https://x.com/i/article/2105211214272024576
Full Stack AI Unpacked
Welcome to the September issue of AI Pulse.
This month was quite a busy one for us. On September 1, we became dual-primary listed in Hong Kong and on Nasdaq. Now, our Hong Kong-listed shares are available to Chinese mainland investors through the Stock Connect programs.
These developments come at an important point in our AI journey, after years of investment in chips, cloud, models, and applications. Those parts are not developing in isolation, and the value increasingly comes from how they work together as a full stack.
That connection becomes especially important as AI moves into very different kinds of real-world work. A productivity agent working across files, a digital human responding to a shopper in real time, and an industrial agent optimizing a port do not ask the same things of the technology beneath them. They need different combinations of speed, reasoning, context, reliability, and cost.
So this issue looks at the full stack through the agents running on top of it: what different kinds of agents need, and how those needs are beginning to shape the technology underneath them.
Why the Full Stack Matters
An AI application has to work as a complete system. It needs computing power to run, a platform that makes that computing power accessible, a model suited to the task, and an application that people can actually use. We have spent years building each of those parts: Kunlunxin chips, Baidu AI Cloud, ERNIE, and the products and agents built on top. As our CEO Robin Li has said, our “full-stack, end-to-end AI architecture is a key differentiator.” It also underpins the growth of our AI-powered business, which accounted for 50% of our General Business revenue in Q2, reaffirming our transition "from an internet-centric company to an AI-first company."
For a full-stack AI company, the goal is not simply to sit at the top of every model leaderboard. The value of the full stack comes from how those four layers work together. On performance, this means being able to bring different models and tools to bear on what a user is actually trying to do. A live customer conversation may prioritize speed and low latency, while a complex research or optimization task may need deeper reasoning. Because chips, cloud, models, and applications can be optimized together, the system can be tuned around what the application actually needs.
The same applies to efficiency. Because the layers are connected, optimization can work in both directions: application needs can help determine which models and computing resources are used, while improvements in models and infrastructure can make those applications more efficient to run. We can route different tasks to the most suitable models or match computing resources more closely to the workload. This matters because, as a recent McKinsey analysis argues, the breadth of AI adoption may depend as much on infrastructure economics as on advances in raw model capability.
That feedback loop becomes even more important as AI moves from training into real-world use. Gartner forecasts that inference will account for 55% of AI-optimized IaaS spending in 2026, surpassing training as organizations move AI into production applications and workflows. Agents are part of that shift: multistep, autonomous tasks require repeated inference, increasing demand on the infrastructure underneath.
Our CFO Henry He made the same case in recent interviews with Bloomberg and CNBC, describing the model as the link between infrastructure and applications and saying that the layers of the stack become more valuable when they work together. As those years of investment turn into meaningful revenue streams, he sees Baidu entering a second growth curve.
All of this gives us a different way to look at the application layer. Instead of simply asking which agent has the longest feature list, we can ask what kind of work it is trying to do, and what that work demands from the rest of the stack.
What Different Agents Ask of the Stack
We have long argued that AI creates its value through applications. Robin Li singled out agents as a promising direction back in 2024. As those agents become more specialized, however, the technical challenge changes with the job.
A general-purpose work agent needs to hold context across tasks and tools. A commerce agent needs to respond quickly enough for a live conversation. An industrial agent may need to explore thousands of possible decisions before recommending one. Those differences help explain why the application layer and the infrastructure underneath it cannot really be separated.
Productivity: Context, Tools, and Working Software
Productivity is one of the clearest examples because even within a single category, workloads can look very different.
DuMate works across apps and files on multistep tasks. Kuku AI brings agents into Baidu Wenku and Baidu Drive, where the system can work with documents, files, and accumulated context. Kooko AI gives international users a workspace for research, analysis, documents, and data, with memory that can carry context between projects.
These products need more than a model that can generate a strong response. They need access to context, tools, and files, and they need to keep track of what has already happened as a task moves from one step to the next.
Miaoda, known internationally as MeDo, puts a different kind of pressure on the system. Here the agent is not simply producing information. It has to turn simple instructions into working software, let the user test what was built, and respond when the requirements change. That matters most for one-person businesses, where the same person is often the founder, designer, and developer. Miaoda's latest upgrade adds agents for design, app generation, and testing, as well as a marketplace connecting businesses with creators for templates and custom development. A single prompt can now be the first step toward building a business.
That is a very different technical problem from summarizing a document or answering a question. Yet all four sit under the broad label of productivity agents. The category may be the same, but what each application needs from models, tools, context, and compute can be quite different.
