As AI progresses from 'talking and writing' to 'doing and acting', cloud computing's primary users, SaaS pricing models, and the foundational infrastructure for enterprise AI implementation are being redefined by intelligent agents.
During the World Artificial Intelligence Conference 2026, at the Alibaba Cloud-hosted "Agentic Cloud: Infrastructure for the Agent Era Forum" in Shanghai, Alibaba Cloud and industry partners discussed topics including Agent Infra, Agentic Products, AI Native Cloud, Data Plane, and Security, revealing multiple product advancements, industry case studies, and infrastructure data.
The forum did not provide new financial revenue growth rates, orders, or performance guidance. However, from a market perspective, the core signal from Alibaba Cloud is that the competition in AI cloud is expanding beyond pure model capability to include inference efficiency, data platforms, Agent runtime, security governance, and delivery capability focused on business outcomes. In other words, the commercial potential for AI cloud is moving further from "selling compute power and model calls" towards "supporting agents to complete business results."
The Core Shifts: From Model-Centric to Outcome-Driven
During a dialogue session, Alibaba Cloud Intelligent Group Chief Technology Officer Li Fei-Fei and PyTorch Foundation Executive Director Mark Collier discussed the evolution of AI infrastructure.
Mark offered a key judgment: "The model itself is no longer the sole core. The model is certainly very important, it is the brain of the agent." He believes enterprises will use multiple models simultaneously, develop proprietary models, or customize open-weight models into smaller, larger, or more specialized versions. "You not only need to orchestrate the agent itself, you also need to decouple the agent from the models it uses."
Li Fei-Fei further pointed out from a business model perspective that the pricing logic of the SaaS industry may change: "Looking across the entire SaaS industry, enterprises purchase software services based on subscription for results, not priced on resource consumption—CPU cycles or memory usage."
He predicts the market will quickly migrate to an agent service model: "These agents are designed to solve specific tasks and specific problems, and the customer pays for the final result."
Mark offered a judgment in the same direction: "The customer will say: I don't care about the cost, I don't care how many GPUs are needed, I just want to pay for the result. But someone has to pay for all those GPUs and data centers."
This means that for cloud providers, inference efficiency and cost optimization of backend infrastructure remain critical. Mark mentioned that a DeepSeek inference software optimization case showed "token costs were reduced by 60% within six months," and "achieved nearly three times the throughput on exactly the same hardware."
New Metrics Emerge: Token, Agent, Database Instances
In the opening speech, Deputy Director Sheng Lei of the National Information Center stated that agents are moving from concept to scaled implementation, and AI is advancing from "talking and writing" to "doing and acting."
He proposed that the core measurement method in the intelligent economy era is changing: "The industrial economy looks at kilowatt-hours, the internet economy looks at traffic, at bits. When we reach the intelligent economy era, Token is becoming an important yardstick for measuring AI's actual productivity."
Sheng disclosed that China's average daily Token call volume "exceeded 140 trillion times within two years." He also stated that future computing service models will become internationalized, exploring "new forms such as offshore computing, inbound data processing, and Token export."
On the cloud resource usage side, Alibaba Cloud Database Head Yang Xinjun gave a more direct change: the cloud's primary user is shifting from humans to agents.
He said that Databricks data shows "80% of database instances are created by agents, 97% of branches are created by agents." Alibaba Cloud has observed a similar trend: "Recently, 80% of our PostgreSQL databases were created by agents."
More specific data is: "Over the past five full years, the entire PostgreSQL database created about 40,000 instances. Recently, in the past few months, we've seen agents create nearly 120,000 databases."
Yang Xinjun stated that the month-on-month growth rate for agents autonomously creating databases reached 300%. He summarized: "Data is increasingly becoming the wealth for enterprise AI implementation."
Infrastructure Innovations and Market Share
Alibaba Cloud Intelligent Group R&D Vice President and Elastic Compute Head Wu Jiesheng stated in a speech that Alibaba Cloud is "one of the few cloud service providers globally with full-stack AI infrastructure capabilities," covering IDC, self-developed hardware and chips, cloud product services, the Qwen large model, and Bailian inference service.
He disclosed that Alibaba Cloud's full-stack AI cloud service market share is 40.1%, and stated that "at least half of the large model companies in China run on Alibaba Cloud."
On the training side, Alibaba Cloud released the "Super Node Instance" based on the T-Head Zhenwu M890 AI chip, ICN Switch 1.0 chip, and Panjiu AI super server node. A single super node is configured with 64 cards, with inter-card interconnect speeds reaching 800GB. Each M890 chip is configured with 144GB of memory, with the overall super node memory reaching 9TB.
Wu Jiesheng said the latest M890 chip has shown training performance three times that of the previous generation 8101 in some large model training tests.
On the inference side, Alibaba Cloud released a new KV Cache Store storage system to address the explosive growth of KV Cache data under long-context, multi-turn dialogues. Wu stated that in customer POC validation, KV Cache Store "can increase cache hit rate by up to 20%," thereby reducing inference cost and improving inference efficiency.
He also mentioned that Alibaba Cloud's TokenWorks can reduce model launch time "from days to tens of minutes"; the container-based inference solution can reduce cold start latency by up to 90%, with some customers seeing GPU utilization increase by 35% and average output time per token decrease by 60%.
For Agent runtime, Alibaba Cloud launched capabilities like a secure isolation sandbox and Agentic FS. Wu stated that Alibaba Cloud can support "rapidly launching 100,000-level sandboxes per minute," with single-region support for million-level sandbox scale; cold start time can reach the hundred-millisecond level, with hot start at the ten-millisecond level.
