--- title: "Alibaba Cloud rebuilds its stack so AI agents can plan, act, and remember" date: 2026-10-19 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-corp-alibaba-agentic-cloud language: en --- # Alibaba Cloud rebuilds its stack so AI agents can plan, act, and remember > At its 2026 Apsara Conference, Alibaba Cloud laid out an "Agentic Cloud" strategy — AgentCore to build and govern agents, a 100,000-accelerator domestic supernode, and a "Context Engine" that turns enterprise data into real-time memory. For two years the cloud pitch was simple: rent a GPU, spin up a model, build a chatbot. That story is quietly ending. At its 2026 Apsara Conference (云栖大会) in late September, Alibaba Cloud stopped talking about models as the product and started talking about the **environment an agent needs to actually get work done** — plan it, run it, and remember what it learned. The banner term is **"Agentic Cloud" (Agentic Cloud)**, and the strategy rests on three pillars the company now repeats like a mantra: **Model, Harness, Context**. ## From Q&A to execution The throughline of the keynote, delivered by Alibaba Cloud CTO Li Feifei (李飞飞), was a shift in expectation. Earlier large models (大模型) were judged on how well they answered questions. Agents are judged on whether they finish a task: book the trip, reconcile the ledger, file the ticket. That demands more than a smarter model — it demands infrastructure that can hold a long task open, recover from failure, and feed the agent the right data at the right moment. Alibaba's answer is a rebuilt technology stack spanning compute, storage, networking, runtime, identity, and enterprise data, organized around those three pillars. ## The three layers: Model, Harness, Context - **AI Native Cloud (the Model layer)** handles large-scale training and inference. Here Alibaba showed off its **Lingjun Zhenwu M890 (灵骏真武M890)**, described as a domestically developed 100,000-accelerator supernode cluster. The company says it already serves models like Qwen3.8 Max and Kimi K3 that exceed 2 trillion parameters. - **Agent Native Cloud (the Harness layer)** is where agents are built, run, and secured. The headline product is **AgentCore**, a governance platform for the agent lifecycle. - **Context Engine (the Context layer)** turns fragmented enterprise data — documents, business systems, chat logs, multimodal files — into live, queryable memory so an agent knows not just "what" but "what to do right now." A coming upgrade, the **Lingjun Zhenwu V900** supernode, is pitched at roughly three times the performance of its predecessor, with a single cluster scalable toward 500,000 accelerators. ## AgentCore: one control room for every agent **AgentCore** is the piece most relevant to enterprises. Rather than letting teams scatter agents across notebooks and scripts, it offers one place to build, register, run, and watch them. Its stated capabilities include unified identity and permissioning for both humans and agents, multi-agent and human-agent team orchestration, failure recovery for long tasks, and full-link tracing of the call chain, tool chain, and decision chain. The pitch is operational maturity: agents move from prototypes to production when they can be governed like any other piece of business software. Alibaba also announced an **Agent Security Center** covering the agent lifecycle with real-time threat detection. ## A storage and network rethink for agent workloads Agents are brutally I/O-hungry — they re-read context, cache results, and hit memory constantly. Alibaba's new-generation **CPFS** storage, paired with a KVCache management system called **Tair KVCM**, is tuned for exactly this. The company claims the storage tier cuts AI storage cost by about **69%** and delivers throughput at the 100 TB/s class. On the network side, **HPN 8.0 Pro** is described with 100 Pbps of bandwidth and more than 130,000 800G ports in a single cluster, with the share of service impact from network upgrades or failures roughly halved. On the training side, the **PAI** platform added asynchronous AgenticRL support, which Alibaba says can complete a state-of-the-art post-training cycle in about **5 days** when paired with Qwen models. ## The context problem is the real product Perhaps the most telling claim is about the **Context Engine**: in knowledge-intensive scenarios like customer service, AI coding, and data analysis, Alibaba says it can cut token usage by up to **67%** by giving agents better, compressed memory instead of stuffing raw context into every prompt. If true at scale, that is where the money is — tokens are the new compute bill. According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the strategic move is treating context engineering — not just bigger models — as the product layer; whoever organizes enterprise memory cleanly is positioned to win the agent era. ## Honest limitations Most of the headline efficiency numbers (69% storage saving, 67% token reduction, 5-day training cycle, 3x supernode performance) are **vendor claims from a launch keynote**, not yet validated by independent benchmarks we could locate. The "2-trillion-parameter" framing for served models reflects Alibaba's own conference statements and should be read as marketing language until corroborated. AgentCore and the Context Engine are tightly coupled to Alibaba Cloud's own ecosystem, so the governance benefits come with platform lock-in. And "agents that remember" raise fresh data-governance questions — condensed enterprise memory is still enterprise data, and its handling needs its own compliance story. ## What readers can do now - **Map your "context debt."** List the data an agent would need to do a real task in your company and how fragmented it is — that gap, not model choice, is usually the bottleneck. - **Pilot one governed agent on a platform with audit trails** before spreading agents across teams, so you learn the permissioning model early. - **Track token spend per task, not per model**, since context engineering — not raw model size — is becoming the main cost lever. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-alibaba-agentic-cloud Free to quote with attribution and a link to the original.