China's DataCanvas Files for a Hong Kong IPO on a 221% Revenue CAGR and a Standardized 'Compute Unit'

Beijing-based artificial intelligence infrastructure provider DataCanvas (JiuZhang YunJi) filed an application for a main board listing on the Hong Kong Stock Exchange on Sept. 29, with China Galaxy International as sole sponsor. The company is positioning itself as the first "NeoCloud" — a compute-cloud specialist, rather than a server reseller or data-centre landlord — to list in Hong Kong.

The numbers in the filing are the pitch. Revenue went from 106 million yuan in 2023 to 395 million yuan in 2024 and 1.098 billion yuan in 2025, a compound annual growth rate of 221.5% and roughly a tenfold increase in two years. In the first half of 2026 revenue reached 588 million yuan, up 92.0% from a year earlier. Client count for the higher-margin compute cloud business moved from two enterprise customers in 2024 to 75 in 2025 and 84 in the first half of 2026; cloud services accounted for 26.1% of first-half 2026 revenue.

DataCanvas runs two businesses: turnkey compute-centre infrastructure and on-demand compute cloud. The company's most distinctive claim is a unit of account. It defines "1 DCU" as 312 teraflops of FP16 compute sustained for one hour, which consultancy China Insights Consultancy describes as the Chinese compute industry's first quantifiable billing unit. The idea is to make capacity from different chip vendors and server architectures comparable, so it can be metered and sold the way electricity is. On top of that standard it sells virtual container services billed by actual consumption and dedicated cluster containers.

Delivery runs through what the company calls AI factories — a training factory, where the filing claims 54.1% model FLOP utilization on large-scale training, and a token inference factory that abstracts scheduling and model deployment behind standard APIs so customers do not have to stand up their own clusters. The orchestration layer is Alaya NeW OS, which the company says is the first domestic compute operating system certified by the China Academy of Information and Communications Technology across all four domains of scheduling, model training, model inference and data processing. Its serverless architecture places Kunlunxin, Hygon and Iluvatar cores on the same scheduling plane as international accelerators.

Losses are narrowing on an adjusted basis. Non-IFRS adjusted net loss improved from 91.27 million yuan in 2025 to 14.88 million yuan in the first half of 2026. The company reports 125 registered patents and 117 software copyrights, and says it has published research at ICLR and NeurIPS.

Capital efficiency is the other half of the story. DataCanvas uses a joint-development model in which infrastructure partners fund part of the underlying build while the company contributes data-centre architecture, systems integration, scheduling and operations, alongside selective self-build. It says it controls more than 29,000 petaflops of capacity across seven Chinese provinces, locked in through irrevocable agreements of three to eight years with no unilateral early-termination clause — a structure that keeps the balance sheet lighter than owning the metal outright but commits it to long-dated payments.

Notably, the company says it will not commercialize a proprietary foundation model and will not build AI consumer applications, arguing that a neutral infrastructure position means it never competes with its own customers — a pointed contrast with the compute clouds operated by model labs.

The bet underneath the filing is that compute in China is turning from a scarce engineering service into a metered utility, and that the winner is whoever defines the meter. That is a plausible thesis for a market where chip supply is fragmented across domestic and foreign silicon and buyers want to avoid lock-in. It is also a thesis that depends on heavy, long-dated commitments landing before demand does. The coming test is whether public-market investors price a compute utility like infrastructure — or like the capital-intensive, thin-margin business it can easily become.