According to people familiar with the matter, Alibaba Group finalized a decision on January 22 to support its semiconductor subsidiary T-Head in pursuing an independent public listing. Founded in 2018, T-Head is a wholly owned Alibaba subsidiary.
Over the past eight years, the company has developed a portfolio of chips spanning computing, storage, and networking. Some of these products rank among the top performers in their respective categories. Sources told 36Kr that T-Head’s chips have become an important component of China’s artificial intelligence computing market, though no independent market share data was disclosed.
Chip design has long been regarded as one of the most technically demanding fields in the technology industry. Before establishing T-Head, Alibaba had little experience in hardware R&D, and no Chinese internet company had previously attempted large-scale in-house chip design. Internally, the technical difficulty was often compared to Alibaba’s early push into cloud computing, which required years of sustained investment before becoming commercially viable.
Despite the odds, Alibaba maintained its commitment to semiconductors, mirroring the long-term approach it took with Alibaba Cloud.
From the outset, T-Head’s strategy was tightly aligned with cloud data centers, focusing on chips designed for Alibaba Cloud’s specific workloads. This approach contrasted with independent chipmakers, which typically build standardized products for a broad customer base. While more constrained in scope, the strategy allowed T-Head to scale quickly within the semiconductor value chain.
Traditional chip companies often spend months or years gathering customer requirements before beginning design, a process that can delay commercialization. T-Head, backed by Alibaba Cloud, had direct access to real-world usage data and computing demands, significantly shortening the design-to-deployment cycle.
In 2019, just one year after its founding, T-Head released its first chip, the Hanguang 800, an AI inference processor based on a self-developed architecture. The chip achieved an inference performance of 78,563 images per second, setting performance and energy-efficiency records in its category at the time, according to Alibaba. It was later deployed in Alibaba’s search infrastructure during the Singles’ Day shopping festival.
That same year, T-Head began work on a more complex project: a general-purpose CPU. At the 2021 Apsara Conference, Alibaba unveiled the Yitian 710, its first general server processor. The company said the chip outperformed industry benchmarks by 20% while delivering more than 50% better energy efficiency, though it did not specify which benchmarks were used.
General-purpose CPUs are widely regarded as one of the most challenging areas of semiconductor design, with only a small number of companies worldwide capable of developing them at scale. Alibaba became the first Chinese internet company to complete the full CPU R&D and design process.
Today, Yitian 710 is used across Alibaba Cloud services, including video encoding and decoding, high-performance computing, and gaming. In the US, Amazon has followed a similar path, with EC2 instances powered by its in-house Graviton processors becoming a core offering of AWS. Together, the two companies demonstrated that tightly integrated, cloud-driven chip development can be commercially viable.
That model has since been adopted by other Chinese internet companies.
Around 2020, as the previous deep learning boom began to slow, a new wave of Chinese chip startups emerged, including Biren Technology, Moore Threads, and MetaX. During the same period, T-Head quietly began developing a general-purpose PPU accelerator card.
By late 2022 and early 2023, the project had reportedly completed development and scenario validation but remained undisclosed publicly.
In September 2025, US media outlet The Information reported that the first-generation PPU card delivered performance comparable to Nvidia’s H20, while an upgraded version exceeded the A100. According to the report, Alibaba had already used the chip to train small and midsized AI models internally.
Later that month, specifications for the chip were disclosed during a broadcast on state-run CCTV. The program said the card features 96 gigabytes of HBM2e memory, 700 gigabytes per second of interconnect bandwidth, a PCIe 5.0 ×16 interface, and 400 watts of power consumption. The specifications exceeded those of Nvidia’s A800 and most domestically produced GPUs, and were broadly comparable to the H20 on paper.
Beyond raw specifications, the PPU card’s main advantage is said to be usability. Stable performance and relatively straightforward deployment have contributed to positive feedback from the industry.
Over the past year, demand reportedly exceeded supply. Industry insiders said T-Head’s PPU card became one of China’s bestselling domestically designed GPUs in 2025.
T-Head has continued to expand beyond compute chips. In 2023, it launched its first storage product, the Zhenyue 510 SSD controller, which Alibaba positioned as a competitor to Samsung’s offerings and designed to meet the low-latency, high-bandwidth requirements of AI training and inference.
The company is also expected to release a networking chip this year, a move that would place it among a small group of Chinese firms offering a full data center chip stack. On the device side, its Yuzhen series has shipped hundreds of millions of units, according to Alibaba.
Despite its low public profile, T-Head’s full-stack semiconductor capabilities may represent one of Alibaba’s most overlooked assets in the AI race. While Alibaba Cloud’s infrastructure and its open-source Qwen model have drawn attention, T-Head fills a remaining gap in Alibaba’s broader AI strategy.
The timing also coincides with a surge in investor interest in Chinese chipmakers, which has driven a flurry of public listings.
In August, Cambricon Technologies, one of the few semiconductor firms listed on China’s A-share market, briefly surpassed Kweichow Moutai in intraday market capitalization, despite reporting revenue of RMB 4.6 billion (USD 644 million) in the first half of 2025. Its valuation has since stabilized at around RMB 600 billion (USD 84 billion).
That rally helped trigger a broader rush to market. Since late last year, several Chinese GPU companies have gone public in quick succession, further inflaming an already overheated semiconductor sector:
- Moore Threads became the fastest company to receive IPO approval on the STAR Market, completing the process in 88 days before listing on December 5, 2025. It is now valued at about RMB 300 billion (USD 42 billion), after reporting RMB 780 million (USD 109.2 million) in revenue over the first nine months of 2025.
- MetaX followed 12 days later, debuting on the STAR Market with shares rising nearly sevenfold on its first day of trading. The company now has a market capitalization exceeding RMB 240 billion (USD 33.6 billion) and reported nine-month revenue of RMB 1.24 billion (USD 173.6 million).
- On January 2, Biren Technology listed in Hong Kong, becoming the first domestic GPU company to do so. Its market value now exceeds RMB 70 billion (USD 9.8 billion), despite half-year revenue of RMB 58.9 million (USD 8.2 million).
- A week later, Iluvatar CoreX also debuted in Hong Kong, reaching a valuation of roughly RMB 40 billion (USD 5.6 billion) on first-half 2025 revenue of RMB 324 million (USD 45.4 million).
- Another contender, Enflame Technology, has completed IPO counseling and is preparing a STAR Market listing. Its prospectus shows half-year revenue of RMB 2.83 billion (USD 396.2 million).
Together, these five companies command a combined market capitalization of about RMB 1.25 trillion (USD 175 billion), roughly 4% of Nvidia’s current valuation.
T-Head, by contrast, operates across a broader range of computing segments, giving it a stronger base to deliver end-to-end solutions. Already positioned as a market leader in several categories, its eventual listing is likely to further influence how investors assess China’s domestic semiconductor sector.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Qiu Xiaofen for 36Kr.