Commerce: Speed in Real Time
Creating something is one kind of work. Responding to another person as a conversation unfolds is another. In commerce, Baidu Yijing shows what that can look like. Its digital humans can host livestreams, make videos and video podcasts, and interact with viewers in real time.
The key word here is "real time." A livestream cannot stop for a long reasoning process every time someone asks a question. The system has to understand what the shopper wants, generate an appropriate response, and keep the interaction moving naturally. That puts much more emphasis on latency and consistency than many other agent tasks.
Speed is only part of it. A business still has to respond to interest, explain an offer, and help someone decide what to do next. For some merchants, an interactive digital human can help them stay present through more of that process, including at hours when a human presenter is unavailable.
Some users start with a pilot and expand their use after seeing what the technology can deliver. One of its clients, for example, expanded its digital human livestreaming deployment to about 2.5 times its previous level after just one quarter of use.
Industry: Reasoning Through Constraints
Some of the hardest work for an agent happens where there is no single obvious answer. A port has to coordinate equipment, cargo, and time. A production line has to fit work around capacity and changing demands. Finding a better plan means weighing many possible choices against the constraints of a real operation.
Famou, our self-evolving decision agent, is built for problems like these. It focuses on production scheduling, process optimization, and logistics planning. Rather than follow one fixed set of instructions for every situation, it can explore possible solutions and refine them against the objective it has been given.
That means the challenge is different again. Speed still matters, but so do deeper reasoning, optimization, and the ability to work within real operational constraints. The business expert also remains essential: someone has to define the problem, set those constraints, and decide what improvement would actually matter.
Benchmarks are one test. Famou Agent 2.0 took first place on nine of MLE-Bench's 15 hardest tasks. The more telling result came from an automated port, where Famou helped improve a key performance measure by 10.21%.
The port example also shows how different the stakes can be across agent applications. An app prototype can be tested and revised before it reaches users. An industrial recommendation has to fit equipment, schedules, and people who are already at work. Success has to be something an expert can evaluate in the operation itself.
What This All Stacks Up To
Put these examples next to each other and "AI agent" starts to look like a very broad category, one with no single technical requirement that defines it. That's the point. Each application pulls differently on the models, cloud, and compute beneath it. And as those applications grow, their demands feed back into decisions about which models to run, how to route workloads, and where inference capacity is needed.
That is where the value of the full stack becomes easier to see. It is not simply about having chips, cloud, models, and applications. It is about being able to connect those layers around the needs of very different agents and adapt the technology underneath as those needs change.
As agents take on more roles in everyday work, commerce, and industry, the applications themselves may increasingly help determine how the rest of the AI stack evolves.
Meet AI, Evolving
- We launched AI, Evolving, our new podcast series exploring where AI agents are heading and what that means in the real world.
- Episode 1 follows Famou into a pine wilt disease research project and asks whether research agents could become part of the infrastructure of discovery. Episode 2 turns to DuMate to look at what it takes for work agents to earn our trust.
Miaoda Levels Up for Businesses and Creators
- We upgraded Miaoda, our no-code app builder, for both enterprises and individual creators. To date, Miaoda has served more than 40 million users, who have created 5 million business apps.
- The upgrade brings stronger AI agents for design, app generation, and testing, plus enterprise tools for private deployment and team collaboration. A new marketplace also connects businesses with creators for templates and custom development.
Kooko AI Takes on the Whole Workflow
- Formerly Oreate AI, the newly introduced Kooko AI has reached more than 10 million international users across nearly 200 countries and regions in its first year, with a focus on helping users turn complex briefs into finished deliverables.
- Built for professional productivity, Kooko AI brings research, planning, files, AI assistance, and content creation into one workspace, producing editable documents, presentations, spreadsheets, websites, visuals, videos, and more.
Apollo Go Hosts Hong Kong Delegation in Beijing
- In September, Hong Kong Secretary for Transport and Logistics Mable Chan, members of the Legislative Council Panel on Transport, and government representatives visited Apollo Park in Beijing, continuing the conversation around autonomous driving between Hong Kong and Beijing.
- Having previously ridden in an RT6 in Hong Kong, Secretary Chan said its performance was equally impressive in Beijing. We share her hope that the fully driverless trial on Airport Island will soon move into passenger service, bringing Apollo Go closer to welcoming its first riders.
Have a question about Baidu's latest developments, or something you'd love us to cover next? Leave a comment or DM us!
Until our next roundup, keep up with our latest AI developments and innovations by following us on LinkedIn and X.