T-Head is coming for the enterprise AI chip market and it wants its own IPO to do it

Alibaba's chip unit just got a lot more ambitious than filling an Nvidia-shaped gap

Alibaba's semiconductor subsidiary T-Head has unveiled the Zhenwu M890, a new AI accelerator built for agentic workloads, alongside a roadmap stretching to 2028 and plans to eventually list the unit publicly. Taken together, the announcements mark a shift in how Alibaba is positioning T-Head: less internal chip supplier, more independent platform contender.

The M890 is a meaningful hardware step. The chip carries 144GB of HBM3 memory, up from 96GB in its predecessor the Zhenwu 810E, with interchip bandwidth of 800GB per second. It handles both training and inference on a single chip and is designed specifically around the demands of AI agent workloads, where models must hold long context windows and coordinate with one another in real time.

Alibaba is also accelerating its upgrade cadence, now planning annual Zhenwu releases, roughly aligning itself with Nvidia's product rhythm. The V900 follows in Q3 2027, the J900 in Q3 2028, each expected to deliver roughly another threefold performance jump.

The specs matter, but the signal matters more. T-Head is not building a fallback option for customers who cannot access Nvidia hardware. It is building a product line designed to compete on its own terms, with a roadmap long enough to make procurement decisions around.

The IPO angle changes the calculus

Bloomberg reported in January that Alibaba is preparing to restructure T-Head as a partially employee-owned business ahead of a potential IPO, timing unclear. The M890 launch, read in that context, is as much a commercial prospectus as it is a product announcement. T-Head needs to demonstrate that it has real external traction, a credible roadmap, and a differentiated position in a market that already includes Huawei's Ascend line and Cambricon's Siyuan chips.

On traction, the numbers are solid. T-Head has shipped more than 560,000 Zhenwu units to over 400 customers across 20 industries, including automakers and financial services firms. That is production-scale deployment, not pilot hardware. An analyst at SemiAnalysis told CNBC the chips are "becoming one of the more popular platforms among Chinese domestic AI hardware chips", an independent data point that T-Head will be happy to have on record ahead of any investor roadshow.

A stack, not just a chip

What makes the M890 announcement harder to dismiss is the software layer sitting alongside it. Alibaba announced Qwen 3.7-Max simultaneously, a large language model optimised for coding and long-running agent tasks, capable of operating for up to 35 hours without performance degradation. The chip and the model are designed for the same workload class. That is deliberate.

Alibaba is positioning T-Head silicon, Qwen models, and Bailian cloud delivery as a closed loop, reducing enterprise dependency on any single external vendor at any layer of the stack. The Panjiu AL128, a rack-scale system packaging 128 M890 accelerators, is available now through Alibaba Cloud's Bailian platform to Chinese enterprise customers. Broader availability timelines have not been confirmed.

Half a million chips shipped. A roadmap to 2028. An IPO in the works. T-Head is not positioning itself as a domestic stopgap, it is positioning itself as a business.