Alibaba's T-Head Unveils Zhenwu V900: Triple Performance Leap, Mass Production in 2027

On September 22, 2026, Alibaba's T-Head officially unveiled the unified training-and-inference AI chip Zhenwu V900 at the Apsara Conference 2026 in Hangzhou. Its compute power is three times that of the previous-generation Zhenwu M890. It features 216GB high-bandwidth memory, 1,200GB/s inter-chip interconnect bandwidth, and native support for FP8/FP4 low-precision formats. A single cluster built on V900 can scale to as many as 500,000 cards, supporting training and inference for frontier large models. This is currently the highest-performance commercial AI chip with publicly disclosed specifications in China.

The full-stack supernode server Panjiu, equipped with V900 as well as the ICN Switch intelligent switching chip, Panmai smart NIC, and Zhenyue SSD controller chip, is scheduled for official mass production and market launch in Q1 2027. What was unveiled today are specifications and a timetable, not a product available for immediate purchase.

Full-stack synergy: V900 is not a single-chip story

The launch of Zhenwu V900 is not isolated. T-Head simultaneously showcased a complete integrated compute-storage-network hardware stack: the ICN Switch chip handles high-speed interconnect at the 10,000-card level, with its 1,200GB/s inter-chip interconnect bandwidth about 50% higher than the previous generation; the Panmai smart NIC and Zhenyue SSD controller chip handle network and storage acceleration, respectively. This full-stack in-house development approach means Alibaba Cloud's technical control over computing infrastructure has become further concentrated—and the room for coordinated optimization across components has expanded accordingly.

At the conference, Alibaba Group CEO Eddie Wu said that the M890 AI supernode has been deployed at scale in Alibaba Cloud data centers this quarter and is already capable of supporting inference for large models with over 2 trillion parameters. Alibaba Cloud's Lingjun M890 supernode instance GP9A is already available to external customers, making it China's first supernode-form computing power to successfully run a large model with over 2 trillion parameters in commercial operation. The V900 launch, building on M890's scaled deployment, publicly sets a performance benchmark for the next-generation product—this cadence reduces the technical risk of "launch first, validate later," and for V900's claimed 3x compute increase, there is at least a mass-produced M890 as a traceable comparison baseline.

The procurement timing question behind 650 enterprise customers

T-Head's AI chip product line has now accumulated more than 650 enterprise customers, covering fields such as autonomous driving, finance, large models, embodied intelligence, energy, and manufacturing. Eddie Wu expects that as T-Head's chip product line matures and customers adopt it widely, annual shipments will increase substantially. At present, the main products used by these 650 customers are still M890 or earlier-generation products, and V900 will not enter these customers' data centers until Q1 2027.

For enterprise buyers, this means Q4 2026 to early 2027 remains an M890 window. If computing capacity expansion plans can wait, purchasing after V900 enters mass production and passes initial stability validation is the more prudent choice. For users urgently needing to scale training clusters, M890's currently validated inference capability for 2T-parameter models provides a sufficiently solid basis for real-world deployment.

Positioning against the domestic competitive landscape

The industry backdrop to V900's launch is that the competitive landscape of China's AI chip market has undergone visible structural changes over the past year. According to industry analysis data, in China's AI chip market in 2025, Huawei Ascend led with about 49% market share, while Alibaba T-Head ranked second with about 16%. Entering 2026, Huawei Ascend's share adjusted to about 37%, while Cambricon ranked first among specialized AI chip vendors with 8%. Overall competition intensified and shares were redistributed, but the top-tier landscape remained relatively stable.

Based on publicly available information, the Huawei Ascend series has planned three generations—950, 960, and 970—covering 2026 to 2028, with computing power commitments continuing to double with each iteration. Cambricon, meanwhile, demonstrated the explosive potential of another path—in 2025, Cambricon's revenue exceeded RMB 6.497 billion, a year-on-year increase of more than 450%, achieving its first profit since going public. Competition among domestic AI chips has moved from "whether we have them" to "whose are better to use."

T-Head's differentiator is ecosystem integration: V900 is not a standalone chip but is bundled into Alibaba Cloud's complete computing product system, from chips and full-stack supernode servers to cloud instances, forming a coordinated and optimized commercial path. This contrasts clearly with Cambricon's strategy, which mainly targets third-party integrators, and Huawei Ascend's focus on government and carrier markets, forming a clear divergence in approach.

J900 in 2028: The real bet on in-house architecture

At the conference, T-Head also previewed the next-generation chip Zhenwu J900, which will feature an in-house parallel computing architecture, positioned as a "leapfrog innovation," and is scheduled for launch in March 2028. Compared with V900, J900 represents a deeper level of architectural autonomy—shifting from engineering optimization of existing architectures to redesigning parallel computing logic from the ground up. This is also the most ambitious step on T-Head's roadmap to date.

With the cadence of "V900 mass production in Q1 2027 and J900 launch in March 2028," T-Head is roughly maintaining an annual generational iteration frequency. This cadence can be compared in timing to NVIDIA's transition cycle from the H series to the B series.