Alibaba Zhenwu V900: What T-Head Announced at Apsara 2026

Alibaba’s T-Head introduced the Zhenwu V900 at the 2026 Apsara Conference as one accelerator for both high-precision model training and low-precision inference. Alibaba says it delivers three times the performance of the earlier Zhenwu M890; that figure is a company claim, not an independently verified benchmark result. 1812

V900 specifications and release timing

The announced specifications are 216 GB of memory, 1,200 GB/s of chip-to-chip bandwidth and native support for multiple numerical precisions, including FP8 and FP4. Those formats are part of T-Head’s pitch for using the same processor across training and inference workloads. Alibaba plans mass production and commercial release in the first quarter of 2027. 1819

How it fits into larger AI systems

T-Head says its ICN Switch chips can connect V900 accelerators into supernodes with unified memory addressing. Alibaba also says a V900-based cluster could support up to 500,000 accelerator cards for training and inference. That is a stated scaling capability, not confirmation that a cluster of that size has been deployed. 911

Alibaba’s existing customer figures need the same distinction. The company reported that more than 650 external customers across over 20 industries had adopted the M890 through Alibaba Cloud services. That figure does not establish V900 adoption ahead of its planned commercial release. 3119

What remains unproven

The disclosed V900 details do not provide a FLOPS figure, power consumption or process node, and the reporting cited here does not identify an independent benchmark validating the threefold gain. The announcement therefore describes Alibaba’s intended capabilities and production schedule, but does not yet support a measured comparison with rival accelerators. 612