Produced by | He Xi

Layout by | Ye Yuan

On May 13, Baidu hosted the Create 2026 conference. During the event, Li Yanhong introduced a new value term: DAA—Daily Active Agents.

'Tokens only represent costs, not benefits,' he said.

Li's statement punctured a layer of window dressing. Over the past two years, the entire industry has been caught up in a computing power race: whoever buys more chips or burns more tokens is considered the winner. However, ByteDance's net profit decline, Baidu's AI investment eroding profits, and Alibaba Cloud's slowing growth reveal that the computing power 'arms race' is straining financial results.

How can the current predicament (I'll keep this term as is for now, assuming it refers to a specific industry challenge) be escaped? Li Yanhong offered an alternative metric: instead of comparing who burns the most money, compare who generates the highest number of 'daily active intelligent agents.'

Shen Dou took up this metric. He announced that Baidu Intelligent Cloud would fully upgrade to a new full-stack AI cloud for intelligent agents, transforming from a computing power provider to a builder of 'intelligent agent factories.'

DAA is not just a concept but Baidu's redefinition of the competitive rules for the second half of the AI era.

Next, the author will discuss why DAA is a new metric for value in the AI era and how Baidu is using it to rewrite the rules of the game, starting with this 'standard-setting.'

Why does Li Yanhong say tokens don't represent benefits? What is the relationship between DAA and DAU? How does Shen Dou's new full-stack AI cloud transform intelligent agents from 'toys' into 'productive forces?'

These three questions correspond to the three core dimensions of the DAA era. First, why DAA is needed—computing power inflation forces a shift in value metrics. Second, how DAA is implemented—supported by the dual Infrastructures of the new full-stack AI cloud. Third, what DAA means for the industry—the battle for definition rights and Baidu's ambitions.

01

Computing Power Inflation Forces a Shift: Why Tokens Are No Longer 'Hard Currency'

First, let's examine the industry background behind DAA—computing power consumption is rising while profits are shrinking.

The token mindset is a legacy of the computing power era: it only measures input, not output. From 2024 to 2025, nearly all AI companies flaunted 'average daily token usage' as a core metric. Whoever's model generated content faster, had more users, or engaged in more frequent dialogues was considered the winner. However, by 2026, this logic began to crumble.

The reason is that the marginal cost of tokens is not zero. For traditional internet products, a tenfold increase in users might only double server costs. For large-scale model products, a tenfold increase in users leads to a nearly proportional increase in computing power costs. ByteDance's Doubao saw its average daily token usage soar from 120 billion in May 2024 to 120 trillion in March 2026—a 1,000-fold increase in two years. During the same period, ByteDance's capital expenditures on AI computing power exceeded 150 billion yuan, while net profits plummeted.

This is not just a ByteDance problem. Although Baidu Intelligent Cloud's AI revenue growth rate increased to 200%, it still couldn't cover the massive R&D and computing power investments. Alibaba Cloud also faces scrutiny over its AI computing power ROI. The entire industry is grappling with the same question: Where has all the computing power money gone?

Li Yanhong's insight lies here. He points out that tokens are merely a cost center, not a profit center. Users chatting or generating content with tokens do not directly create commercial value. What truly creates value is how many tasks intelligent agents complete and how many results they deliver—for example, an agent automatically processing 1,000 invoices for a business or completing 500 quality inspections on a factory production line.

Daily Active Agents (DAA) is the metric for measuring such 'output.' Analogous to DAU in the mobile internet era, it directly corresponds to commercial monetization capabilities—advertisers pay for daily active users, and capital values them accordingly. Li predicts that global DAA could exceed 10 billion in the future, meaning intelligent agents will penetrate every industrial sector, just as apps did in their time.

This prediction is not mere fantasy. Baidu Intelligent Cloud has already accumulated real-world data on intelligent agent implementation across multiple industries: 20 million intelligent driving vehicles, 1,000 AI hardware manufacturers, 800 financial institutions, and 80% of central enterprises. In these scenarios, intelligent agents are 'working' every day. DAA is not a concept but a reality in the making.

The token mindset is a legacy of the computing power era; the DAA mindset is the ticket to the intelligent agent era.

02

New Full-Stack AI Cloud: The 'Power Plant' for DAA

After discussing the necessity of shifting metrics, let's examine how DAA is implemented—Baidu's new full-stack AI cloud serves as the 'power plant' for DAA.

