Alibaba Cloud’s Agentic Cloud Strategy at Apsara 2026

At the 2026 Apsara Conference, Alibaba Cloud presented Agentic Cloud as more than a place to run AI models. Its proposed stack combines model infrastructure, a managed environment for enterprise agents, and the data and memory those agents need to work with. The company organized that strategy around three scenarios—Model, Harness and Context—and three corresponding layers.34

Three scenarios, three cloud layers

• Model → AI Native Cloud: Infrastructure for large-scale model training and inference.3
• Harness → Agent Native Cloud: Services for deploying, running and securing enterprise AI agents.34
• Context → Context Engine: Real-time data and long-term memory intended to give agents information beyond an individual interaction.34

The distinction matters because Alibaba is pitching an operating stack for agents, not just access to a model. The model supplies intelligence; the harness supports agent execution and management; and the context layer supplies information the agent can use over time.34

AgentCore: Building, running and governing agents

AgentCore sits in the Agent Native Cloud layer. Alibaba describes it as an enterprise platform for building, running and managing agents across their lifecycle, including oversight of human–agent collaboration and monitoring agent performance.34 A conference report also describes support for long-running tasks, retries, resuming interrupted work, asynchronous execution, isolated execution environments, identity permissions and audit trails.7

Agent security is a related concern, but Agent Security Center should not be mistaken for a verified guarantee that agents are safe. Alibaba identifies it as a product in its agent-security portfolio, and describes Agent Loop as one of its log-analysis modules.2530 The provided conference evidence does not establish which Agent Security Center controls are integrated with AgentCore or quantify their effectiveness in production.

CPFS: Storage for model workloads

Alibaba positions its next-generation Cloud Parallel File Storage (CPFS) for AI model training. It claims hundred-terabyte-per-second-class throughput, hundred-million-class input/output operations per second, and a 69% reduction in enterprise AI storage costs.17 A conference report adds a claimed 100-PiB single-file-system scale, five times the previous scale, alongside 50% shorter average model-start times and 30% higher peak compute utilization.7

Those are vendor-reported results, not independently verified benchmarks. They also should not be assumed to describe every CPFS deployment: Alibaba’s existing CPFS product documentation lists different, configuration-specific throughput limits.23

HPN 8.0 Pro: Network capacity and resilience

For the network beneath the stack, Alibaba says HPN 8.0 Pro delivers 100 petabits of bandwidth and supports more than 130,000 800G ports in a single cluster.17 A conference report says its redundant design is intended to prevent optical-transceiver or link failures from interrupting service. It also reports that the impact of network upgrades and failures fell from 50% to 25% compared with the previous generation, without defining precisely how that impact was measured.3

Taken together, the announcements show Alibaba’s intended path from model training to managed agents with persistent context. Whether the stated storage, networking and governance benefits hold for a particular enterprise workload requires more detail than the conference claims provide.3717