At the Yunqi Conference in Hangzhou on May 20, Alibaba unveiled three new products: a chip, a model, and a server rack. But the highlight was a 35-hour demonstration that defied conventional logic.
A counterintuitive demo
Alibaba placed its newly released Qwen3.7-Max model on the brand-new Zhenwu M890 chip and tasked it with an audacious job: writing its own low-level kernel driver to optimize the Extend Attention operator from scratch. The chip had no external documentation.
After 35 hours, the results were:
- Tool calls: 1,158
- Kernel evaluations: 432
- Architecture redesigns: 5 rounds
- Final speedup (geometric mean): 10×
In plain terms, the model figured out how to write drivers for a chip it had never seen, boosting performance tenfold. Such a task typically takes a team of chip engineers one to two months.
The Zhenwu M890 chip
Key specs:
- HBM3 memory: 144 GB (50% more than the previous Zhenwu 810E)
- Interconnect bandwidth: 800 GB/s
- Overall compute: 3× the previous generation
The 144 GB memory positions the M890 between Nvidia's H200 (141 GB) and B200 (192 GB), making it suitable for large model inference and agentic workloads requiring long contexts and multiple steps.
Alibaba also disclosed that it has shipped 560,000 Zhenwu chips to over 400 customers across 20 industries, indicating production-scale deployment.
The Qwen3.7-Max model
Key features:
- Context window of 1 million tokens (up from 256,000 in Qwen3.6-Max-Preview)
- Optimized for long-duration tasks and code
- Co-designed with Zhenwu M890
While not the longest context window available, 1 million tokens is sufficient for agentic tasks that run for hours without losing context.
Why this matters
Three factors converge:
First, Alibaba now offers a full-stack solution with its own chip, model, and server rack (Panjiu AL128), covering training to inference.
Second, the model writing its own driver is a clever demonstration that tells customers the chip is capable and the model is intelligent. It also reduces the need for Alibaba to hire low-level engineers.
Third, 560,000 chips shipped is not a PowerPoint promise.
Together, these points suggest Alibaba is building a Chinese version of Nvidia plus OpenAI: its own chips, models, and integrated solutions, funded by its own revenue and customer base.
How it compares to Nvidia
| Zhenwu M890 | Nvidia H200 | Nvidia B200 | |
|---|---|---|---|
| Memory | 144 GB HBM3 | 141 GB HBM3e | 192 GB HBM3e |
| Interconnect | 800 GB/s | NVLink 900 GB/s | NVLink 1.8 TB/s |
| Ecosystem | Building CUDA alternative | CUDA | CUDA |
Hardware specs are competitive with the H200, but the software ecosystem gap remains. Qwen3.7-Max's ability to write its own kernels could help bridge that gap: the model acts as a super-compiler, generating drivers as needed. While not a complete solution, it offers a potential path for Chinese firms to bypass CUDA.
What to watch
Two things:
First, how much Zhenwu compute Alibaba Cloud opens to external customers in the second half of the year. If most of the 560,000 chips remain internal, the offering is for self-use; if over 30% is external, it could reshape China's cloud market.
Second, whether Qwen3.7-Max's driver-writing capability can be generalized to other chips. If the model can adapt to any chip, it would become industry common sense that chipmakers need their own LLM teams, and vice versa.
Alibaba has staked a position on both fronts.
Sources: Alibaba unveils Zhenwu M890 chip and Qwen3.7-Max LLM (Let's Data Science); Qwen3.7-Max Wrote Its Own Chip's Software in 35-Hour Run (TechTimes); Alibaba reveals more powerful Zhenwu AI chip, CocoLoop, new LLM (CNBC); Alibaba Bets Big on AI Agents With New Zhenwu M890 Processor (SQ Magazine)