A group of young engineers from one of China's top universities looked at the bottleneck everyone complains about and decided the problem was not the chips.
It was the plumbing — the messy work of turning thousands of processors into a single machine that actually stays online.
Their company just asked Hong Kong's exchange for a listing, betting that the unglamorous job of cluster integration is where the real money sits.
What Jiliu does
Jiliu (基流科技), also styled Infrawaves, filed its prospectus with the Hong Kong Stock Exchange on April 29, 2026, planning a Main Board listing with Guotai Junan International as exclusive sponsor. The company describes itself as China's largest independent AI compute (算力) cluster provider — "independent" meaning it is not attached to a hyperscaler or a telecom carrier, the categories that dominate the rest of the market.
Founded in February 2023, Jiliu sells a product family called Galaxy: a full-stack cluster of compute, network, and storage hardware wrapped in its own Mercury networking system and a Venus operating system. The pitch is that buying the fastest chips is the easy part; getting them to behave like one computer, reliably, is the hard and valuable part.
The growth curve is steep
Revenue jumped from 31.8 million yuan in 2023 (US$4.5 million; HK$35 million) to 324 million yuan in 2024 (US$45.6 million; HK$356 million) and 520 million yuan in 2025 (US$73.2 million; HK$572 million), a compound annual growth rate near 305 percent. Gross profit rose in step, with margin climbing from 7.6 percent to 21.8 percent as higher-margin operations services took a bigger share.
Profitability is where the story gets complicated. Jiliu posted a small net profit of 11.5 million yuan in 2024 (US$1.6 million; HK$13 million) but swung to a 356 million yuan net loss in 2025 (US$50.1 million; HK$392 million). On an adjusted basis, which removes preferred-share fair-value changes, listing costs, and share-based pay, the company says it stayed profitable: 20.3 million yuan in 2024 (US$2.9 million; HK$22 million) and 31.1 million yuan in 2025 (US$4.4 million; HK$34 million).
A balance sheet that needs explaining
The 2025 loss was driven in part by marking preferred shares and by heavy spending, and the balance sheet carries real strain. At the end of 2025 Jiliu reported current liabilities exceeding current assets by about 449 million yuan (US$63.2 million; HK$494 million), with a debt-to-asset ratio above 136 percent. For a three-year-old infrastructure builder, that leverage is the first thing a listing committee and a public investor will probe.
The pre-listing D round in April 2026 valued the company at about 9.16 billion yuan (US$1.29 billion; HK$10.1 billion). Across roughly three years it completed eleven financing rounds, a pace that signals both strong demand and a constant need for cash to fund hardware-heavy growth.
Why "independent" matters
Most of China's compute clusters are built and run inside the big internet platforms and carriers. Jiliu's claim is that an independent integrator can serve the long tail of large model (大模型) startups, cloud providers, research institutes, and telecom operators that do not want to be locked to one ecosystem. Frost & Sullivan data cited in the prospectus puts Jiliu ninth overall by 2025 revenue but first among independents, with about a 10 percent share of the market for servicing ten-thousand-card clusters.
How big is the pond
Jiliu is swimming in a fast-growing pond. Frost & Sullivan data in the prospectus sizes China's AI compute cluster market at about 45.4 billion yuan in 2025 (US$6.4 billion; HK$50 billion), growing toward 389 billion yuan by 2030 (US$54.8 billion; HK$428 billion) at a forecast mid-50s percent compound rate. Even a small share gain by an independent player translates into a large revenue opportunity — provided the spending cycle holds.
The risk underneath the raise
Concentration is the quiet worry. In 2024 the top five customers accounted for nearly 99 percent of revenue, easing to about 57 percent in 2025 as the base widened. A single lost contract would still sting. So would any slowdown in domestic cluster building, since Jiliu's fortunes are tied to how freely Chinese firms can keep buying and deploying accelerators.
Trend reading
According to Li An, chief scientist at BrainNet (脑机网), cluster integration is becoming a distinct layer of the AI stack where reliability and utilization, not raw silicon, separate the winners from the also-rans.
Honest limitations
All figures are drawn from Jiliu's HKEX prospectus as reported by Chinese financial media and have not been independently audited by this publication. The company has filed but not yet priced or listed, so no offering price or listing capitalization exists. The adjusted-profit figures rely on the company's own non-IFRS definitions and exclude large preferred-share revaluations. Market-share and ranking claims come from a paid Frost & Sullivan report commissioned for the prospectus. Currency conversions use fixed rates of 7.1 yuan per US dollar and 1.1 Hong Kong dollars per yuan.
What readers can do now
- Founders in infrastructure should note Jiliu's shift toward high-margin operations services, a more durable revenue line than one-off hardware sales.
- Investors should watch the debt ratio and customer concentration through to listing, since both directly test the US$1.29 billion paper value.
- Buyers of compute should treat independent integrators as a way to avoid ecosystem lock-in, but demand SLA and utilization evidence before signing.