Baidu World 2026 opened in Beijing on November 12. Founder Robin Li's thesis: "The next act of AI is the intelligent organization." He called it the "new factory floor" of the AI era and laid out three principles for organizational evolution — dissolve job boundaries, capture tacit knowledge, and design closed-loop processes.

Dou Shen, Baidu executive vice president and president of the Baidu AI Cloud Group, picked up those three principles on stage but redirected them: organizational evolution and infrastructure development advance together, and intelligent organizations in the AI era need an intelligent cloud. In his words, the three principles of organizational evolution are also three new standards for cloud services.

One side is management vocabulary, the other is infrastructure vocabulary, and the two have to meet in the middle. Whether they do depends on a more basic question: what is an intelligent cloud actually serving? Dou's answer is "specific tasks," not "IT systems."

From Serving IT Systems to Serving Tasks

The distinction sounds like a matter of phrasing, but it describes two different businesses. Traditional cloud serves a company's IT systems and delivers resources and stability — so many cores, so much storage, so many nines of availability. The acceptance criteria can be written into a contract, and what runs on top is the customer's problem.

Agents executing tasks don't fit that structure. As Baidu puts it, an agent needs computing power, models and a runtime environment, but it also needs to call the tools, data and systems the enterprise already has. In other words, much of what the intelligent cloud has to deliver isn't in its own hands: inventory, approval workflows, historical chat logs — all of it has to be callable before a task counts as done.

Hence one of the heavier lines of the day: the intelligent cloud will penetrate deep into an enterprise's underlying resources, information systems and business systems, letting agents call on organizational capabilities as needed. "Penetrate" is an unusual word in cloud computing — the boundary between cloud and a customer's internal systems has traditionally been drawn sharply, because clear boundaries are what make the product sellable. Talking about penetration moves the delivery boundary out of the server room and into the business process.

Three Principles, Translated Into Three Service Standards

The translation is line by line. Dissolving job boundaries, applied to infrastructure, becomes "task-centric: organizing computing power, models, data and security into unified infrastructure." Those four things traditionally sit under different procurement lines, different budgets and different owners; now they have to be repackaged around the task.

Capturing tacit knowledge, applied to products, becomes "helping enterprises store and manage organizational context, turning individual experience into organizational capability." This is the hardest of the three. Tacit knowledge has never lived in documents; it lives in the judgments people make in specific situations — why this batch of材料 was rejected, why this customer got an exception. Making it callable requires somewhere to store it, a format to store it in, and a way to retrieve and reuse it.

Designing closed-loop processes, applied to service delivery, becomes "serving customers with a long-term mindset, supporting them as they keep closing process loops and evolving." The first two are about product capability; this one is about the shape of the business. Process loops don't close in a single engagement, delivery cycles stretch out, and pricing has to follow. Writing it down as a service standard amounts to admitting this isn't a one-off sale.

Baidu as Its Own Test Bed: Four Multiples in a Year

Behind the pitch are numbers Baidu generated internally. Over the past year, daily active agents grew 6x, completed effective tasks grew 11x, total cloud sandboxes grew 315x, and concurrency grew 60x. All four figures are Baidu's own; none has been independently verified.

Two things stand out when the four are read together. Task volume grew nearly twice as fast as agent count, meaning the average agent is carrying more work — a departure from the early pattern in which one agent did one thing.

The other is that total sandboxes rose 315x while concurrency rose only 60x. The gap can only come from turnover: more sandboxes created and destroyed per unit of concurrency, which means each sandbox lives a shorter life. Tasks are becoming smaller, more numerous and more disposable.

The sandbox is the least conspicuous element of this story and the most revealing. To execute a task, an agent runs code, calls tools and makes mistakes inside an isolated environment, then the environment is destroyed and the next task spins up a fresh one. Sandbox turnover is, in effect, the tempo of "task" as a unit of delivery.

Three Tiers of Agents, One Entry Point

On the product side, agents are organized into three tiers. The general tier is Baidu Dazi, positioned as the single entry point through which an enterprise calls and manages all its agents. The professional tier is a set of scenario-specific agents: Famou, Yijian, Yijing, Shengsuan, Miaoda, Kuku, Comate and Hogee. The custom tier is left to an enterprise's own unique scenarios, built on the Baidu Dazi development platform.

