Alibaba M890 chip claims 3x Nvidia H20 performance
Key insights
- Alibaba has already shipped 560,000 Zhenwu chips to over 400 customers across 20 industries before this public announcement.
- Qwen3.7-Max supports 35-hour continuous autonomous agent operation and 1,000+ sequential tool calls without performance degradation.
- The Panjiu AL128 Supernode integrates 128 M890 accelerators in a single rack at petabyte-per-second internal bandwidth.
Why this matters
Chinese hyperscalers building credible alternatives to Nvidia Hopper-class hardware undermines the core premise of U.S. export controls as an AI capability ceiling, which changes the risk calculus for any Western company or government assuming compute access is a durable moat. The 35-hour autonomous agent runtime spec for Qwen3.7-Max sets a new public benchmark for long-horizon agentic reliability that Western frontier labs will now face pressure to match or discredit. At 560,000 units already deployed, the M890 has passed proof-of-concept into real production scale, meaning Chinese enterprises are no longer just planning around chip shortages but actively substituting around them.
Summary
Alibaba used its annual Cloud Summit on May 19 to announce two significant bets on domestic AI infrastructure: the Zhenwu M890 accelerator and the Qwen3.7-Max large language model, both positioned as direct answers to U.S. export controls that have cut Chinese firms off from Nvidia's latest hardware.
The M890 claims 3x the throughput of Alibaba's current 810E chip, ships with 144 GB of HBM3 memory and 800 GB/s interchip bandwidth, and supports the full precision stack from FP32 down to FP4. The Panjiu AL128 Supernode Server packs 128 M890 units into a single rack at petabyte-per-second bandwidth, signaling that Alibaba is building toward full-stack independence from Western silicon.
Essentially: (Alibaba, Nvidia) are now in direct hardware competition inside China, with 560,000 units already shipped to 400+ customers.
- Qwen3.7-Max sustains autonomous agent operation for up to 35 hours and handles over 1,000 sequential tool calls without degradation.
- The M890 targets Nvidia's Hopper-class accelerators, the last generation Chinese firms could legally import before tightened controls.
- 20 industries across 400+ customers already running Zhenwu units suggests real deployment volume, not vaporware.
If Alibaba's performance claims hold under independent benchmarking, the assumption that U.S. export controls would cap Chinese AI compute indefinitely is due for a serious revision.
Potential risks and opportunities
Risks
- Nvidia's China revenue, already under pressure from H20 restrictions, faces further erosion if M890 benchmarks validate at even 70% of claimed performance, accelerating customer migration away from legacy Hopper inventory.
- Western AI labs and cloud providers that have built agentic product roadmaps assuming a Chinese compute gap may face faster-than-expected competitive pressure from Alibaba Cloud customers running equivalent workloads domestically.
- U.S. policymakers could accelerate additional controls targeting HBM memory and advanced packaging, creating supply chain disruption for Korean and Taiwanese memory suppliers (SK Hynix, Samsung) currently caught between both markets.
Opportunities
- Chinese AI application developers (Baidu, ByteDance, Moonshot AI) gain a domestically sourced, high-bandwidth alternative to constrained H20 supply, potentially accelerating their own model training and inference buildouts.
- Benchmark and evaluation firms (Epoch AI, Artificial Analysis) have a clear opening to publish independent M890 vs. H100 comparisons, which would become highly cited references across both policy and investment communities.
- Sovereign AI infrastructure programs in the Middle East, Southeast Asia, and Africa that want non-U.S.-controlled compute stacks now have a credible second vendor to put into procurement conversations alongside Nvidia.
What we don't know yet
- Whether the 3x performance claim over the 810E holds on independent benchmarks against Nvidia H100 or H800, not just the internal 810E baseline Alibaba chose as the comparison point.
- Which of the 400+ customers are running M890 at scale versus pilot deployments, and whether any have published real workload performance data.
- How Alibaba sources HBM3 memory at 144 GB per chip given DRAM export restrictions targeting advanced memory for Chinese AI applications.
Originally reported by cnbc.com
Read the original article →Original headline: Alibaba Unveils Zhenwu M890 AI Chip and Qwen3.7-Max LLM at Cloud Summit — 3× H20 Performance, 144 GB HBM3, 35-Hour Autonomous Agent Operation