing and cloud-connected services. The vehicle’s onboard systems handle perception, planning, and immediate driving decisions. The cloud handles fleet management, route assignment, regulatory geofencing, and — critically — remote human intervention when the vehicle encounters a situation it can’t resolve autonomously.
In theory, a cloud outage shouldn’t cause vehicles to stop driving. The onboard systems should be capable of completing a trip or at least navigating to a safe location independently. But in practice, the degree of cloud dependency varies significantly between operators. Baidu’s system, according to technical documentation the company has published in academic papers and at industry conferences, relies heavily on its cloud layer for real-time HD map updates, V2X (vehicle-to-everything) data integration, and dispatch commands. When that layer becomes unreachable, the vehicles’ default behavior is conservative: stop.
This is a defensible safety design in isolation. No one wants an autonomous vehicle improvising without its full information stack. But at fleet scale, a conservative default becomes a systemic hazard. Twelve hundred vehicles simultaneously executing a minimal risk condition don’t just inconvenience passengers. They disrupt urban traffic patterns, block emergency access, and erode public confidence in ways that take months to rebuild.
Some researchers have argued for more decentralized architectures that give individual vehicles greater autonomous decision-making capability during cloud outages. “The question isn’t whether cloud connectivity will fail — it will,” said Dr. Zhang Wei, a professor of intelligent transportation systems at Tsinghua University, in an interview with Chinese state broadcaster CCTV on Tuesday evening. “The question is what the vehicle does in the 30 minutes or three hours before it’s restored. Right now, the answer for most operators is: nothing. That needs to change.”
Baidu isn’t alone in facing this challenge. Waymo experienced a smaller-scale incident in San Francisco in 2023 when a group of its vehicles clustered in a residential neighborhood and became temporarily unresponsive, though that event was attributed to a mapping error rather than a full cloud outage. Cruise, General Motors’ autonomous vehicle unit that suspended operations in late 2023 after a pedestrian dragging incident, had also faced questions about the reliability of its remote assistance systems.
But the scale of Tuesday’s Baidu outage is unprecedented in the commercial robotaxi industry. No operator has ever had this many vehicles go offline simultaneously in active service areas. It is, in a sense, the first true stress test of what happens when a city-scale autonomous fleet encounters a backend catastrophe.
The results are not encouraging.
For Chinese policymakers, the incident creates a difficult balancing act. Beijing has made autonomous driving a strategic priority, embedding it in the country’s broader industrial policy goals alongside AI, semiconductors, and electric vehicles. Cities like Wuhan have been especially aggressive in welcoming robotaxis, with Wuhan granting Baidu permission to operate over 400 fully driverless vehicles — no safety driver, no remote copilot physically in the car. The city has positioned itself as a global showcase for autonomous mobility.
That showcase had a very public malfunction.
Local officials in Wuhan have not commented publicly beyond acknowledging the traffic disruptions and saying they were “resolved with the cooperation of the operating company.” But behind the scenes, according to a report from Chinese business publication Caixin, regulators are discussing whether to mandate minimum onboard autonomous capability standards — essentially requiring that vehicles be able to complete trips or return to designated parking areas even during extended cloud outages.
Such a requirement would be technically demanding and expensive. It would force operators to invest in more powerful onboard computing hardware, more comprehensive local map caching, and more sophisticated fallback driving algorithms. It would also shift the balance of intelligence from the cloud to the edge — from centralized efficiency to distributed resilience. For Baidu, which has built its entire Apollo platform around the premise of cloud-centric intelligence, this would represent a significant architectural pivot.
Whether the company is willing or able to make that pivot quickly enough to satisfy regulators and reassure the public remains an open question.
What Comes Next
Baidu held an internal all-hands meeting Tuesday evening, according to two people familiar with the matter who spoke to TechCrunch on condition of anonymity. Robin Li reportedly addressed staff directly, calling the outage “unacceptable” and ordering a full technical audit of Apollo Go’s cloud infrastructure with results due within two weeks. The company is also expected to brief regulators in Beijing, Wuhan, Shenzhen, and other operating cities in the coming days.
For the autonomous vehicle industry globally, the Baidu incident offers a case study in the operational risks that emerge only at scale. Pilot programs with 10 or 50 vehicles can absorb a cloud outage quietly. A fleet of 1,200 in dense urban environments cannot. The gap between demonstration and deployment is not just about technology readiness. It’s about infrastructure resilience, failure mode engineering, and the unglamorous work of building systems that degrade gracefully rather than catastrophically.
Competitors are watching closely. Waymo, which operates roughly 700 vehicles across San Francisco, Los Angeles, Phoenix, and Austin, declined to comment on the Baidu incident specifically but pointed to its own redundancy protocols in a statement to reporters. Pony.ai, which operates in both China and the U.S., said in a Weibo post that its systems “maintained normal operations throughout the day” — a subtle but unmistakable bit of competitive positioning.
The Chinese public, meanwhile, has reacted with a mixture of alarm and dark humor. The hashtag #BaiduRobotaxiFreeze trended on Weibo for most of Tuesday, accumulating over 280 million views. Some users posted memes of the frozen vehicles with captions like “even the robots are tired of commuting.” Others were less amused. “I was in one of these cars on the way to a job interview,” one user wrote. “It just stopped on the highway ramp. I had to call a regular taxi. I was 40 minutes late. This isn’t funny.”
No, it isn’t. And for Baidu, the cost of this failure extends well beyond one bad morning. The company’s credibility as a safe, reliable operator of fully driverless vehicles — the credibility it needs to expand into new cities, attract more riders, and eventually turn Apollo Go into a profitable business — took a serious hit. Rebuilding that credibility will require more than a two-week audit. It will require demonstrating, concretely and publicly, that the system architecture has been hardened against the kind of single-point failure that brought an entire fleet to its knees.
The robots stopped. The real question is what Baidu builds to make sure they don’t stop again.
WebProNews is an iEntry Publication