Alibaba Cloud's Agent-Era Data & AI Infrastructure: From Multi-Modal Training to Self-Evolution

At the 2026 Yunqi Conference, Alibaba Cloud CTO Feifei Li identified Model, Context, and Harness as the three key elements for agent deployment, positioning Alibaba Cloud's "Agentic Cloud" to deliver four pillar capabilities: full-domain cognitive engine, autonomous operation and continuous evolution, trusted collaboration and governance, and real-time intelligent decision-making.

Data-Intelligence-Application Flywheel

Alibaba Cloud VP Junhua Wang described a self-driving flywheel: data flow produces intelligence, intelligence materializes as agents, and agents generate new data to drive system evolution. The flywheel's speed depends on three efficiencies — model production, intelligence deployment, and system evolution — around which Alibaba Cloud's big data AI platform released full-link product upgrades.

1. Multi-Modal Data Processing & Large-Scale Training

As AI enters the physical world, data types expand from text to video, point clouds, motion trajectories, and sensor signals. Model intelligence ceilings depend on data quality and processing speed.

MaxCompute AI Compute Engine Upgrade

MaxCompute now unifies scheduling of CPU, GPU, and LLM token compute resources. MaxFrame distributed Python engine and SQL AI Function engine deeply integrate Qwen flagship models, enabling multi-modal data understanding and computation with a single line of code. For data development, autonomous driving, and embodied intelligence scenarios, MaxCompute introduced industry Skills that build complete data pipelines from one natural-language sentence, drastically improving development efficiency.

Open-Source Big Data Platform: "Ingest-to-Train" for Multi-Modal Data

EMR Serverless Spark provides a multi-modal embodied intelligence data processing platform supporting Spark, Ray, and Daft engines, natively connecting to Paimon and Iceberg lake table formats. Combined with serverless second-level elasticity, it reduces embodied intelligence data processing time by 40% and directly outputs LeRobot/RLDS-format datasets from PB-scale raw collected data.

EMR Serverless StarRocks adds full-text search, vector search, and KV point lookup to its extreme OLAP capability, enabling multi-route recall in a single SQL — "one dataset, one engine, one query" — widely adopted in intelligent driving and vehicle-to-everything scenarios.

PAI-DLC 3.0 for Large-Scale Distributed Training

PAI-DLC 3.0 natively supports the Zhenwu M890 64-GPU super-node, combined with HPN 8.0 and CIPU 2.0 to build a multi-level topology, 10,000-GPU interconnected compute foundation. A new intelligent diagnostic Xfuse module fuses insights from the compute foundation, training framework, and task runtime, allowing an agent to automatically perform root-cause analysis. In practice, multi-modal model end-to-end training accelerated 2.4x, and Qwen 3.8 reinforcement learning throughput increased 3.8x.

PAI Embodied Intelligence Full-Stack Solution

Data layer (PAI-EgoX): Connects first-person data pipelines, supports mono/stereo, pinhole/fisheye, multi-view video processing, ViPE/VITRA dual pipelines, with pre/post quality inspection to output high-quality datasets.

Training layer (PAI-RoboX): Provides VLA model and world model training frameworks, supporting supervised fine-tuning, reinforcement learning, and evaluation. Enables joint training with real-robot, simulation, and first-person data so diverse sources collectively serve model learning.

Compute layer (PAI-TurboX): Optimizes across data loading, scheduling, compilation, operators, precision, and storage — six dimensions — to accelerate large-scale training and fully unleash compute power.

2. Agent Data Plane & Enterprise-Grade Inference

Powerful models alone do not guarantee agent value. Agents must connect to real enterprise business data and run on efficient, secure inference infrastructure.

OpenLake Evolves to Agentic Lake

OpenLake uses unified, real-time open lake storage to carry multi-modal data, pushing big data, search, and AI engines to cross-engine collaboration and agent-readiness. A cross-engine semantic layer built with DataWorks empowers agent services, while the AI-native big data service ADA supports multi-agent collaboration. Enterprise multi-modal data built on OpenLake can serve BI, AI, and agents, ushering in the multi-modal open data lake agent era.

