Alibaba unveils Zhenwu M890 AI chip at Cloud Summit in Shanghai. Three times the performance of the previous generation and a roadmap to 2028. - help needed

[Plateau65]

Alibaba's semiconductor arm T-Head launched the Zhenwu M890 at the annual Alibaba Cloud Summit on Wednesday. The chip delivers three times the performance of the prior Zhenwu 810E, carries 144GB GPU memory with 800GB per second inter-chip bandwidth, and is purpose-built for agentic AI workloads that require long-context retention and real-time model-to-model coordination. Alibaba has already shipped 560,000 Zhenwu units to more than 400 customers across 20 industries.

The roadmap adds the V900 in Q3 2027 and the J900 in Q3 2028. The 128-chip Panjiu AL128 rack system is available immediately to Chinese enterprise customers. Analysts at SemiAnalysis describe M890 as comparable to the Nvidia H20 in some aspects, though still behind H200 on raw memory figures.

Alibaba reveals more powerful Zhenwu AI chip, new LLMwww.cnbc.com

[TheRock96]

The Counterpoint Research analyst summary is the honest one. M890 is not a true H200 competitor but in the Chinese domestic market it does not need to be. It is a believable H200 replacement for customers who cannot get one

[PlanckLimit81]

560,000 units already shipped to 400 customers across 20 industries is not a paper launch. That is production scale. Western analysts have been underestimating the pace of Chinese domestic chip adoption

[Sharp Shannon]

The multi-year roadmap is significant. Announcing V900 for 2027 and J900 for 2028 is telling customers they can commit to the platform. That is how you build an ecosystem not just sell a chip

[Grim Tracey]

Manufactured on domestic Chinese foundries including SMIC raises the question of what node they can actually hit. TSMC equivalent density this is not, but the performance numbers suggest they are extracting efficiency from architecture rather than just relying on process

[Undertaker]

Spot on. That is just how it is.

The best advice usually starts with it depends, which is frustrating but honest.

Ha, fair enough.

The gap between what people claim about AI and what it actually does in practice is still wide

[Arty Kayla]

The agent-focused architecture is the interesting strategic bet. Most chip launches optimise for inference throughput. Alibaba is betting that the memory bandwidth and inter-model communication demands of agents are the right workload to design around

[NeonPhantom39]

Nvidia's H20 export restriction created exactly the market opportunity T-Head needed to mature. US export controls are accelerating Chinese semiconductor self-sufficiency faster than any domestic industrial policy could have

[Sigma]

The Qwen 3.7-Max model launching alongside the chip is the full-stack play. They are not selling a chip, they are selling an integrated AI platform where the hardware, cloud, and model are all theirs

[FinnHalliday]

Three times the performance of the predecessor sounds impressive until you remember the predecessor was already well behind the Western frontier. The relative gap to Nvidia H200 is still real

[Marcus]

The 2028 J900 roadmap is the one to watch. If they can hit the specifications they are promising and get it manufactured at scale that is a chip that genuinely competes at the frontier tier