Alibaba Zhenwu M890 AI Chip: Specs, Infrastructure and Where It Fits in 2026

Table of Contents

Alibaba is expanding its AI infrastructure stack with its own accelerator chips, high-density server systems and AI models designed for long-running agent workloads.

At the center of the company’s 2026 hardware push is the Zhenwu M890, developed by Alibaba’s semiconductor subsidiary T-Head. Alibaba says the M890 delivers three times the performance of its previous-generation Zhenwu 810E and is designed for both AI training and inference. Alibaba Group

The company has also introduced the Panjiu AL128 Supernode Server and Qwen3.7-Max, linking its chips, servers and software into a broader AI infrastructure platform.

What Is the Zhenwu M890?

The Zhenwu M890 is an AI processor developed by T-Head, Alibaba’s semiconductor design subsidiary.

Alibaba introduced the chip at its Cloud Summit on May 20, 2026. The company positioned it specifically for the computing requirements created by modern AI agents, where applications may make large numbers of model calls and require high memory capacity and fast communication between accelerators. Alibaba Group

Alibaba says the M890 delivers three times the performance of the previous Zhenwu 810E.

The chip includes:

144GB of memory
Up to 800GB/s inter-chip bandwidth
Support for precision formats from FP32 to FP4
Support for AI training and inference
Architecture designed for large-scale AI workloads

These specifications come from Alibaba’s own disclosures and have also been reported by independent technology publications. AliHome

Why AI Agents Need Different Infrastructure

Traditional AI applications often generate a relatively predictable number of inference requests.

Agentic systems can behave differently. An AI agent may call multiple models and tools while completing a longer task. That creates additional requirements for:

Memory capacity
Inter-chip communication
Low-latency networking
High inference throughput
Efficient handling of many concurrent requests

Alibaba says the M890 was designed around these requirements.

The company demonstrated this approach with Qwen3.7-Max, which Alibaba says can operate continuously for up to 35 hours and perform more than 1,000 tool calls during long-running agent tasks. These figures are based on Alibaba’s own testing and should therefore be treated as company-reported results rather than independent benchmark results. Alibaba Group

Panjiu AL128: Putting 128 Accelerators Together

Alibaba’s hardware strategy goes beyond the individual processor.

The company introduced the Panjiu AL128 Supernode Server, a rack-scale system that combines 128 AI accelerators and is designed for large-scale model training and agent inference. Alibaba says the system is available through its Bailian model-service platform. AlibabaCloud

The system is intended to address the communication demands created when many AI accelerators work together.

This is important because large AI systems depend not only on accelerator performance but also on how quickly those processors can exchange data.

Alibaba’s Zhenwu Roadmap

Alibaba has also outlined a multi-year roadmap for its Zhenwu chip family.

The company announced:

Zhenwu M890 — introduced in 2026
Zhenwu V900 — planned for the third quarter of 2027
Zhenwu J900 — planned for the third quarter of 2028

The roadmap shows that Alibaba is treating its AI silicon program as a continuing hardware platform rather than a single chip release. Alibaba Group

The company has subsequently introduced the Zhenwu V900 at its 2026 Apsara Conference and said the new chip delivers three times the performance of the M890. Alibaba Group

That means the M890 should now be viewed as one stage in Alibaba’s broader AI accelerator roadmap rather than its final hardware offering.

How Widely Is the M890 Being Used?

Alibaba said in its May announcement that T-Head had shipped more than 560,000 Zhenwu chips and that more than 400 external customers across 20 industries had deployed them. AliHome

Alibaba has separately reported that its Zhenwu chips have reached more than 560,000 units shipped and that more than 400 external customers across 20 industries had deployed them. Those figures cover the broader Zhenwu family and should not be interpreted as M890-only shipment numbers.

This distinction is important when evaluating adoption because a chip family’s total deployments are different from the number of customers using one specific accelerator model.

Alibaba’s Full-Stack AI Strategy

The M890 is only one part of Alibaba’s AI infrastructure strategy.

The company’s current approach covers several layers:

AI chips: Zhenwu accelerators

CPUs: Yitian processors

Networking: Panmai and ICN technologies

Servers: Panjiu systems

AI models: Qwen family

Cloud services: Alibaba Cloud and Bailian/Model Studio

Alibaba says it is working on end-to-end optimisation across chips, servers, networking, models and inference engines. Alibaba Group

This approach allows Alibaba to optimise more of the AI stack rather than depending entirely on third-party accelerator hardware.

What the M890 Means for AI Infrastructure

The M890 illustrates an important change in AI infrastructure.

Cloud providers are increasingly looking beyond individual AI models and focusing on the complete computing stack required to operate them.

For enterprises, the relevant questions are therefore not only:

How fast is the accelerator?
How much memory does it have?
Which models can it run?

They also include:

How easily can the hardware scale?
What software ecosystem supports it?
How efficiently can multiple accelerators communicate?
Can the infrastructure support long-running agent workloads?
Is the hardware available through a cloud service rather than requiring direct procurement?

These factors can be just as important as a single benchmark number when evaluating AI infrastructure.

M890 vs the Next Zhenwu Generation

The M890 is already being followed by newer hardware.

Alibaba’s 2026 Apsara Conference announcement introduced the Zhenwu V900, which the company says delivers three times the performance of the M890 and is intended for large-scale training and inference. AlibabaCloud

This means businesses evaluating Alibaba’s infrastructure should pay attention to the generation of hardware being offered rather than assuming that the M890 represents Alibaba’s latest accelerator.

What Enterprises Should Check

Organizations evaluating AI infrastructure should consider several factors before choosing a platform.

Workload

Determine whether the system is primarily required for:

AI training
High-volume inference
Agent workloads
Model fine-tuning
Large-scale enterprise applications

Memory and networking

Large models can place significant demands on memory and accelerator-to-accelerator communication.

Software compatibility

Hardware performance is only useful if the required models, frameworks and deployment tools are supported.

Cloud availability

Accessing an accelerator through a cloud service can be substantially different from purchasing and operating physical hardware.

Long-term roadmap

A clear hardware roadmap can matter when organizations are planning infrastructure investments over several years.

What the M890 Means for AI Infrastructure

Alibaba’s Zhenwu M890 is part of a broader attempt to build an AI infrastructure stack covering chips, servers, networking, models and cloud services.

The M890 itself is notable for its 144GB memory, 800GB/s inter-chip bandwidth and Alibaba’s claimed threefold performance improvement over the Zhenwu 810E. The Panjiu AL128 extends the concept to large-scale multi-accelerator infrastructure. AliHome

The subsequent Zhenwu V900 announcement shows that Alibaba is continuing to develop the platform beyond the M890. For businesses, the more significant development may therefore be the full-stack approach rather than one individual chip. Alibaba is attempting to connect its processors, servers, networking technology, AI models and cloud services into a single infrastructure ecosystem.

Sources & Further Reading

Sources used for this article:

Alibaba Group — Alibaba Unveils New AI Chip, Flagship Model, and Rebuilt Cloud Stack AI for Agentic Era Alibaba Group official announcement
Alibaba Cloud — Comprehensive Full-Stack AI Upgrade for the Agentic Era Alibaba Cloud official announcement
Alibaba Group – New AI Chip and Full-Stack AI Strategy — Official Alibaba
Data Center Dynamics — Alibaba announces new AI chip for training and inferencing workloads Data Center Dynamics report
TrendForce — T-Head Zhenwu M890 and future chip roadmap TrendForce report