AiCybr

Alibaba Zhenwu V900 AI Chip: 216GB Memory, 1.2TB/s Interconnect and Q1 2027 Roadmap

By AiCybr Editorial · Published September 24, 2026

Alibaba unveiled the Zhenwu V900 AI accelerator at its Apsara Conference on September 22, 2026, positioning the T-Head-designed processor for large-scale AI training and inference. Alibaba says the V900 delivers three times the performance of the Zhenwu M890 and can scale to clusters containing as many as 500,000 accelerators.

Technical details reported from the conference list 216GB of memory, 1,200GB/s (1.2TB/s) chip-to-chip bandwidth, and native support for FP8 and FP4 low-precision formats. Commercial release and mass production are planned for the first quarter of 2027.

The V900 is part of a broader Alibaba infrastructure roadmap spanning accelerators, CPUs, networking, storage, cloud capacity and future Qwen models. The company is also targeting more than 20GW of global Alibaba Cloud data-center capacity by 2032.

Zhenwu V900 specifications announced so far

Specification Zhenwu V900
Developer T-Head / Alibaba
Workloads AI training and inference
Memory 216GB
Chip-to-chip bandwidth 1,200GB/s
Low-precision formats FP8, FP4
Claimed performance 3× Zhenwu M890
Maximum stated cluster scale Up to 500,000 cards
Commercial timing Q1 2027 roadmap

Alibaba's official keynote confirms the threefold M890 comparison and 500,000-card cluster target. The 216GB memory, 1,200GB/s interconnect and FP8/FP4 specifications were reported from T-Head's conference disclosures by current industry coverage.

The company has not published a standardized independent benchmark behind the 3× performance figure. That number is therefore best treated as Alibaba's generational comparison until production hardware can be tested on disclosed workloads.

216GB memory targets frontier-model workloads

Memory capacity is one of the V900's most consequential published specifications. At 216GB per accelerator, the chip is designed to keep larger portions of model weights, activations and inference state close to compute while reducing the amount of model partitioning required across devices.

The capacity figure alone cannot establish real training or inference throughput. Effective performance also depends on memory bandwidth, compute throughput, interconnect topology, software kernels, precision, batch size and model architecture.

Alibaba presents the V900 as one component of a larger compute system. T-Head's surrounding stack includes Alibaba's ICN interconnect technology and a supernode architecture intended to connect large numbers of accelerators for frontier-model workloads.

1.2TB/s chip-to-chip bandwidth and 500,000-card clusters

The reported 1,200GB/s chip-to-chip bandwidth is aimed at scaling the accelerator beyond a single device. Alibaba says a V900-based cluster can support up to 500,000 cards for frontier-model training and inference.

The 500,000-card figure is the architecture's stated maximum scale. Alibaba's current M890 infrastructure is already being deployed commercially, while V900 remains on the 2027 roadmap.

At the conference, Alibaba described end-to-end co-optimization across chips, servers, supernodes, networks, models and inference engines. This approach targets aggregate token throughput and utilization across the company's full AI infrastructure stack.

FP8 and FP4 support

V900 natively supports FP8 and FP4, according to conference reporting. These lower-precision formats can reduce memory traffic and increase effective compute throughput when models and kernels can operate at reduced precision without unacceptable accuracy loss.

The practical benefit varies by workload. Training, prefill and decode can stress different parts of an accelerator, and application-level performance depends on the complete hardware and software path.

Q1 2027 is the commercial target

Current reporting places V900 mass production and commercial release in Q1 2027. Alibaba has announced the hardware and roadmap specifications; production availability remains a future milestone.

Alibaba's September 22 keynote says its existing T-Head product lines are already deployed across customers and that the company expects annual AI-chip shipments to grow significantly. V900 pricing was not provided in the keynote.

For buyers and infrastructure teams, the next useful evidence will be production availability, cloud-instance configurations, measured throughput, power characteristics and software compatibility on shipping V900 systems.

Alibaba's full-stack AI infrastructure roadmap

V900 sits inside a larger strategy that combines Alibaba's own silicon with cloud infrastructure and Qwen models. CEO Eddie Wu said Alibaba Cloud plans to exceed 20GW of global data-center capacity by 2032 as demand for AI compute expands.

Alibaba also confirmed that Qwen 4 is in training and outlined plans for future models at substantially larger scale. Those model and infrastructure targets remain roadmap statements.

The company is simultaneously developing server CPUs, networking and storage components. The goal is a vertically integrated stack in which models, inference engines, accelerators, interconnects and cloud systems can be optimized together.

How to interpret the 3× performance claim

Alibaba describes the V900 as delivering three times the performance of the M890. The official keynote does not publish a workload, precision, power envelope or standardized benchmark for that comparison.

The available evidence is insufficient for a direct performance comparison with NVIDIA, AMD, Huawei or other accelerator families. Memory capacity and interconnect bandwidth are useful architectural inputs; cross-vendor performance requires matched workloads, precision, software stacks and system configurations.

V900 represents Alibaba's next major accelerator generation, with a large memory configuration and scale-out architecture aimed at frontier training and inference. Production benchmarks will determine how those specifications translate into throughput and efficiency.

Bottom line

Zhenwu V900 expands Alibaba's in-house AI compute roadmap with 216GB memory, 1.2TB/s chip-to-chip bandwidth, FP8/FP4 support and a stated path to very large clusters. Alibaba's official generational claim is 3× M890 performance, while commercial availability is targeted for Q1 2027.

The announcement establishes an infrastructure roadmap built around V900 supernodes, a 20GW-plus cloud-capacity target and future Qwen training. Its competitive position will become clearer when production V900 systems provide measured workload, power and software data.

Sources

Continue reading

Artificial Intelligence Dnotitia VDPU: Vector-Search ASIC Enters Silicon Testing for AI Retrieval

Dnotitia's first VDPU ASIC samples are back from fabrication, with Q4 2026 silicon evaluations planned after FPGA tests reported up to 5.77x CPU-server vector-search throughput.

AI Infrastructure NVIDIA and MediaTek Expand NVLink Fusion Partnership for Custom AI Chips

NVIDIA and MediaTek are expanding their partnership across rack-scale AI infrastructure, local AI and automotive. MediaTek will use NVLink Fusion as a platform for custom XPUs, while NVIDIA is investing $3.5 billion in MediaTek convertible bonds.

Hardware CXMT Says Fifth-Generation DRAM Platform Is in Mass Production, Shows 24Gb LPDDR5X

CXMT says its fifth-generation DRAM technology platform has entered mass production, using quadruple patterning to reach 11.95nm feature spacing and support denser memory products. The company also introduced new 24Gb LPDDR5X dies.