Alibaba rolls out its Zhenwu M890 AI accelerator – delivering triple the performance of its predecessor and built for the heavy demands of agentic AI workloads.
Developed by Alibaba’s chip unit T-Head, the M890 carries 144GB of GPU memory and 800GB/s of interchip bandwidth — specs aimed squarely at closing the gap with Nvidia’s export-restricted silicon. Alongside the chip, Alibaba unveiled the Panjiu AL128 server system, packing 128 M890 accelerators into a single rack, and outlined a multi-year cadence: the V900 chip in Q3 2027, promising another roughly threefold performance jump, followed by the J900 in late 2028 — explicitly modeled on Nvidia’s own annual upgrade rhythm. T-Head says it has shipped more than 560,000 Zhenwu units to date across 400 external customers and 20 industries.
The timing is the real story..
Bloomberg’s reporting places the announcement days ahead of a scheduled Trump-Xi summit, and Alibaba’s roadmap sits alongside CEO Eddie Wu’s plan to develop a model with 5 to 10 trillion parameters — a scale that would dwarf most current frontier systems — as the company targets 20 gigawatts of underlying data center capacity by 2032.
For executives assessing supply-chain and geopolitical risk in AI compute, Alibaba’s cadence commitment matters more than any single chip’s benchmark numbers — it signals China’s hyperscalers now expect to compete on innovation speed, not just cost, in a market increasingly bifurcated by export controls on both sides.
