Inside Baidu’s AI Stress Test
Baidu is turning AI into agents, chips, cloud, and robotaxis, but scale is now testing the whole system
Baidu is one of the most complete and contradictory examples of AI commercialization in China: it is standing on the front line of both the digital world and the physical world, trying to turn agents into an entry point for office work and knowledge production while proving that autonomous driving can withstand the real pressure of city-scale operation.
This essay is part of What China’s Industry Media Is Really Talking About.
Executive Summary
Baidu is becoming one of China’s most complete AI commercialization test cases, because it spans agents, cloud infrastructure, AI chips, autonomous driving, and real-world urban operations.
Baidu’s AI-to-consumer strategy is shifting toward workflow-based productivity tools through Wenku, Netdisk, GenFlow 4.0, Office Agent, and OpenClaw, rather than relying only on chatbot-style interaction.
GenFlow 4.0’s reported 100 million monthly active users and 200 million monthly task deliveries show that Baidu is trying to turn AI from a demonstration product into a recurring task-execution layer.
Kunlunxin’s listing process strengthens Baidu’s full-stack AI narrative by connecting applications, models, cloud infrastructure, and AI chips into a more vertically integrated system.
Apollo Go gives Baidu a rare position in physical-world AI, but the Wuhan vehicle stoppage exposed the operational, regulatory, and public-trust challenges of city-scale Robotaxi deployment.
Baidu’s central challenge is to prove that AI can become a reliable production system across both digital workflows and urban infrastructure, not only a collection of strong technical assets.
Baidu Is Back at the Center of China’s AI Narrative, but the Test Has Changed
Baidu has returned to the center of China’s AI discussion over the past month, not because it has suddenly become the loudest large-model company, but because its AI system is now facing two very different stress tests at the same time. One is taking place in the digital world, where agents, cloud, chips, search, Wenku, and Netdisk must prove that they can be organized into a sustainable productivity system. The other is taking place on real roads, where Robotaxi operations must prove that they can handle communications, dispatching, failures, rescue procedures, regulation, and public trust at the city level. Baidu’s special position is that it is not only building models or applications. It is one of the few Chinese technology companies pushing AI from the screen into the road, and from software into urban infrastructure.
Over the past few years, Baidu has often occupied an awkward position in China’s AI debate. It invested early in AI, autonomous driving, large models, search, maps, Wenku, Netdisk, intelligent cloud, Kunlunxin, and Apollo. Yet capital markets and public discussion have often focused more on ByteDance’s product speed, Alibaba’s cloud and e-commerce scenarios, Tencent’s social and gaming ecosystem, DeepSeek’s model shock, and the rise of new large-model companies. Baidu’s AI story looked complete, but it also looked scattered.
Chinese industry-media coverage over the past month has pulled this scattered story back together. Create 2026, Baidu’s AI Developer Conference, was held in Beijing from May 13 to May 14. The theme was “All Things as One,” and the conference was expected to showcase Baidu’s latest progress in its full-stack AI layout across chips, cloud, models, agents, industrial applications, and AI ecosystems. This year’s Create conference also combined the Baidu AI Developer Conference and the Cloud Intelligence Conference, with Baidu AI Cloud expected to release new products in both AI Infra and Agent Infra.
This shows that Baidu is actively moving its AI narrative from standalone model capability toward full-stack system capability. “Chip-cloud-model-agent” is not a lightweight slogan. It represents Baidu’s attempt to reorganize AI chips, intelligent cloud, large models, agents, and application entrances into one commercialization chain. This is also where Baidu differs most from many other Chinese AI companies. Its pressure does not simply come from whether its model is strong enough. It comes from whether these long-accumulated assets can be turned into an AI system that users, enterprises, and cities can all verify.
Agents Are Becoming Baidu’s Real AI-to-Consumer Entry Point
Baidu’s recent AI-to-consumer progress is centered on workflow-oriented products such as Baidu Wenku, Baidu Netdisk, and GenFlow, rather than on a standalone chatbot. Wenku and Netdisk naturally contain large amounts of documents, materials, courseware, images, videos, contracts, résumés, learning content, and personal knowledge assets. This makes them much closer to real task scenarios than a generic chat interface. A user who opens a chatbot may simply ask a question. A user who opens Netdisk or Wenku usually already has a task involving files, materials, writing, organization, archiving, search, editing, or office work.