What matters about this structure isn't the number of agents but the entry point. All three tiers funnel into one place, so enterprises don't have to remember eight product names. The more agents there are, the more valuable that entry point becomes, because the cost of choosing is absorbed by the platform.

Below that sit two layers of infrastructure. Qianfan Agent Infra runs the agents: support for million-scale sandbox concurrency, with thousands of sandboxes launching simultaneously in hundreds of milliseconds. It also provides Qianfan MaaS model services and Ciyuansuo, a large-model trading platform matching token supply with token demand.

Baige AI Infra handles training and inference: 2 million sandboxes spun up in minutes, end-to-end reinforcement learning throughput of 52,000 tokens per second, and train-inference divergence held to one part in ten thousand.

That last metric deserves a note. Training and inference use different precisions, parallelism strategies and frameworks, so the same model often behaves differently on each side. When the gap widens, everything looks fine in training and the model gives nonsense answers in production. One part in ten thousand is Baidu's official consistency figure, and consistency is one of the few areas where platforms like this can put a number on the table.

Tianchi 1,024-Card: A Spec Sheet, Not a Shipment

Baidu also unveiled the Tianchi 1,024-card supernode, which will go on sale in 2027. The gap matters: what was announced is a product definition and a spec sheet, not something you can order today.

On paper, it uses optical engine technology for card-to-card optical interconnect, with interconnect bandwidth triple that of the previous generation. A single supernode can support training of 5-trillion-parameter models and scale to 10,000 cards.

Read "on sale in 2027" and "5-trillion-parameter training" together and it's clear this product line isn't aimed at today's inference business. Models at that scale demand cluster-wide scheduling, parallel partitioning and fault recovery; card count and bandwidth are merely the price of admission. Between now and 2027 lie three hurdles — tape-out, a mature software stack, and customer validation. Miss any one of them and the spec sheet stays a spec sheet.

It also explains why a supernode was announced at this particular event. The whole conference argued that the intelligent cloud delivers tasks; the ceiling on those tasks is set by the silicon underneath — how many sandboxes an agent can open at once, how long a context it can run, how many expert modules fit in memory. It all lands on this hardware.

Manufacturing and Hotels First, Three Paths for Partners

Only two industries were named: manufacturing and hotels. No customer names, no scale figures. For a business that has just been reorganized, that's unremarkable — prove it works first, then talk about how much of it is running.

Partners are offered three modes of collaboration. Standardized integration: developers of data, tools and Skills plug their capabilities into Baidu Dazi via MCP and Skills, making them available to more users and tasks. Co-built workbenches: professional SaaS vendors and industry systems build workbenches jointly with Baidu Dazi, bringing complete professional applications into task flows. Joint industry AI products: partners supply domain knowledge, specialist tools and proven processes, while Baidu Dazi supplies the Harness engine, general office capabilities and agent execution.

The third is the most interesting, because the promise to partners is that they "don't have to build general-purpose capabilities from scratch." Translated: an industry software company that wanted to do AI used to have to pick its own model, build its own agent framework and wire up its own tool calls. Now the platform takes that over, leaving partners with what they already do best — domain knowledge and process. It's the standard playbook for widening supply on a platform: cut the fixed cost of joining in exchange for a greater variety of tasks in the ecosystem.

In a Factory, the Real Money Is in Power and Water

Back to Robin Li's metaphor. A factory floor carries an implicit premise: the building doesn't make anything — the production lines and workers inside do. If the intelligent organization really is the new factory floor, Baidu AI Cloud wants to be the utilities and the production line: it doesn't decide what you make, but it decides whether you can start and how fast you run.

Whether the story holds up comes down to a plain question: who judges whether a task was done well? Acceptance criteria for traditional IT systems are resources and availability, written in black and white. "Task completed" is far fuzzier — tell an agent to "process this batch of invoices" and finishing is not the same as getting it right. Baidu's three-tier split into general, professional and custom agents answers who builds them, not who signs off on the result.

"Intelligent cloud" has been a buzzword for years, usually meaning little more than a layer of AI bolted onto a cloud. What's different this time is the unit of measurement: from selling resources to selling completed tasks. Change the unit and pricing, acceptance and liability all have to be rewritten — and that part went unaddressed.