跳到正文
原文
钛媒体· Chelsea_Sun·· 4 天前AI 评分59

阿里计划分拆芯片子公司平头哥赴港上市,能否独立定价成焦点

Alibaba’s Chip Unit Heads for Hong Kong With Cloud Still in the Way

文章要点
  • 1阿里宣布计划分拆芯片子公司平头哥赴港交所主板上市,交易完成后仍将作为集团子公司存在
  • 2估值预期分化明显,研究区间上限看至约850亿美元,另有市场人士认为仅在200亿至300亿美元
  • 3管理层称真实云负载中超60%的平头哥芯片最终服务外部商业客户,但这是渗透率口径而非直销数字
  • 4行业数据显示2026年上半年平头哥约占中国AI加速器出货量7%,居国产第二,仅次于华为昇腾
决策要点

这次分拆的实质是用上市机制倒逼市场回答一个问题,嵌在云收入里的芯片资产能否按独立芯片公司定价。

  • 影响:对阿里,分拆为芯片研发开辟独立融资渠道,缓解自由现金流由流入转为流出的压力;员工持股意在留住芯片人才;对算力采购方,芯片仍以云服务形式交付,短期难以按独立供应商评估。
  • 参考:评估国产AI加速器选型时,可把平头哥与寒武纪、海光、昆仑芯的客户集中度披露放在一起比较;若采购依赖阿里云算力,值得等招股书披露关联销售与芯片直销收入后再定长期供应商策略。
  • 依据:管理层称真实云负载中超60%芯片服务外部客户,属渗透率口径而非直销数字;约7%出货份额为2026年上半年行业数据;850亿美元及200亿至300亿美元估值均来自研究机构与市场人士说法。
正文

NextFin News — Alibaba Group said it intends to spin off T-Head, its semiconductor arm, for a main-board listing in Hong Kong. The chip unit would remain a subsidiary after the deal. The exchange filing thresholds Alibaba cited put the transaction below levels that would force a major disclosure event for the parent—an odd quiet for one of the group’s most technology-heavy assets.

Valuation talk immediately split. Research ranges have stretched from the high tens of billions of dollars into the mid-eighties at the top end, while other market participants have talked closer to $20–30 billion. The spread is not mainly about process nodes. It is about whether T-Head can be priced as a stand-alone chip vendor or only as a captive supplier wrapped inside cloud revenue.

Reports around the plan say Alibaba would first reorganize T-Head as an entity with employee ownership, then pursue an IPO. Chip designers at listed peers already hold tradable stock. T-Head’s engineers, many drawn from earlier domestic teams and from firms such as Qualcomm, AMD, and HiSilicon, have largely lived on group-level incentives. Turning them into T-Head shareholders is a retention move as much as a governance one. The sequencing suggests the harder problem is identity, not a term sheet.

T-Head did not start as a merchant GPU house. Its first AI accelerator, launched in 2019, was built for Alibaba’s own search and shopping load. The Yitian server CPU followed to cut cloud cost on video and high-performance jobs. A more general GPU effort came later. That history still shapes the business. Chips ship into Alibaba Cloud clusters and are sold onward as rented compute. Automakers training driving models and banks running document recognition sign cloud contracts, not chip purchase orders.

Management has said that in real cloud workloads, more than 60% of T-Head silicon ultimately serves external commercial customers. That is a penetration statistic, not a direct sales figure. Cambricon, Hygon, and Baidu’s Kunlunxin publish clearer customer concentration and product-sale lines. T-Head’s top-line remains folded into a broader “AI cloud and computing services” segment, which makes external revenue hard to audit.

The product map is still broad by domestic standards: Zhenwu AI accelerators with cumulative shipments in the hundreds of thousands, server CPUs, SSD controllers, and interconnect chips. Industry data for the first half of 2026 put T-Head at roughly 7% of China’s AI accelerator unit shipments—second among domestic suppliers behind Huawei’s Ascend. Inference, where customized silicon can beat general GPUs on cost, is the segment the company is built to serve. The highest-volume line historically, the Xuantie RISC-V processor IP business, sits outside the listing perimeter after an earlier group reorganization.

Alibaba has committed enormous capital to cloud and AI infrastructure—hundreds of billions of yuan over a multi-year plan, with a large share already spent. Group free cash flow has swung from solid inflow to outflow as investment accelerated, and the company has raised equity to help fund the build-out. A listed T-Head could finance research and capacity on its own balance sheet and ease some of that pressure.

Depth of integration cuts both ways. Co-design with Tongyi models and Alibaba Cloud gives T-Head real deployment volume and iteration speed. The same integration keeps delivery packaged as cloud service rather than chip sale, raises the share of internal settlement, and leaves investors asking who buys the product when the parent is not the channel. Alibaba continues to purchase other accelerators, including Huawei’s, which only sharpens the question of when T-Head becomes the default for training as well as inference.

Kunlunxin’s path is the closest comparison: heavy early reliance on Baidu, then public concentration ratios and large external tenders that let the market track the decline in parent share. T-Head has not yet put equivalent numbers on the table. Until related-party sales, direct chip revenue, and non-Alibaba channels are broken out, the valuation debate will stay wide.

A listing can reprice talent and open a funding channel. It cannot, by itself, turn a cloud cost center into a merchant semiconductor franchise. That conversion is what investors will underwrite—or refuse to underwrite—when the books finally open.

更多精彩内容,关注钛媒体微信号(ID:taimeiti),或者下载钛媒体App

来源:钛媒体 · tmtpost.com