Chinese media reports on GenFlow 4.0 show that Baidu is making the connection among Wenku, Netdisk, Office Agent, OpenClaw, and private-domain data more concrete. Baidu Wenku and Baidu Netdisk jointly released the general-purpose agent GenFlow 4.0, upgraded Office Agent, and emphasized that both individuals and teams can deploy OpenClaw inside Netdisk, turning Wenku and Netdisk into an AI workspace. The reported numbers are important: GenFlow 4.0 already has 100 million monthly active users and delivers 200 million tasks per month.
These numbers matter. What Chinese AI applications now lack is not another product demonstration on stage, but measurable usage frequency, task delivery, and paid conversion. Baidu Wenku and Baidu Netdisk already have stable user scenarios and accumulated content. When AI floats above a search box or a chat box, it can easily remain a one-off Q&A tool. When AI enters Netdisk, Wenku, and office documents, it has a chance to become a document-processing layer, task-execution layer, and personal knowledge-management layer.
Baidu’s real AI-to-consumer battlefield is the continuous workflow of finding, reading, writing, revising, editing, storing, and delivering. This path does not have the viral spread of a social product or the daily stimulation of short video, but it is closer to paid productivity. Baidu’s historical accumulation in search and knowledge content can become a foundational asset in the agent era. The question is whether Baidu can turn these entrances into tools that are smooth enough, frequent enough, and trusted enough, rather than simply adding AI features to old products.
Kunlunxin Pulls Baidu’s AI Story Back to the Compute Layer
If one only looks at Wenku, Netdisk, and agents, Baidu’s AI story can look like an application-layer product upgrade. Kunlunxin’s capital-market move pulls the story back to the compute layer. On May 8, TMTPost reported that Kunlunxin formally began STAR Market listing tutoring on May 7, 2026, with CICC as the tutoring institution. Filing information showed that Kunlunxin was founded in June 2011 with registered capital of about RMB 412 million, and its controlling shareholder is Baidu China, with a 57.67% stake. The report also noted that Kunlunxin had submitted an A1 listing application to the Hong Kong Stock Exchange in confidential form as early as January 2026. The STAR Market tutoring process means the company is pursuing an A+H dual capital-platform layout.
This is important for Baidu. China’s AI competition is no longer only a model competition. It is increasingly a competition in compute organization. Large-model training, inference costs, cloud-service margins, enterprise deployment, and concurrent agent calls all require support from underlying compute and software-hardware adaptation. Alibaba can rely on Alibaba Cloud. Tencent has Hunyuan and its cloud ecosystem. ByteDance has massive traffic and recommendation systems. Huawei has Ascend and a strong government-enterprise ecosystem. Baidu’s differentiated position lies in its early AI chip layout, search and knowledge scenarios, intelligent cloud, and autonomous-driving system.
If Kunlunxin can move beyond dependence on Baidu itself, Baidu’s “chip-cloud-model-agent” narrative will look more like an industrial system. If Kunlunxin remains mainly an internal compute supplement for Baidu, the full-stack AI story will still lack external market validation. This is the core of Baidu’s current AI commercialization pressure. It does not lack assets, historical investment, or technical labels. But each layer of its assets must be reconfirmed by real customers, real revenue, and a real ecosystem.
Baidu AI Cloud Is Carrying the Second Main Line of Enterprise AI Deployment
The Create 2026 agenda also shows that Baidu wants to emphasize not only consumer agents, but also enterprise AI infrastructure. The conference included an opening ceremony, two main forums, and more than 20 specialized forums covering AI Infra, OpenClaw, digital humans, embodied intelligence, AI plus industry, and other directions. The “Technology and Product” main forum hosted by Baidu AI Cloud was expected to release full-stack AI product progress, explain the path of AI value deployment, and help enterprises enter the agent era.
This line is also important. Since 2023, Chinese large-model companies have generally moved from model launches to application deployment, and from financing narratives to revenue pressure. Enterprise customers no longer care only about model benchmarks. They care about whether AI can reduce customer-service costs, improve R&D efficiency, transform marketing workflows, connect internal data, ensure security and compliance, and control inference costs. If Baidu AI Cloud wants to regain presence in this round of competition, it must package large-model capability into a system that enterprises can deploy, developers can call, and industry partners can integrate.
