Alibaba Elevates "Cloud + AI + Chip" Strategy as Proprietary PPU Shipments Reach Hundreds of Thousands
Alibaba has formalized a strategic integration of its AI infrastructure capabilities, positioning itself as one of China's few tech giants with full-stack AI competencies. The company has internally branded this initiative "Tongyunge" — a portmanteau combining Tongyi Laboratory (its large language model division), Alibaba Cloud, and chip subsidiary T-Head, which together form what Alibaba calls its "AI Golden Triangle."
The strategic framework was first articulated by Alibaba founder Jack Ma during an April 2025 internal exchange with the company's technology teams. According to sources familiar with the matter, Ma coined the "Tongyunge" term and emphasized that the full-stack AI capability represents both an advantage and a responsibility for Alibaba. Group CEO Eddie Wu participated in the same discussion, describing the cloud-AI-chip combination as the most critical triangular support for implementing Alibaba's technology strategy.
Alibaba has now disclosed its proprietary high-end AI chip "Zhenwu 810E," a Parallel Processing Unit (PPU) developed by T-Head. Cumulative shipments have reached hundreds of thousands of units, surpassing Cambricon Technologies and positioning T-Head in the first tier among domestic GPU manufacturers. One week prior, Alibaba decided to support T-Head's future independent listing.
The strategic consolidation comes as Alibaba's cloud business shows accelerating growth driven by AI demand, with quarterly revenue growth expected to reach 30% to 40%, and most stock price increases linked to AI developments.
Integrated Infrastructure Approach
The "Tongyunge" concept represents a departure from fragmented deployment toward a unified AI infrastructure system. "Previously we played cards separately, dealing out a straight flush one card at a time. Now what people can see is the entire straight flush being played together," said a person close to Alibaba's core management.
The integration reflects Alibaba's ambition to become foundational infrastructure for the AI era. In MiniMax's recent IPO prospectus, the AI startup disclosed planned computing power purchases from Alibaba Cloud totaling $375 million over the next three years (2026-2028). Combined with previous procurement, MiniMax's cumulative spending on Alibaba Cloud computing power approaches or exceeds Alibaba's equity investment in the four-year-old large language model company.
Alibaba Cloud Intelligent Group Senior Vice President Liu Weiguang previously indicated that AI accelerates customer cloud migration, with momentum driven not only by GPU usage or large model API calls but broader workload shifts.
Cloud serves as the starting point for "Tongyunge." Alibaba Cloud was established in 2009 after the company recognized two critical factors: the importance of software-hardware integration for cost efficiency, and the necessity of multi-chip compatibility to avoid single-vendor dependence. The 2018 establishment of T-Head represented an inevitable choice in cloud development, according to an Alibaba Cloud employee with over 10 years tenure.
Within Alibaba's technology architecture, if AI applications are plants, cloud is the soil, large models are the air, and chips are the water in the soil — all essential elements for AI applications.
Alibaba Cloud generated revenue of 39.82 billion yuan (US$5.5 billion) in Q3 2025, up 34% year-over-year, while revenue from external customers rose 29% to 20.8 billion yuan (US$2.9 billion). The company is reportedly considering increasing its three-year investment in AI infrastructure and cloud computing from 380 billion yuan (US$52.6 billion) to 480 billion yuan (US$66.4 billion).
T-Head's Commercial Expansion
In early 2025, T-Head's proprietary AI chip "Zhenwu 810E" entered scaled commercialization. The chip serves inference computing demands on Alibaba Cloud while T-Head launched small-scale resale operations in March 2025, marking a transition from internal use to external commercial operations.
The Zhenwu PPU has served over 400 customers including State Grid Corporation of China, Chinese Academy of Sciences, Xpeng, and Weibo. Most customers accessed the chips through Alibaba Cloud's computing services. Alibaba Cloud remains T-Head's most important customer.
T-Head is actively expanding external markets. In 2025, the company secured two major external orders — both Xpeng and BYD ordered more than 10,000 PPU units. For 2026, intelligent driving, embodied intelligence, and AI training and inference represent key expansion directions for T-Head's PPU.
T-Head initiated development of the Zhenwu PPU series in 2020, approximately two years after its establishment. China's domestic general-purpose GPU "Four Dragons" — Biren Technology, Moore Threads, MetaX Integrated Circuits, and Iluvatar CoreX — were all founded around 2020.
T-Head's current chip portfolio includes the Hanguang 800 AI inference chip launched in 2019, capable of processing 78,000 images per second and now deployed in Taobao's main search scenarios; the Yitian 710 general-purpose server chip released in 2021, widely used through Alibaba Cloud for video encoding/decoding, high-performance computing, and gaming; the Zhenwu 810E training and inference AI accelerator chip for AI training, inference, and autonomous driving, already deployed at scale for Qwen model training and inference; and the Zhenyue 510 SSD controller chip with shipments exceeding 500,000 units.
Performance Positioning and Ecosystem Competition
According to industry sources, T-Head's Zhenwu 810E performance in certain typical workloads has entered the first tier of domestic computing chips, comparable to Huawei's Ascend 910 series. In specific inference or constrained computing scenarios, its comprehensive performance can exceed Nvidia Corp.'s A800 and approach the H20, though significant generational gaps remain versus Nvidia's H100 and H200 in core metrics including general computing scale and memory bandwidth.
Industry observers note that expanding actual usage of domestically-developed chips reflects not only gradual performance improvements but pragmatic choices under supply constraints and long-term external uncertainties.
Nvidia has built a highly closed yet sticky developer ecosystem through "hardware first-mover advantage + CUDA ecosystem." Huawei Ascend has formed high barriers in government cloud and smart city scenarios through "full-stack proprietary development + phased openness."
By contrast, T-Head focuses on system-level coordination across "chip-cloud-model," leveraging deep integration with Alibaba Cloud and the Qwen large model to validate requirements in real business scenarios and drive rapid chip-software iteration.
For 2026's first half, Alibaba's computing power supply will comprise two main components: existing inventory chips adapted for inference scenarios to meet basic computing demands, and T-Head's Zhenwu series chips, which will serve as the primary force supporting Alibaba Cloud's inference computing requirements.
Future domestic GPU competition will center on who can create an ecosystem that is "useful, versatile, and easy to develop" to achieve meaningful market substitution for Nvidia.