Alibaba Cloud is pushing deeper into the AI hardware race with its newly revealed Zhenwu M890, an in-house AI chip built for agentic AI workloads, alongside a new large language model called Qwen3.7-Max.
For Malaysian and SEA readers, this is not just another “big tech flex” from China. AI chips directly affect the cost and availability of the tools many developers, startups, game studios, content teams, and cloud platforms rely on. If more serious AI silicon exists outside the NVIDIA ecosystem, cloud AI could become more competitive — and hopefully less painful on budgets.
What is the Zhenwu M890?
The Zhenwu M890 is based on Alibaba’s own PPU, or Parallel Processing Unit, architecture. It also includes a Transformer-focused core engine, which tells you exactly where Alibaba is aiming this thing: AI inference, agent workloads, and large model deployment.
According to the details shared, the M890 delivers 0.6 PFLOPs of FP16 compute. Alibaba claims this puts it around NVIDIA A100-class performance and roughly three times faster than NVIDIA’s Hopper H20 solution. The company also says it is three times faster than its previous-generation offerings.
On paper, the memory setup is pretty serious. The M890 comes with 144GB of HBM3 memory, up from 96GB on the older Zhenwu 810E. Interconnect bandwidth is also higher at 800GB/s, compared to 700GB/s previously.
The chip supports FP32, FP16, FP8, and FP4 formats, which matters because modern AI workloads increasingly depend on lower-precision formats to squeeze more performance and efficiency out of hardware.
Alibaba is not just selling one chip
The more interesting bit is that Alibaba is building a wider ecosystem around the M890.
The company also introduced ICN Switch 1.0, an interconnect chip offering 25.6Tb/s speeds with peer-to-peer latency under 150ns. That is meant to support massive agent concurrency — basically, many AI agents running and communicating at scale.
Alibaba is also tying this together with its Yitian Arm-based host CPU, Panmai networking cards, and a new Panjiu AL128 Supernode Server. That server is designed to pack 128 AI accelerators into a single rack, with bandwidth reaching PB/s scale.
T-Head, Alibaba’s chip division, says around 560,000 Zhenwu AI chips have already shipped, with more than 400 external customers across 20 industries.
Roadmap: V900 in 2027, J900 in 2028
Alibaba is already talking about what comes next. The Zhenwu V900 is planned for Q3 2027, with a new architecture, a claimed 3x performance jump, 216GB memory, and 1200GB/s bandwidth.
After that, the Zhenwu J900 is targeted for Q3 2028, with further architecture and performance upgrades.
That roadmap matters because AI infrastructure is becoming a long game. For SEA companies choosing cloud providers, the question is no longer just “who has GPUs now?” It is “who can keep scaling AI compute for the next three years without prices going gila?”
Qwen3.7-Max joins the party
Alibaba is also launching Qwen3.7-Max, a new LLM focused on agentic coding, reasoning, and long-running tasks. The model is described as being able to handle complex software engineering, office productivity workflows, and multi-agent operations.
Alibaba claims it can run long-horizon agentic tasks for up to 35 hours and manage more than 1,000 tool calls without performance degradation. It is also said to be optimized for agent frameworks including OpenClaw, Hermes Agent, Claude Code, Qwen Paw, and Qoder.
The model will be made available through Alibaba’s Model Studio platform for global developers.
Why SEA should pay attention
Malaysia and the wider SEA market are increasingly plugged into AI tools for coding, commerce, content, customer support, and game development. If Alibaba can offer competitive AI compute through its cloud stack, regional teams may get more options beyond the usual NVIDIA-heavy providers.
Of course, claimed performance is one thing. Real-world pricing, availability, software support, and developer experience will decide whether this becomes a serious option or just another impressive spec sheet.
Still, more competition in AI hardware is good news. If it helps bring down inference costs or improves access for SEA builders, then yes bro, we are watching.
Source: Wccftech Gaming