System-Level Solutions Over Single Chips
T-Head Semiconductor Vice President Gao Hui stated that agent applications are reshaping data center workloads, with high-frequency collaboration, tool calling, and long-context reasoning all challenging compute utilization and TCO.
She emphasized: "These problems cannot be solved by upgrading one chip or two chips; a system-level holistic solution is needed."
T-Head's answer is a full-stack upgrade of compute, storage, and network. Gao disclosed that as of April this year, Zhenwu chips have shipped 560,000 units, serving over 20 industries and more than 400 customers, becoming "the most widely applied AI chip in China in terms of scenarios."
Simultaneously, T-Head formally announced the open-sourcing of the Seal software stack. Gao explained: "The real explosion of the AI era will not come from one or two truly powerful chips, but from a more open, collaborative, and efficient full-stack computing system."
She also provided ecosystem adaptation data: there are 3,277 mainstream AI repositories on GitHub with more than 10 stars, and Seal has adapted 3,248 of them; it also covers over 260 mainstream training and inference frameworks including PyTorch, TensorFlow, vLLM, and SGLang.
Industry Implementation Cases
Full Truck Alliance Group AI Algorithm Director Gao Yiming provided a case of agent implementation from the logistics industry. He stated that the essence of logistics platform agentification is: "from humans operating software to humans entrusting goals."
In freight scenarios, drivers do not come to the APP to "browse loads," but have clear goals: where to go, how much to earn, whether they can load, whether they are willing to travel empty. Gao said: "The driver's agent does not answer every sentence, but must understand context, maintain state, and transform goals into executable plans."
After Full Truck Alliance launched the relevant agent system, driver usage rate increased from 15% to 35%, with next-time retention after one use at 71%. He also disclosed that drivers spend an average of 169 minutes daily on the platform looking at loads, "about 2.5 hours, 10% of their life."
Gao Yiming stated that if a million drivers entrust their goals to the system, platform supply would "increase tenfold"; the response time after a load is posted is also expected to shorten from over ten or twenty minutes to 3 minutes, 5 minutes.
He summarized the future form of the platform: "Agents are responsible for handling uncertainty, while engineering is responsible for executing certainty. Ultimately the platform defines how to collaborate credibly."
Bridging the Gap to Production
Alibaba Cloud Intelligent Group Cloud Native Application Platform Head Zhou Qi pointed out that over the past year the market has seen many exciting Agent Demos, but "there is a very deep chasm between demo and production."
He believes the real challenge for enterprises is not whether they have an Agent, but how to let Agents "cross roles, cross systems, cross boundaries" into enterprise processes and collaborate stably.
Alibaba Cloud proposed a three-layer enterprise agent architecture: Infra, Desktop, and Platform. Infra provides a trusted runtime environment, Desktop lets Agents enter real work environments, and Platform is responsible for construction, governance, collaboration, and evolution.
Zhou disclosed that within the Alibaba Cloud cloud native team, an agent system has been used to connect open-source development, governance, Q&A, and quality tracking. Currently, 15 agents are running, providing 7x24 online service, handling 85% of technical Q&A, reducing operational support time by 90%, and compressing demand response time from 7 days to 1 day.
He said: "What the enterprise ultimately possesses is not a batch of Agents, but an enterprise workflow that continuously produces Agents and continuously optimizes to unleash their value."
Application Delivery and Security Imperatives
On the application generation front, Alibaba ATH Business Unit MASS Business Line Miaowu Deputy Head Zhou Hengmin launched Miaowu Team Edition. He stated that Miaowu is positioned as an "AI application creation platform for everyone," where users can generate websites, mini-programs, apps, etc., via natural language and publish them as accessible production-grade apps with one click.
Zhou Hengmin disclosed that as of the forum day, tens of thousands of users create and publish on Miaowu daily, most without technical backgrounds, including product managers, operations personnel, teachers, students, designers, and entrepreneurs.
Team Edition addresses enterprise concerns like unified procurement, resource sharing, asset transfer, permission management, and team collaboration. Zhou Hengmin said: "Miaowu Team Edition completes the upgrade from a personal tool to an organizational productivity platform."
Andr Pro Product Head Long Dongheng emphasized from a "Delivery Agent" perspective that generating code is only the first step; real business systems also need data, deployment, permissions, security, and continuous operation.
He said: "Humans don't need to understand Infra, but Agents must understand Infra better." In his view, when a user says "help me make an ordering system," the Agent must not only create the webpage but also "set up the database, assign the domain name, deploy the website." He summarized: "Coding Agent focuses more on code, while Delivery Agent focuses more on the overall picture."
Security as a Prerequisite for Scale
Alibaba Cloud Intelligent Group Cloud Security Product Head Zhu Jianyue warned that the more autonomous the Agent, the more important the security boundary. He stated that the number of machine identities has already surpassed humans, and enterprises are in an environment of human-Agent symbiosis.
He cited data stating that 57% of enterprises have deployed agents, but Gartner predicts nearly 40% of projects will fail due to security risk control issues.
Zhu Jianyue emphasized: "Agent security is no longer an option today; it is a necessity in business development. Otherwise, your Agent will only become a ticking time bomb."
Alibaba Cloud proposed three-layer unified protection at the Infra layer, model inference service layer, and Agent application layer, and launched Agent Security Center capabilities including asset identification, vulnerability detection, AI Red Team, runtime security, and log tracing. Zhu stated that Alibaba Cloud has launched 150 detection capabilities targeting Agent vulnerabilities, and integrated runtime security capabilities into Bailian and AI security gateways, with full-link latency controlled within 100-120 milliseconds.
Comments