Shen Dou announced that Baidu Intelligent Cloud would fully upgrade to a new full-stack AI cloud for intelligent agents, establishing dual Infrastructures: Agent Infra and AI Infra. The former reduces the 'unit intelligence cost' of agents, while the latter provides cost-effective computing power. Their synergy makes DAA scaling possible.

Agent Infra: Making Each Token 'Smarter'

The core of Agent Infra is the 'Token Factory' and 'Harness Engineering.'

Token Factory is a complete reconstruction of the traditional MaaS (Model as a Service). MaaS simply packages models into APIs and charges based on token usage. However, Token Factory adopts an Agent-first philosophy, reducing redundant computations from the ground up. For example, when an agent processes a continuous dialogue, it often needs to repeatedly understand the same context. Token Factory improves inference generation speed by 25% compared to the market average through intelligent caching and context reuse. This means the same token can produce more 'intelligence.'

Harness Engineering is a set of capability modules for complex agent tasks. It covers key areas such as long-context management, persistent memory, tool invocation, sub-agent scheduling, evaluation feedback, and runtime environments. These modules are not simply stacked but deeply synergized. When handling office tasks using tools like browsers and Office, Harness Engineering achieves a 95% success rate. More importantly, thanks to better context management and task orchestration, it reduces dialogue rounds and token consumption for the same tasks—23% fewer tokens compared to solutions like OpenClaw.

For developers, this means two things: first, the barrier to developing intelligent agents is significantly lowered, as they no longer need to build complex memory and tool invocation systems from scratch. Second, operational costs are substantially reduced, allowing the same budget to support more daily active agents.

AI Infra: Making Computing Power 'Cheaper'

At the AI Infra level, Baidu Intelligent Cloud has achieved multiple breakthroughs in model training and inference, cluster construction, and data center development.

First is Kunlun Core, Baidu's self-developed AI chip. The P800 model has completed large-scale validation, with multiple 10,000-card clusters delivered since 2025. A 10,000-card cluster consists of 10,000 AI acceleration cards for training ultra-large models. On Kunlun Core's fully domestic cluster, the Wenxin large model version 5.1 completed training with an effective training rate of 97% and a linear scalability of over 85% for 10,000-card clusters. This means that when scaling from 1,000 to 10,000 cards, computational efficiency loss is controlled within 15%, meeting the requirements for large-scale training of cutting-edge models in terms of computational precision, operator stability, framework adaptation, and long-term operation.

The upcoming Tianchi 256-card super node, launching in June, further improves inference efficiency. Compared to the previous generation, it offers 25% higher throughput and 50% better inference efficiency. This super node has been adapted for mainstream models like Wenxin, DeepSeek, GLM, and MiniMax, optimizing end-to-end latency by 50%.

At the data center level, Baidu Intelligent Cloud unveiled an upgrade plan for gigawatt-scale AIDCs (Artificial Intelligence Data Centers). Through a 'network-centric layout' design philosophy, the distance between computing nodes and network nodes is minimized to maximize computational efficiency. The large-scale implementation of a wind-liquid compatible architecture shortens the overall data center construction cycle by about 30%. This means Baidu can deliver more computing power resources in less time.

The logic of dual Infra synergy is simple: do more with fewer tokens and run more agents at lower costs. This is the mathematical foundation for DAA scaling. Analogous to the electric power era, Token Factory improves generator efficiency, Harness Engineering optimizes the transmission network, and Kunlun Core and AIDCs reduce coal costs. Together, they drive down the 'electricity price' of intelligent agents to the threshold of mass adoption.

Baidu is not just upgrading its cloud but building a 'power plant' for intelligent agents—tokens are the coal, and DAA is the electricity.

03

The Battle for Definition Rights: How DAA Rewrites AI Competition Rules

After discussing the foundation supporting DAA, let's explore what this 'standard-setting' means for the industry—the battle for definition rights.

Definition rights represent the highest dimension of competition. In the mobile internet era, Google used 'search volume' and Facebook used 'daily active users' to define platform value. Whoever defines the KPIs controls the flow of capital and resources. In the PC era, Microsoft used 'install base' to define software industry value standards; in the mobile era, Apple used 'App Store downloads' to define ecosystem prosperity.

Now, Baidu is attempting to make DAA the 'DAU' of the AI era.

Why does Baidu dare to set this standard? Because it has already accumulated a verifiable DAA foundation across multiple industries.

