Alibaba introduces the Zhenwu M890, marking a strong push in China to create an alternative to NVIDIA.

      T-Head's new GPU is now subject to US export controls, coinciding with a summit between Trump and Xi regarding AI chips, and the company asserts that the domestic accelerator market in China is already in large-scale production.

      On Wednesday, Alibaba's T-Head chip division revealed comprehensive specifications for the Zhenwu M890, its latest GPU-class AI chip intended as a domestic substitute for NVIDIA's accelerators. An Alibaba executive indicated that T-Head's proprietary GPU chips have entered large-scale production. This announcement comes amid a notably active period for discussions surrounding Chinese AI chips.

      The South China Morning Post provides a detailed account of the Zhenwu M890's specifications. This chip represents T-Head's most advanced product to date and is aimed at competing with NVIDIA's H100 generation rather than the newer Blackwell series. Although independent benchmark reviews suggest there remains a significant performance gap between it and NVIDIA's flagship models, the disparity with the H100—now unavailable for Chinese customers due to US export restrictions—is narrower. This situation, where customers cannot purchase the H100 and need a credible domestic option, informs Alibaba's announcement.

      The claim of "large-scale production" is a noteworthy operational detail. Coverage from EE Times highlights how this production-line ramp is the type of disclosure Western analysts have desired from Chinese chip designers for the past two years. An Alibaba executive's willingness to make this claim publicly indicates the company is confident in its supply-chain resilience, ready to face the technical scrutiny that may ensue.

      Reports indicate that the chip is produced using process nodes that Chinese foundries can manufacture without US-controlled lithography equipment, which has been a major constraint impacting the entire Chinese domestic-chip ecosystem.

      The financial context is significant. T-Head, Alibaba's chip division officially named Pingtouge (which means "honey badger" in Chinese), was founded in 2018 and delivered its first AI chip, the Hanguang 800, in 2019, consistently functioning as an internal supply unit within Alibaba Cloud ever since. T-Head plans to pursue an IPO to support a more assertive infrastructure investment strategy, potentially putting it into direct competition with Cambricon and Huawei’s Ascend lineage in the domestic accelerator sector. Thus, the announcement of the Zhenwu M890 also serves as the operational substance for whatever competitive claims T-Head makes in its eventual prospectus.

      The competitive dynamics are often understated in media coverage. Chinese hyperscalers, foundational model labs, and AI deployment clients have been increasing their purchases of domestic AI chips through the first two quarters of 2026, even as NVIDIA's H200 has been approved for sale to a limited number of Chinese clients under the new export licensing framework. These two trends are not at odds; Chinese clients seek alternatives from both markets. With the Zhenwu M890, Alibaba aims to establish T-Head as the domestic default choice, akin to how Huawei has positioned Ascend and Cambricon has done with the Siyuan line. The market, based on current evidence, seems sufficiently large to accommodate all three at scale.

      The geopolitical context emerged earlier this month with the Trump-Xi summit in Beijing. AI policy negotiations between the US and China are now being conducted at the head-of-state level, with discussions on H200 export licensing and AI regulatory measures on the agenda. None of the H200 chips authorized for ten Chinese buyers have yet been shipped.

      In the current vacuum of procurement, Chinese customers are increasingly turning to domestic alternatives. The announcement of the Zhenwu M890 is a strategic commercial response to this situation.

      The broader trend of developing alternatives to NVIDIA extends beyond China. Takeover talks between Tenstorrent and Intel and Qualcomm in the US, along with Google's recent $25 billion TPU-cloud joint venture announcement with Blackstone, represent similar challenges in the market.

      Given current capacities, the AI compute supply curve cannot sustain hyperscaler demand solely on NVIDIA silicon. Alibaba’s positioning of T-Head represents the Chinese solution to this allocation issue.

      Regarding operational details that subsequent reports will need to address: the pricing per chip, total shipment volumes of Zhenwu M890 so far, named customers outside of Alibaba's own cloud services, the division between internal and third-party shipments, and T-Head’s IPO timeline. However, none of these specifics have been formally disclosed yet.

      What is currently available is the technical information and the production status claim. The next announcement of named customers, particularly those beyond Alibaba Cloud's infrastructure, will be a crucial indicator of T-Head's success in moving the chip beyond internal usage.