Skip to content
NewsProduct launch

Alibaba launches AgentCore enterprise agent platform and full-stack Agentic Cloud at Apsara 2026

· by Pondero Newsdesk

The short version

At its annual Apsara Conference in Hangzhou on September 23, Alibaba unveiled AgentCore, a three-layer agentic cloud stack, a next-gen AI chip, and disclosed that Qwen 4 is in training.

Alibaba launches AgentCore enterprise agent platform and full-stack Agentic Cloud at Apsara 2026

Alibaba used its annual Apsara Conference in Hangzhou on September 23, 2026 to reveal the most complete public picture yet of a non-Western hyperscaler building chip-to-agent infrastructure entirely outside the US cloud ecosystem.

What

The centerpiece is AgentCore, a new enterprise platform for building, running, and monitoring AI agents across their full lifecycle, per the Alibaba press release published by HPCwire. AgentCore sits within a three-layer agentic cloud architecture. The bottom layer, AI Native Cloud, handles large-scale model training and inference. Above it, Agent Native Cloud contains AgentCore plus an Agent Security Center for runtime threat detection and compliance management. A third layer, the Context Engine, provides agents with real-time context and long-term memory via a service called Agent Context, which per Alibaba cuts token usage by up to 67 percent in knowledge-intensive scenarios such as customer service and AI coding.

A data platform component called OpenLake was upgraded into a multi-modality data lakehouse covering structured, semi-structured, unstructured, vector, and streaming data in a single copy. Per Alibaba, OpenLake reduces total infrastructure costs by 38 percent and query response times by 40 percent compared with traditional architectures.

On the hardware side, Alibaba's chip unit T-Head disclosed the Zhenwu V900, its latest AI training and inference accelerator. The chip delivers three times the performance of its predecessor, the Zhenwu M890 released in May 2026, with 216 GB of GPU memory and 1,200 GB/s of inter-chip bandwidth. Mass production is scheduled for Q1 2027. The company's networking stack, HPN 8.0 Pro, supports over 130,000 ports at 800G speeds across a single cluster with 100 petabits of aggregate bandwidth.

Alibaba also disclosed that Qwen 4 is currently in training, with a roadmap that extends through Qwen 4.5 and Qwen 5, projected to reach 5 to 10 trillion parameters. CEO Eddie Wu said the company targets 20 gigawatts of global data center capacity by 2032.

Teams avoiding US-hosted APIs get a full-stack option

Enterprise and government teams that cannot route inference traffic through US-hosted model APIs now have a clearer option. Alibaba has announced the full vertical stack at once: T-Head silicon, Qwen foundation models, cloud-native agent infrastructure, and enterprise security tooling, all under Alibaba Cloud's management plane.

The RSI data point is the most technically striking item in the announcement. Per Alibaba, Qwen3.8-Max ran 33 automated iterative training cycles spanning pipeline design, data validation, and error diagnosis, boosting its Artificial Analysis score from 40 to 45 without human involvement. In a separate chip design run, the model made over 10,000 EDA tool calls across 60-plus hours and reduced chip bus area by 42 percent with no performance regression. These are verifiable-sounding benchmarks, and they signal that Alibaba is conducting live recursive self-improvement research on Qwen ahead of the Qwen 4 cycle.

For enterprise developers, AgentCore's practical value depends on integration depth. The platform abstracts lifecycle management, retry logic, checkpointing, and audit trails, which are exactly the gaps that make rolling your own agent runtime painful at scale. How it handles interoperability with existing MCP-compatible tooling is not yet public.

What to watch next

Whether Alibaba publishes an AgentCore developer SDK with standard MCP tooling support will determine how quickly it gets adoption outside of China and Southeast Asia. Qwen 4 benchmark results are the other near-term signal: if the model reaches the top of the Artificial Analysis leaderboard on release, the case for the Alibaba stack as a viable US alternative strengthens materially. Both are expected before the end of 2026.

Sources