Alibaba Is Building AI Chips Specifically for AI Agents
Alibaba just unveiled a new AI processor built specifically for AI agents. And this changes what the race is actually about.
Let me break down what they’re doing and why it matters.
The Chip: Zhenwu M890
Alibaba’s semiconductor subsidiary, T-Head, developed the Zhenwu M890. It delivers three times the performance of its predecessor, the Zhenwu 810E.
But here’s the interesting part. The performance jump isn’t the real story.
The M890 is purpose-built for AI agents. That means software systems that need to:
- Retain long stretches of context
- Coordinate with other models in real time
- Execute complex multi-step tasks with limited human intervention
Those demands are different from what standard inference chips are optimized for. Standard chips focus on raw compute. Agent workloads need memory bandwidth and inter-model communication.
Alibaba isn’t designing around today’s dominant use case. They’re building for the workload profile they expect to define enterprise AI over the next several years.
The Roadmap
More significant than the chip itself is the roadmap Alibaba put alongside it.
- M890: Available now
- V900: Third quarter of 2027 (expected 3x performance gain)
- J900: Third quarter of 2028 (another 3x gain)
That’s a deliberate, sustained cadence of in-house silicon upgrades. It mirrors the kind of tick-tock product cycles Nvidia has used to maintain its lead in AI accelerators.
Alibaba is playing the long game.
The Bigger Picture: Huawei Parallel
The parallel to Huawei is worth noting.
Huawei laid out a similar chip roadmap for its Ascend line last year. Both announcements reflect the same underlying reality: Chinese technology companies have concluded that depending on foreign silicon is a structural risk they cannot accept.
The response has been to treat semiconductor development as a long-term capability-building exercise rather than a procurement problem.
The Investment
Alibaba’s commitment isn’t shallow. The company pledged more than 380 billion yuan (about US$53 billion) on cloud and AI infrastructure over three years last year. That’s their largest-ever investment commitment to the sector.
The M890 and its successors are downstream of that spending.
Real-World Traction
This isn’t lab hardware. T-Head has shipped more than 560,000 Zhenwu units to date. Over 400 external customers across 20 industries are deploying the chips, including automakers and financial services firms.
That’s a material production footprint. And it provides Alibaba with real-world deployment data at scale ahead of the M890’s rollout.
The new chip will be available to Chinese enterprise customers through Alibaba Cloud’s domestic model platform, Bailian, packaged inside the Panjiu AL128 server system that stacks 128 M890 accelerators into a single rack.
The Software Side
Alongside the hardware, Alibaba announced Qwen 3.7-Max, the latest version of its flagship LLM. It’s engineered for advanced coding and long-running agent tasks.
The company said the model can operate continuously for up to 35 hours without performance degradation. That’s a capability specification that only makes sense if you’re designing for extended autonomous operation.
The timing is deliberate. Releasing a chip and a model optimized for the same workload class on the same day is a platform play.
The Strategy
Alibaba is building a closed loop:
- Its own silicon (T-Head)
- Its own model (Qwen)
- Its own cloud delivery (Bailian)
Each component reinforces the others. The combined stack is designed to reduce enterprise customers’ dependence on any external vendor.
“This push into purpose-built silicon for agents pairs naturally with the broader cost strategy I covered recently. Alibaba and DeepSeek are already slashing AI model costs to make AI more accessible. Now they’re building the hardware to match.”
More than half a million chips have been shipped. A successor is arriving in 2027, with another planned for 2028. T-Head is not hedging.
The Bottom Line
Alibaba is building AI chips specifically for AI agents. The M890 is designed for memory bandwidth and inter-model communication, not just raw compute. They’ve shipped over half a million units already. They have a multi-year roadmap. And they’re pairing the hardware with their own LLM optimized for agent workloads.
At some point, building around US export controls stops being a workaround and starts being a strategy. Alibaba appears to have crossed that line.
