Alibaba's new Zhenwu M890 chip is 3x faster and aimed straight at agent workloads

Alibaba showed the Zhenwu M890 at its Cloud Summit on May 19. The chip runs about three times faster than the current 810E and ships 144 GB of GPU memory and 800 GB per second of interchip bandwidth. It’s built for multi-step agent workloads, not batch inference.

T-Head, Alibaba’s in-house silicon design subsidiary, has shipped more than 560,000 Zhenwu units to over 400 external customers across 20 industries, including automakers and financial-services firms. The M890 isn’t a one-off launch either. Alibaba paired it with a multi-year roadmap that points to a successor called the V900 in Q3 2027, with another step-up to the J900 in Q3 2028. The V900 is expected to deliver another roughly threefold gain over the M890.

What the chip actually is

The M890 is fabricated in China and aimed at filling the gap U.S. export rules have opened in the Chinese AI hardware market. Nvidia’s most capable training accelerators are blocked from sale into China, which has forced Alibaba, Baidu, and ByteDance to either route compute through offshore subsidiaries or build their own silicon. Alibaba has chosen both paths in parallel, and the M890 is its bet on the second one.

The specs are not at Nvidia H200 or B100 parity. Analysts surveyed by CNBC noted Alibaba’s chips still trail Western parts on memory capacity and bandwidth in the absolute. The interesting number isn’t 144 GB on its own. It’s 144 GB next to a credible domestic supply chain that ships in volume. A China-designed accelerator a Chinese hyperscaler can buy in quantity without an export-license review is the product, not the spec sheet.

CEO Eddie Wu told investors at the summit that Alibaba’s AI work had “moved beyond the initial investment phase” and was now progressing toward commercialization at scale. The chip and the model land together because Alibaba wants customers to think of them as a stack, not separate purchases. Cloud Intelligence Group revenue grew 38% year over year to roughly $6.1 billion last quarter, with AI-related cloud products contributing about 30% of external cloud revenue and triple-digit growth for 11 straight quarters.

Qwen3.7-Max ships next to it

Alongside the chip, Alibaba announced Qwen3.7-Max, the next flagship in the Qwen line. The pitch is engineered for advanced coding and long-running agent tasks, the same workload the M890 is sized for. The preview build has been on the Arena AI leaderboard since May 14 at an Elo score of 1,475, placing it #13 overall in Text Arena. It cracks the top 10 in math reasoning and software-and-IT prompts, and Alibaba is now ranked the #6 lab in Arena’s Text leaderboard.

That’s two notches below GPT-5.5 and Claude Opus 4.7, but it’s also the first time an Alibaba model has approached frontier math performance, and the Plus variant is already #5 globally in Vision Arena. The point of the launch isn’t to claim parity with Opus. It’s to give Alibaba customers a coherent answer to “what do I run on the new chips,” and the answer is now a Qwen flagship the company controls end to end. We covered the open-weights Qwen3.6-35B-A3B when it dropped in April, so the trajectory is clear: bigger flagship at the top of the stack, open-weights coding model in the middle, custom silicon underneath.

What’s still unconfirmed

Alibaba ducked several specifics analysts will want before the Q1 FY27 print on August 28.

• Process node. Alibaba did not disclose the fab partner or the process node for the M890. SMIC’s 7 nm equivalent is the working assumption, but the company has not confirmed it.
• Pricing or unit economics. The 560,000-unit number is cumulative across the Zhenwu line, not M890-specific. Per-chip pricing has not been published.
• Qwen3.7-Max benchmarks. Public SWE-bench, MMLU, and MATH numbers haven’t dropped. The Arena Elo is the only independent signal so far, and arena ratings reward style as much as capability.
• Export controls. The U.S. Commerce Department has not yet ruled on whether the M890 falls inside the current chip-export-control thresholds. Alibaba can sell domestically without that ruling. Selling to Southeast Asia or the Gulf is a separate question.

What this means for you

If you build software that touches Chinese cloud regions or counts Chinese enterprises among your customers, the M890 is the chip your counterparts will be running agent workloads on by next year. Even if you never touch the silicon, the model you compete against probably will. Qwen3.7-Max isn’t an Opus-killer, but it’s a serious agentic coder you can call through Alibaba Cloud or run locally via the Qwen open-weights line, and Western teams that ignore it are leaving an option on the table. For investors, BABA traded up 1.8% on the day of the announcement to $135.64, and the 41-analyst consensus 12-month target sits about 40% above that. The fundamentals look like they’re catching up to the China-AI narrative, not running ahead of it. Watch the V900 disclosure cadence over the next 12 months. If Alibaba is willing to publish a roadmap that goes out to 2028, that’s a company telling you it intends to be in the silicon business through the next administration’s export-control cycle, whoever wins.