Over the past year, the artificial intelligence sector has become more divided.
On one hand, the cost of using large model technology has fallen sharply. As China’s preference for domestic computing architecture has strengthened and algorithmic efficiency has improved, token prices have dropped. The race to build ever larger models has also entered a stage of diminishing marginal returns.
On the other side, after three years of rapid technological development, a basic question remains unresolved: As AI becomes more widely used, how should we measure the amount of work it actually helps complete?
The confusion stems partly from a mismatch in metrics. In the mobile internet era, many companies focused on DAU, or daily active users, because it reflected the capture and redistribution of attention.
In the agent era, the relationship between humans and AI requires a different measure. One term is appearing more often: “collaboration.” A user who spends an hour listening to an AI chatbot may create less value than an agent that works independently for five minutes and delivers a complex business outcome.
At Baidu Create 2026, the company introduced “daily active agents,” or DAA, as a metric for the agentic AI era. “To measure the health of a platform or ecosystem, we should pay closer attention to DAA: the number of [AI] agents actively working and delivering results,” said Robin Li, co-founder and CEO of Baidu.
Why DAU is losing relevance
Under the legacy internet paradigm, there was a mature formula for growth. By 2026, that formula is becoming less effective.
For years, DAU was treated as the lifeline of internet products. In the agent era, that traffic-based logic is weakening.
Anthropic and OpenAI offer one example. Although Claude’s user base is far smaller than ChatGPT’s, Anthropic’s commercial growth has accelerated. At the start of 2026, its run rate revenue had reached about USD 14 billion. In the past, few people believed software users would readily pay USD 200 per month. Claude helped change that assumption.
In the agent era, the ability to complete high-value tasks is starting to matter more than user scale alone. In the consumer internet era, free user acquisition often created network effects. In the AI era, a larger user base can also mean higher inference costs. Without real task delivery, DAU can shift from an asset to a liability.
Tokens, meanwhile, have become a central metric in the AI sector. Model capabilities, inference costs, training scale, and context length are all measured in relation to tokens.
But tokens may only be a transitional metric. A token is the basic unit by which large models process information, similar to a unit of electricity consumption. In the early development of AI, it has reference value, but consumption does not equal output. More code does not guarantee better software. More token consumption does not necessarily mean a task has been completed.
At the conference, Li contended that tokens represent cost, not value, and measure input rather than output. In that framing, token consumption is useful for understanding cost, but less useful for measuring whether AI has produced valuable work. That is why DAA may offer a more useful yardstick.
DAA measures how many agents complete end-to-end task loops in real-world scenarios each day. The underlying shift is from interaction to delivery, and from process consumption to completed results.
Chatbots helped AI solve the problem of information access. In the agent era, users expect AI to complete tasks directly. Li shared this observation at the event:
“Chatbots and general-purpose agents will become two types of entry points. ChatGPT-style chatbots are the first generation of entry point. They mainly solve the problem of information access. The second-generation entry point is general-purpose agents. They solve the problem of task completion. The value ceiling of general-purpose agents is higher than that of chatbots.”
DAA is anchored to that second stage: general-purpose agents completing tasks.
At its core, Baidu’s proposal of DAA is an attempt to define a productivity metric for the AI era. It also points to a restructuring of production relationships, positioning task completion as a more accurate test of AI implementation.
For enterprises, this marks a shift from simple cost reduction to workflow optimization. Organizational structures are changing, and more companies may begin to operate as agent-enabled companies. DAA gives enterprises a way to measure the incremental value created by digital labor. For example, Qingdao Port’s automated terminal used Baidu’s Famou to improve the efficiency of its A-TOS terminal operating system by 10.21%. That is the kind of measurable operational value DAA is meant to capture.
Another Baidu example involves Miaoda, its coding agent. An eight-year-old primary school student used Miaoda to generate a campus umbrella-sharing mini program. Traditional development tools tend to focus on how much code was written. Miaoda, viewed through a DAA framework, emphasizes whether the task was delivered.
Technological progress rarely moves in a straight line. The companies that endure through cycles are often not just the fastest movers, but those willing to pursue unpopular bets before they become consensus.
Baidu’s AI development offers one example of that pattern. As early as 2010, when China’s mobile internet sector was entering a period of rapid growth, the industry was focused on traffic expansion and platform competition. Baidu was already one of China’s largest search engine companies and a major online traffic gateway. Rather than relying only on that advantage, it chose a heavier and slower path, establishing a natural language processing department and beginning systematic investment in AI.
In 2013, Baidu established one of China’s earliest research institutions dedicated to systematic investment in deep learning, with Li serving as dean. What then looked like an unusually strong technical conviction later became part of Baidu’s AI foundation. In 2017, Baidu launched the Apollo autonomous driving platform and the DuerOS conversational AI system, signaling its intent to turn AI into next-generation infrastructure.
Over the following four years, Baidu entered the formation stage of its full-stack AI architecture. Before large models became mainstream, the company had begun building capabilities across chips, cloud services, and models. From the release of the first Ernie model in 2019 to the mass production of its Kunlunxin 2 chip in 2021, Baidu gradually developed AI infrastructure spanning chips, frameworks, models, and cloud services.
This preparation reflected Li’s early view that AI would move beyond experimentation and into large-scale application.
When the large model boom accelerated globally in 2023 and competition centered on model parameters, inference capabilities, and token consumption, Baidu began emphasizing an application-driven path. After releasing Ernie Bot, Li publicly said the value of models lies in applications. At the time, this was not the dominant view, as much of the AI sector remained focused on model capabilities.
By 2024, Li went further, arguing that agents would become the most important application form in the next stage. These judgments looked early at the time. Over a longer timeline, as OpenAI, Google, Microsoft, Anthropic, and other global technology companies began investing heavily in agents, they became more aligned with the direction of the sector.
In 2026, as Ernie models move further toward natively unified omnimodal capabilities and search undergoes AI-driven transformation, an important question remains: Is AI truly creating measurable results?
Retooling for the AI agent era
At the conference, executive vice president Shen Dou announced that Baidu AI Cloud will be upgraded into a “new full-stack AI cloud purpose-built for large-scale agent applications.”
Past AI infrastructure mainly served model training and inference. In the agent era, underlying systems are being redesigned around agents. In the future, the entities actively calling APIs will include not only humans, but also agents. Agents will independently choose models, query data, call tools, and coordinate other agents. That means the main goal of AI infrastructure is shifting from supporting token generation to helping AI agents complete tasks more reliably.
Baidu said it has begun restructuring its infrastructure accordingly. One notable concept is “harness engineering.” Instead of focusing only on model capabilities, Baidu is emphasizing how agents can use fewer tokens and fewer conversation turns to complete more complex tasks through long-context management, persistent memory, tool calls, sub-agent coordination, and runtime support.
For AI infrastructure, Baidu is trying to push token efficiency further by upgrading model-as-a-service into a token foundry model. It is also using hierarchical pooling of KV (key value) cache to raise the context reuse rate, while optimizing long-chain agent inference performance and using an omnimodal training framework and reinforcement learning to accelerate agents’ continuous improvement.
At the same time, Baidu is moving deeper into the foundational layer. From Kunlunxin P800 to the Tianchi 256-card supernode and gigawatt-scale AI data centers, Baidu is restructuring its data center architecture to shift focus from power supply and distribution toward networking and inference efficiency.
Baidu appears to view China’s industrial system, broad base of real-world application scenarios, and enterprise knowledge as resources that agents can absorb and amplify. Accordingly, its goal may no longer be simply to build larger models, but to pair models with agents and infrastructure capable of completing useful work at scale.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Xiao Xi for 36Kr.