Compute engines on OpenLake are fully "Agentic-ized": they not only work around the same data but also expose their capabilities to agents for direct invocation. DataWorks Data Agent provides unified data development, governance, and task orchestration. Users express business goals in natural language; DataWorks orchestrates data processing flows and invokes appropriate engines without requiring users to manage underlying engine operations or prepare separate data copies per engine.

ADA: AI-Native Big Data Service

ADA is an agent-form product using a multi-agent collaboration architecture. Through a unified conversational entry, users invoke offline, real-time, streaming analytics, AI training/inference, search, and other compute engines via natural language, driving multiple agents to collaboratively complete end-to-end tasks across data analysis, engineering, governance, operations, and data science.

Flink Streaming Agent for Real-Time Intelligence

To meet agents' high real-time demands, Flink introduced Flink Streaming Agent, supporting multi-modal data change capture and streaming AI processing/understanding. It provides a complete Agent OS runtime with per-record dynamic orchestration and long/medium/short-term memory.

PAI-RLS & PAI-InferX: Enterprise Model Training & Inference Combo

PAI-RLS offers fully managed cloud reinforcement learning training, combining high-performance RL training architecture, full-chain trajectory tracking, and a self-developed training acceleration engine to lower enterprise model customization barriers.

PAI-InferX is a high-SLO-guaranteed inference platform tailored for agent inference loads. Core capabilities include LLM intelligent routing, cache-aware scheduling, KV cache storage, model pre-warming, and global compute orchestration, enabling enterprises to build proprietary LLM inference services with lower cost, higher SLO guarantees, and stronger autonomy.

3. System Self-Optimization: Accelerating the Application-to-Evolution Flywheel

New data and compute loads generated by agent applications become fresh fuel for continuous intelligence evolution.

PAI-CrystalLLM: High-Assurance Model Simulation Training Engine

Industry-leading models (including Qwen) have evolved from pre-training to AgentRL and toward RSI (Recursive Self-Improvement), posing new challenges for training platforms. PAI-CrystalLLM simulates 10,000-GPU cluster training effects with <1% GPU resource consumption, enabling algorithm engineers to rapidly trial fine-grained training method innovations. In internal Alibaba Cloud practice, it delivered 2%-10% absolute MFU improvement; the related paper was accepted at SOSP'26.

Hologres: Self-Evolving Query Optimizer

Traditional query optimization relies on "heroic" expert tuning — hard to replicate and too slow for changing workloads. Hologres pioneered a new path: workload analysis, candidate set extraction, minimal-change implementation, evolution沉淀, performance gatekeeping, correctness gatekeeping, and continuous feedback iteration — turning tuning into a continuously running engineering closed loop that lets the optimizer self-evolve. In production, Hologres autonomously discovered and completed 28 optimizations in 4 months, boosting optimizer performance 100% and lake query performance 38%, and topped the TPC-H 3000GB benchmark with both performance and cost-effectiveness first place.

Customer Evolution Cases

Yiwu Small Commodity City: Used MaxCompute AI Function for massive product image understanding, processing >1M images daily, doubling performance vs. self-built solution.

Qunche Intelligence: Built cloud Umi Data + AI Pipeline on MaxCompute, Hologres, DataWorks, and PAI, achieving full-process data automation, >10x overall throughput increase, and ~50% model training efficiency gain.

CCTV Sports: Deep collaboration with Flink Streaming Agent for live sports real-time AI commentary. Agents combine video, audio, and events to understand match context in real time and auto-generate synchronized commentary, greatly improving commentary efficiency and viewer experience.

In the Agentic AI era, data, intelligence, and applications form an inseparable, continuously evolving system. Alibaba Cloud's big data AI platform keeps iterating to provide a solid foundation for enterprise agent innovation.