Baidu’s advantage is that it understands traditional enterprise clients and government-enterprise projects better than many new AI companies. It also has years of engineering accumulation in intelligent cloud, maps, search, knowledge graphs, speech, OCR, and autonomous driving. Its weakness is also clear. The enterprise cloud market is intensely competitive. Alibaba Cloud and Huawei Cloud have stronger scale, channels, and ecosystem presence. ByteDance’s Volcano Engine is moving fast in content, recommendation, and model applications. Baidu AI Cloud must clarify its differentiation by proving that it can build stable advantages in Agent Infra, industry data, task orchestration, and AI engineering deployment, rather than simply repeating model-capability language.
Baidu’s enterprise AI opportunity lies in combining large models, agents, knowledge retrieval, industry tools, and low-cost inference into deliverable industry systems. This path is slower and heavier, but it fits Baidu’s long-standing engineering trajectory.
Apollo Go Pushes Baidu AI onto Real Roads
The other line in Baidu’s AI story is more visually powerful: autonomous driving. Compared with Wenku, Netdisk, AI Cloud, and Kunlunxin, Apollo Go pushes AI from the digital world into the physical world. Robotaxi is far more difficult than a normal AI application because it does not face text, images, or office documents. It faces roads, vehicles, pedestrians, traffic police, passengers, weather, communication networks, city traffic rules, and public safety.
Over the past few years, Apollo Go has been one of the most important examples of autonomous-driving commercialization in China. According to recent Chinese and regional media coverage, by February 2026 Apollo Go had expanded to 26 cities, achieved 100% fully driverless operation in its mainland China operating cities, and delivered more than 20 million autonomous ride-hailing orders to the public.
These numbers show that Baidu’s autonomous-driving business has moved beyond the early experimental phase. It is no longer merely running test cars in closed parks or demonstrating a technology route on stage. It has entered multi-city, multi-scenario, long-cycle operation. The true commercial value of Robotaxi also lies here. Once autonomous vehicles can operate reliably, vehicles, algorithms, maps, dispatching systems, operating platforms, and urban traffic networks may form a new layer of mobility infrastructure.
This also means that Apollo Go is no longer facing laboratory risk. It is facing urban-system risk. In single-vehicle testing, a failure can be treated as a technical problem. In city-scale operation, multi-vehicle stoppages, trapped passengers, traffic congestion, and regulatory intervention become public events. The closer autonomous-driving commercialization gets to the real world, the more it must evolve from a technology-company project into an urban-infrastructure system.
The Wuhan Outage Changed the Direction of the Robotaxi Narrative
On the evening of March 31, multiple Apollo Go vehicles in Wuhan reportedly stopped moving, drawing media and regulatory attention. Lianhe Zaobao reported on April 29 that Apollo Go experienced a large-scale vehicle stoppage in Wuhan, with several passengers trapped inside vehicles. Bloomberg, citing people familiar with the matter, reported that after the incident China paused the issuance of new autonomous-driving permits and that regulators asked local governments to conduct comprehensive self-inspections and strengthen safety monitoring to prevent similar incidents.
Fortune China’s May 7 report further stated that on March 31, more than 100 autonomous taxis under Baidu collectively stopped on the streets of Wuhan. Some vehicles were stopped on elevated bridges and elevated roads, and some passengers were trapped for as long as two hours. Several weeks later, according to Bloomberg, Chinese authorities paused the issuance of new autonomous-driving permits.
This incident is a watershed for Baidu and for China’s Robotaxi industry. The metrics the industry used to emphasize were city count, order count, fully driverless ratio, overseas expansion, and declining vehicle cost. After the Wuhan outage, the key questions changed: How should risks be isolated when vehicles experience collective abnormalities? How should vehicles safely degrade when cloud dispatching systems encounter problems? How should passengers escape when communication networks fail? How should malfunctioning vehicles on elevated bridges, expressways, and major roads be handled quickly? How should traffic police, operators, and remote safety staff coordinate? How should regulators set the pace of expansion?
