Alibaba Zhenwu V900

Zhenwu is a family of AI accelerators designed by T-Head, Alibaba’s semiconductor subsidiary, for both training and inference.¹

Multiple generations were announced throughout 2026, though deployment has spanned years.

Zhenwu V900

Zhenwu V900 was unveiled at the 2026 Apsara Conference in Hangzhou on September 22, 2026.¹

Each V900 has:

• 3x the performance of the Zhenwu M890¹ in some unspecified dimension
• 216 GB of HBM²
  • HBM generation not disclosed
  • 1.5x the M890’s 144 GB³
• 1,200 GB/s inter-chip bandwidth¹
  • Unclear if this is unidirectional or bidirectional, per-link or aggregate
  • 1.5x the M890’s 800 GB/s³
• Native FP8 and FP4 support⁴⁵

Manufacturer and process node were not disclosed.⁶ Peak FLOPS, memory bandwidth, and power were also not disclosed at launch.

Mass production and large-scale deployment are scheduled for 1Q2027.⁴²

Scale-up and scale-out

V900 supernodes connect using T-Head’s ICN Switch interconnect chip. This provides memory semantics and unified memory addressing, allowing full-bandwidth interconnect $O(1,000)$ accelerators.⁵ T-Head says more than 1,000 V900s can work together as a single system.¹ This is probably a scale-up domain of 1,024 chips, compared to M890’s 128-way domain.³

A single cluster can scale to 500,000 V900s using Alibaba Cloud’s AI data center network architecture.⁵

Panjiu supernode

The Panjiu supernode server integrates four T-Head-made chips:⁵⁷

• Zhenwu V900 (compute)
• ICN Switch (scale-up interconnect)
• Panmai smart NIC (scale-out networking)
• Zhenyue SSD controller (storage)

It is scheduled to launch in 1Q27.⁷

Timeline

In May 2026, T-Head’s roadmap placed the V900 launch in 3Q27.³ At its September 2026 unveiling, mass production was scheduled for 1Q27.⁴ It is unclear whether the schedule moved up or the two reports describe different milestones.

Zhenwu M890

Zhenwu M890 is the V900’s predecessor, unveiled at the 2026 Alibaba Cloud Summit in May 2026.³ It has:

• 144 GB HBM
• 800 GB/s inter-chip bandwidth
• 3x the performance of the Zhenwu 810E

Lingjun supernode

The M890 supernode puts 128 accelerators in a single scale-up domain using ICN Switch 1.0, with communication latency in the hundreds of nanoseconds.³

Alibaba Cloud’s Lingjun Zhenwu M890 supernode instance (GP9A) was the first supernode platform in China to run a model with more than 2 trillion parameters.⁴ M890 supernodes have run Qwen 3.8 and Kimi K3.⁵

Zhenwu 810E

Zhenwu 810E was unveiled in January 2026, though it was running in production long before then.³ It has:

• 96 GB HBM2e
• 700 GB/s inter-chip bandwidth

Zhenwu J900

Zhenwu J900 is the next-generation accelerator after the V900.⁷

Timeline

Sources disagree on J900 timing. IT之家 reported a planned launch in March 2028.⁷ TrendForce reported 3Q28 in May 2026³ but 3Q27 in September 2026.⁴ The 3Q27 figure is probably an error since it would put the J900 two quarters behind V900 mass production.

Related CPUs

T-Head also announced two Yitian server CPUs for 2027:⁵

• Yitian 720: improves single-core performance, core density, and energy efficiency⁵
• Yitian 730: T-Head’s first fully in-house CPU microarchitecture, with up to 1.4x the single-core SPECint2017/GHz of the Yitian 710⁵

Yitian 750, built on T-Head’s second-generation in-house core, will support the ICN interconnect protocol to connect Yitian CPUs directly to Zhenwu accelerators.⁴

Footnotes

1. T-Head unveils Zhenwu V900 AI chip in Alibaba’s push to expand its AI infrastructure stack (TechNode) ↩ ↩² ↩³ ↩⁴ ↩⁵

2. Zhenwu V900: Alibaba’s most powerful chip links 500,000 units, eyes 10-trillion-parameter AI (Yahoo Tech). Specs attributed to Gao Hui, vice president of T-Head. ↩ ↩²

3. [News] Alibaba T-Head Unveils Zhenwu M890 With 3× Performance; New AI Chips Planned for 3Q27/3Q28 (TrendForce) ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

4. [News] Alibaba Unveils AI Chip Zhenwu V900 for 1Q27 Mass Production, Maps Out New Server CPUs for 3Q27 (TrendForce) ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

5. 平头哥发布中国最强AI芯片真武V900，性能提升至三倍 (Sina Tech) ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

6. Alibaba unveils AI chip and 20GW plan as CEO admits supply squeeze (Techwire Asia) ↩

7. 最强国产 AI 芯片阿里平头哥真武 V900 发布，全新磐久超节点服务器明年 Q1 上市 (IT之家) ↩ ↩² ↩³ ↩⁴