What this is
At the 2026 Yunqi Conference, Alibaba Cloud repackaged five data products — DLF, Flink, EMR Spark/Ray/Daft, StarRocks, and Milvus — into the "Agentic Lake + Agentic Streaming" system, sending a clear signal: the primary customer of data infrastructure is shifting from humans to AI Agents (AI programs capable of autonomously executing tasks).
Wang Feng, head of the open-source big data platform, summarized this evolution through "four shifts" — data structures expanding from text to audio, video, and other full-modality formats; analysis operators (built-in functions that process data) upgrading from BI statistics to AI inference; management approach moving from closed data warehouses to open lakehouses (hybrid architecture combining data lake and data warehouse); and user type shifting from human analysts to AI Agents.
On specific products: DLF uses Apache Paimon as the core lake format (lake format refers to unified specifications for organizing files within a data lake), compatible with Lance and Iceberg, integrating with object storage, relational databases, real-time video streams, and knowledge bases like DingTalk and Yuque; EMR Ray + Daft handles vectorization (converting images and text into numerical sequences AI can understand), annotation, and quality assessment — the AI data preparation work; StarRocks provides scalar + vector + full-text hybrid retrieval (simultaneously querying by multiple criteria); Flink upgrades from a stream processing engine into a Streaming Agent runtime platform.
The system has already been validated in two scenarios: Wuji Technology uses EMR Ray to build a first-person hand 3D reconstruction pipeline; Zhuoyu Technology uses StarRocks for fault backtracking and similar-scenario retrieval in autonomous driving data.
Industry view
Supporters see this as the necessary path. AI-native scenarios like autonomous driving and embodied intelligence (enabling AI to control physical devices such as robots or cars) generate training data daily that mixes images, point clouds (3D coordinate point sets generated by LiDAR), and video. The traditional data warehouse mindset of "only storing tables" cannot sustain this. Alibaba Cloud's Paimon 2.0 natively supporting LeRobot, RLDS, HDF5, and other embodied intelligence industry formats targets the customers with the most intense training data needs.
But there are also sober voices. Several risks worth noting: First, the integration complexity of "five-product coordination" is extremely high — enterprise IT teams must simultaneously maintain lake formats, stream processing, vector retrieval, and AI pipelines — operational costs easily concealed at product launches. Second, Agentic Lake's premise is that Agents genuinely require this level of sophisticated lakehouse architecture. If Agents primarily rely on API calls rather than directly querying data storage, this intermediate platform layer may be redundant. Third, the storage and computational overhead of full-modality data is significantly higher than structured data — AWS and Databricks are pushing similar concepts, but the cost models for all of them remain unclear.
Impact on regular people
For enterprise IT: When planning data platforms over the next two years, teams need to reassess the ratio of "human" and "Agent" users — traditional data warehouse approaches may fail to support AI training scenarios, but blindly deploying "full-modality lakehouse" could also yield an unclear bill.
For individual careers: The skill stack of data engineers and platform architects is being rewritten — terms like Paimon, vector retrieval, and Agent runtime, which previously only appeared in AI infrastructure teams, will gradually enter hiring requirements for broader data positions.
For consumer markets: Impact remains far off for now. The maturity of Agentic Lake determines whether the training data pipeline behind your home robot vacuum and autonomous driving car runs smoothly, but it won't directly change product experience — at least not within the next two years.