Automotive Industry: Baidu Intelligent Cloud has supported the delivery of 20 million L2-level intelligent driving vehicles, with customers including OEMs like Geely, Li Auto, Changan, and NIO, as well as chip and solution providers like Horizon and Neolix. The intelligent computing center jointly built with Changan Automobile, based on Baidu's Baige·AI Heterogeneous Computing Platform, continuously optimizes support for Changan's R&D in large models, intelligent connectivity, and autonomous driving. One hundred percent of mainstream Chinese automakers have chosen Baidu Intelligent Cloud, maintaining its leadership in China's autonomous driving R&D solutions market.

AI Hardware and Embodied Intelligence: Baidu Intelligent Cloud serves over 1,000 AI hardware manufacturers, covering smartphones, AI glasses, robotic vacuums, and smart home appliances. In the embodied intelligence AI cloud market, Baidu's market share exceeds the combined total of the second and third players. This means a vast number of physical-world intelligent agents—from robotic vacuums to industrial robotic arms—are leveraging Baidu's cloud intelligence.

Financial Industry: Kunlun Core helps China Merchants Bank establish a domestic computing power base for its AI applications and supports Shanghai Pudong Development Bank in fine-tuning financial analysis models to improve corporate loan due diligence efficiency. Baidu Intelligent Cloud serves over 800 financial institutions, covering scenarios like investment research, risk control, office operations, and due diligence.

Central Enterprises and Global Expansion: Baidu Intelligent Cloud serves over 80% of central enterprises, including State Grid. In the realm of Chinese brands going global, Baidu unveiled its enterprise-grade intelligent marketing solution, Hogee, and 'Yijian Claw,' which creates exclusive (I'll keep this term as is, assuming it refers to a customized solution) visual intelligent agents for businesses. These are helping Yiwu's 'front-shop, back-factory' merchants transform into 'AI factories' and 'AI store managers.' Through Yijian visual intelligent agents, massive rules can be automatically configured to each camera, enabling natural language rule identification and process disposition with a single sentence. Hogee assists merchants in completing the entire marketing chain—from sales guidance and data analysis to inventory management and promotional recommendations—through built-in marketing skills.

These are not PowerPoint slides but real-world validations of DAA. Every intelligent driving vehicle, AI hardware device, financial institution agent, and Yiwu merchant's digital store manager contributes to DAA.

Li Yanhong said, 'The protagonist of AI is not models but applications.' Behind this statement lies Baidu's transition from a 'model company' to an 'intelligent agent platform company.' DAA is the value metric for this new identity.

Of course, DAA is currently a Baidu-created metric. To become an industry standard, Baidu needs to open its ecosystem, allowing third-party intelligent agents to run on its platform with verifiable data. If DAA remains just an internal Baidu KPI, it loses its metric significance. Shen Dou mentioned in his speech that Baidu Intelligent Cloud supports calls to domestic mainstream models like Wenxin, DeepSeek, GLM, and MiniMax. This open stance is the foundation for DAA to become an industry consensus.

Nevertheless, Baidu has already seized the initiative in this battle for definition rights. If Alibaba Cloud, Tencent Cloud, and others continue to focus on 'computing power scale,' they may face a dimensional reduction attack from Baidu's 'daily active intelligent agents.' When investors start asking, 'How many daily active intelligent agents does your platform have?' instead of 'How many parameters does your model have?' the rules of the game will fundamentally change.

The battle over metrics is essentially a battle for definition rights. Whoever sets the standard rules the world.

Conclusion

Li Yanhong sets the standard; Shen Dou lays the foundation.

DAA is not a word game but Baidu's rewrite of the competitive rules for the second half of the AI era. While the industry is still comparing computing power scale and model parameters, Baidu has shifted its focus to the 'daily active' intelligent agents—how many agents are working and delivering results for humans.

This represents a shift from 'cost thinking' to 'value thinking.' It acknowledges the reality of computing power inflation but refuses to be trapped in the dead end of the computing power race. It redefines the meaning of 'input' through 'output' and redefines 'platform value' through 'daily active intelligent agents.'

Whether DAA becomes a new metric for the AI era depends on whether Baidu can support an open intelligent agent ecosystem with its new full-stack AI cloud. Shen Dou's Token Factory and Harness Engineering, Kunlun Core's Tianchi super nodes and gigawatt-scale AIDCs, and the industry penetration of 20 million intelligent driving vehicles and 1,000 AI hardware manufacturers—all these are paving the way for DAA scaling.

This 'standard-setting' has just begun. But at least, Baidu has taken the first step, and it's a step that hits the industry's most painful point.

Now, it's time to see how other players respond.