Alibaba Unveils AgentCore And Agent Native Cloud At Apsara 2026

At Apsara 2026, Alibaba Cloud unveiled AgentCore, Agent Native Cloud and a stack of new infrastructure - HPN 8.0 Pro, next-gen CPFS and OpenLake - built to run enterprise AI agents at scale.

Alibaba Unveils AgentCore And Agent Native Cloud At Apsara 2026

Hangzhou-based Alibaba Cloud used its flagship Apsara 2026 conference to lay out a full-stack agentic AI roadmap, unveiling AgentCore, an enterprise-grade lifecycle platform for AI agents, alongside the broader Agent Native Cloud. The rollout stretches from silicon and networking through data lakehouses to the Qwen model family, and puts the Chinese hyperscaler in direct competition with AWS Bedrock AgentCore and Google Cloud's Vertex AI Agent Builder.

AgentCore and a Context Engine that cuts tokens 67%

AgentCore is pitched as a governance layer that handles building, running, monitoring and securing agents through their full lifecycle, with a Context Engine that provides real-time context and long-term memory. Alibaba says the Context Engine reduces token consumption by up to 67% on knowledge-intensive workloads by dynamically caching and compressing retrieved documents. Its OpenLake multimodal data lakehouse, released the same week, claims a 38% reduction in total cost and 40% faster query response times against comparable open-source stacks.

Alibaba Cloud agentic cloud stack

Silicon and 100-petabit fabric under the agents

Underneath the agent stack, Alibaba's T-Head unit refreshed its accelerator lineup with the Zhenwu V900 AI processor - 216 GB of memory, 1,200 GB/s inter-chip bandwidth and 3x the performance of the previous M890, with mass production slated for Q1 2027. HPN 8.0 Pro, the new AI networking backbone, ships with 100 petabits of aggregate bandwidth. Chairman Joe Tsai framed the pitch under the theme "Intelligence Goes Beyond," while CEO Eddie Wu committed to more than 20GW of global data-center capacity by 2032, alongside China's broader chip-and-cloud push.

Qwen family gets a big upgrade

Alibaba disclosed that Qwen 4 is already in training, with Qwen 4.5 and Qwen 5 targeting 5-10 trillion parameters. Qwen3.8-Max completed 33 automated self-improvement cycles that lifted its Artificial Analysis score from 40 to 45, and the company shipped Qwen3.8-LiveTranslate for simultaneous interpretation and Qwen-Audio-3.1-TTS-Next for cinematic audio generation - moves reminiscent of Google's Gemini 3.8 Flash TTS drop earlier this week.

Reporting based on coverage from TechAfrica News, Pandaily and HPCwire.

Category: AI & Technology

Tags: Alibaba China AI Agents AI Infrastructure agentic AI Cloud Computing

Related Articles