The commercialization threshold for Robotaxi has moved from “Can it drive?” to “Can it stop, rescue, manage, and recover safely?” This sentence may be closer to the reality of the autonomous-driving industry than any technology demonstration. A truly mature autonomous-driving system is not mature simply because it can complete driving tasks most of the time. It is mature because it can keep losses within an acceptable range when abnormal situations occur.
This is also the core challenge for Baidu’s autonomous-driving business. Apollo Go remains one of the leading commercialization examples in China’s Robotaxi industry. But leadership brings more system responsibility, not less risk. Every city-scale failure magnifies public concerns about safety, regulation, and business models, and pushes the industry from speed competition toward reliability competition.
Baidu’s AI Stress Test Is Taking Place in Two Worlds at Once
Baidu’s current AI stress test has a distinctive structure. Wenku, Netdisk, GenFlow, Office Agent, and AI Cloud operate in the digital world. Their core questions are whether users will use them frequently, whether enterprises will pay, whether agents can genuinely improve productivity, whether inference costs are controllable, and whether product experience is stable enough. Apollo Go operates in the physical world. Its core questions are whether vehicles can run continuously, whether cities can absorb large-scale deployment, whether regulators can build trust, and whether the system can handle failure.
These two worlds point to the same issue: the next stage of AI competition is moving from model capability to system-operating capability. Digital agents require system-operating capability because they must connect user data, tool calls, file permissions, enterprise workflows, and cost control. Robotaxi requires even stronger system-operating capability because it must connect vehicles, roads, maps, cloud systems, communication networks, regulation, and safety redundancy.
Baidu’s advantage comes from this complexity. It has search entrances, knowledge content, Netdisk files, map data, AI Cloud, AI chips, and autonomous-driving fleets. Each asset may not be the strongest in isolation, but the combination is rare. Baidu’s problem also comes from this complexity. The more assets it has, the harder the coordination becomes. The longer the chain becomes, the more any weakness in one link can affect the entire narrative. Baidu does not lack AI assets. It needs to prove that these assets can be organized into a stable, profitable, regulatable, and scalable system.
China’s AI Commercialization Is Entering a More Realistic Phase
Baidu’s recent developments also show that China’s AI industry has entered a more realistic phase. In 2023 and 2024, large-model competition revolved mainly around parameters, benchmarks, open source, price wars, and product launches. Since 2025, the industry has cared more about inference costs, agent workflows, enterprise deployment, AI cloud revenue, chip adaptation, and paid applications. In Baidu’s case, the pressure of AI commercialization has moved further into city roads and public safety.
This is why Baidu’s current narrative is more important than a single model-company story. A large-model company can change market expectations with one technical breakthrough. An application company can win traffic with one breakout product. Baidu must answer heavier questions at the same time: Can agents become a productivity entrance for individuals and enterprises? Can Kunlunxin move toward the external market? Can AI Cloud regain industry relevance in the agent era? Can Apollo Go move from demonstration operation to stable city-level service?
Baidu’s AI story is no longer about whether early investment is finally paying off. It is about whether China’s AI can turn technical accumulation into system capability. This is representative of the entire Chinese AI industry. Chinese companies do not lack model launches, product speed, or engineering iteration. The real test is whether those capabilities can be embedded into daily office work, enterprise processes, industrial sites, urban transportation, and public infrastructure.
Baidu stands exactly at this intersection. Its agent business will test whether AI can restructure knowledge production and office workflows. Its AI Cloud will test whether enterprise AI can move from project-based deployment to platformization. Kunlunxin will test whether self-developed AI chips from Chinese internet companies can form a more independent ecosystem. Apollo Go will test whether autonomous driving can move from technical advancement to urban usability.
In the end, Baidu’s value will not be determined by whether it has the most product announcements or by one model ranking. What Baidu really needs to prove is that AI can be organized into a production system that runs reliably. That is the central competition for China’s AI industry in the next stage.
Source note: This essay is based on Chinese and regional media reporting from Xinhua, Sina Finance, 21st Century Business Herald, TMTPost, Lianhe Zaobao, Fortune China, and related industry reports published between April and May 2026. All translations of short quoted phrases are my own.