The move worth recording from Alibaba Cloud's Apsara Conference 2026: rebuilding the data lake (a technology that centralizes and analyzes all of an enterprise's data) — built over the past decade for data analysts — as a toolbox directly callable by AI Agents (AI programs capable of autonomously completing multi-step tasks). We note this reflects a judgment — that competition in China's enterprise AI is shifting from "whose model is smarter" downward to "whether the underlying infrastructure can catch Agents."
What This Is
Alibaba Cloud's solution is a three-layer stack. The bottom layer is OpenLake upgraded to Agentic Lake (a data lake built for Agents): using DLF's Omni Catalog (a unified catalog — essentially registering scattered data tables and files onto a single central manifest) — structured data (orders, spreadsheets) and unstructured data (documents, images, video) can be discovered under a single system. Real-time data is handled by Fluss, historical data remains in Paimon 2.0, and the two are combined into a unified view through Union Read (a "real-time + historical read simultaneously" approach).
The middle layer consists of a series of Agentic Engines (compute engines callable by Agents) — MaxCompute, Hologres, Elasticsearch and others — exposing Skill (skill interfaces, packaged functional modules) and MCP (a standard protocol allowing Agents to invoke external tools, essentially a "plug-and-play" universal interface) to Agents. The top layer is DataWorks's Data Agent: powered by a "semantic graph" (mapping the relationships between tables, fields, and business meanings into a graph — essentially an enterprise data dictionary for Agents), it understands that "total sales" is not the same thing in finance and business departments, and exposes an OSI interface allowing third-party Agents to invoke it.
Industry View
Supporters argue the direction is sound — industry experience shows that most enterprise Agent failures are not caused by weak models but by data being unfindable, misaligned, or inconsistently defined. Without rebuilding infrastructure, even the smartest Agent can only perform in demos. DataWorks's semantic layer combined with MCP's open strategy is indeed closer to production than the patchwork of each department plugging into its own model.
But we need to keep two doubts. First, complexity: the full architecture involves more than ten Alibaba Cloud products — DLF, Paimon, Fluss, MaxCompute, Hologres, PAI and others — making the deployment barrier far higher than traditional data warehouses (databases). Second, ecosystem: "Agent-ready" is still a self-defined Alibaba Cloud standard, with no cross-cloud, cross-vendor portability verified. Overseas, Databricks and Snowflake are building similar architectures, and the outcome will be decided not by technology but by ecosystem lock-in.
Impact on Regular People
For enterprise IT: Over the next 1-2 years, when selecting a data platform, they will be forced into a decision far heavier than past upgrades — continue reinforcing traditional data warehouses, or rebuild the underlying layer for Agents. Budget and capabilities both need reassessment.
For individual careers: The hard-skill barrier of "knowing SQL (Structured Query Language, the foundational database operation skill)" is being eroded by tools like DataWorks that generate code from natural language. Data analysts must upgrade toward a "business-savvy + Agent-orchestration-savvy" combo, or their value space will be compressed.
For the consumer market: No visible change in the short term. It will take at least another 2-3 years for the dividend of Agentic Lake to reach consumers, depending on how many enterprises actually put Agents into production workflows.