China's AI 'General Contractor' Targets $1.3B Hong Kong IPO

Jiliu Technology, a Chinese AI compute infrastructure startup founded just 3 years ago, has filed for a Hong Kong IPO with a valuation of approximately $1.27 billion. The company, which builds and manages GPU clusters for AI firms without manufacturing chips or developing its own models, saw revenue skyrocket from roughly $4 million to $72 million in under 3 years — making it one of the fastest-growing AI infrastructure plays in China.

Founded in February 2023 by Tsinghua University alumnus Hu Xiaohe, the company positions itself as China's largest independent AI compute cluster provider. Its IPO filing comes at a time when global demand for AI computing power continues to far outstrip supply.

Key Takeaways

• Valuation: ~$1.27 billion (RMB 9.2 billion) in just 3 years of operation
• Revenue growth: From ~$4 million (RMB 30 million) to ~$72 million (RMB 520 million)
• Business model: Builds, deploys, and manages GPU compute clusters for AI companies
• Differentiator: Does not manufacture chips, build cloud platforms, or develop AI models
• Founder: Hu Xiaohe, Tsinghua PhD graduate who studied under UC Berkeley's Scott Shenker
• IPO target: Hong Kong Stock Exchange

The 'General Contractor' Model for AI Infrastructure

Jiliu Technology occupies a unique niche in the AI value chain. While companies like NVIDIA design chips and firms like OpenAI and Baidu build large language models, Jiliu focuses exclusively on the infrastructure layer between them — assembling, configuring, optimizing, and managing GPU clusters at scale.

Think of it as a general contractor for AI data centers. When an AI company acquires thousands of GPUs, it still faces the enormous challenge of networking them together, ensuring efficient utilization, managing thermal loads, and optimizing workloads. This is where Jiliu steps in.

The model is particularly compelling in China's current market environment. During the height of the 'Hundred Model War' — when dozens of Chinese companies raced to build their own foundation models — many startups discovered that having the budget to buy GPUs was only half the battle. Actually building a functioning, high-performance compute cluster required deep expertise in networking, systems engineering, and distributed computing.

A Founder Built for This Moment

Hu Xiaohe's background reads like a blueprint for this exact business. He spent a full decade at Tsinghua University from 2010 to 2020, earning his undergraduate and doctoral degrees. His PhD advisor was Li Jun, a prominent researcher in network security.

During an exchange program at UC Berkeley, Hu studied under Scott Shenker, a member of the National Academy of Sciences and one of the pioneers of Software-Defined Networking (SDN). SDN fundamentally changed how networks are designed and managed by separating the control plane from the data plane — a concept directly applicable to orchestrating large GPU clusters.

This combination of network systems expertise and high-performance computing knowledge positioned Hu to recognize a gap that most AI entrepreneurs were overlooking. While the spotlight shone on model developers and chip designers, the critical middleware layer — connecting hardware to workloads efficiently — remained underserved.

Revenue Growth Signals Strong Market Demand

The company's financial trajectory tells a story of explosive demand. Growing from approximately $4 million to $72 million in annual revenue within 3 years represents roughly 18x growth — a pace that rivals some of the most successful enterprise infrastructure startups in the West.

For context, CoreWeave, the U.S.-based GPU cloud provider that went public in early 2025, followed a similar trajectory of rapid revenue scaling driven by insatiable AI compute demand. While CoreWeave operates a different model — owning and renting GPU capacity rather than building clusters for others — both companies are riding the same macro wave.

Several factors are driving Jiliu's growth:

• Surging demand for AI training compute as Chinese companies develop increasingly large foundation models
• Complexity of cluster deployment that exceeds the internal capabilities of most AI startups
• GPU scarcity due to U.S. export controls on advanced NVIDIA chips, making efficient utilization of available hardware even more critical
• Cost pressure that forces companies to maximize performance per GPU rather than simply buying more hardware
• Rising inference workloads as deployed AI applications scale to millions of users

Export Controls Create an Unexpected Tailwind

U.S. semiconductor export restrictions, which have progressively tightened since October 2022, have paradoxically strengthened Jiliu's value proposition. With Chinese companies unable to freely purchase NVIDIA's most advanced H100 and B200 GPUs, every available chip becomes more precious.

This scarcity premium means that efficient cluster management is no longer a nice-to-have — it is mission-critical. A poorly configured cluster might waste 30-40% of its theoretical compute capacity through inefficient networking, scheduling, or cooling. Jiliu's optimization services can potentially recover much of that lost performance, effectively 'creating' additional compute capacity without requiring new hardware.

The dynamic mirrors what happened in the oil industry during supply crunches: when the resource itself becomes scarce, the companies that help extract maximum value from existing supply become disproportionately important.

Risks and Challenges on the Road to IPO

Despite the compelling growth story, Jiliu faces several significant risks that investors will scrutinize:

• Customer concentration: As an infrastructure service provider, Jiliu likely depends on a small number of large AI companies for the majority of its revenue
• Geopolitical exposure: Further tightening of U.S. export controls could disrupt supply chains or limit the types of hardware Jiliu can work with
• Competitive pressure: Major cloud providers like Alibaba Cloud, Huawei Cloud, and Tencent Cloud all offer managed GPU services and could vertically integrate into Jiliu's space
• Margin sustainability: Service-based businesses in infrastructure typically face margin pressure as they scale
• Technology risk: Rapid shifts in AI hardware architectures — such as the move toward custom ASICs — could reduce demand for GPU cluster assembly

The Hong Kong IPO market itself also presents timing risks. While the exchange has been actively courting tech listings, investor sentiment toward Chinese tech companies remains volatile due to regulatory uncertainties and macroeconomic headwinds.

What This Means for the Global AI Infrastructure Market

Jiliu's emergence and rapid scaling highlight a broader trend: the AI infrastructure stack is unbundling. Just as the cloud computing era spawned specialized companies for storage, networking, security, and orchestration, the AI era is creating distinct value layers between chip fabrication and model training.

In the U.S., this trend is visible in companies like CoreWeave (GPU cloud), Lambda (AI compute), Run:ai (GPU orchestration, acquired by NVIDIA), and Together AI (inference infrastructure). Jiliu represents the Chinese equivalent of this infrastructure specialization.

For Western observers, the company offers a useful data point on the state of China's AI ecosystem. Despite export controls, Chinese companies are finding creative ways to build competitive AI infrastructure — not by designing better chips, but by extracting more performance from available hardware.

Looking Ahead: Can the Growth Sustain?

Jiliu's IPO filing will be closely watched as a barometer for AI infrastructure investment appetite in Asia. If successful, it could open the door for similar specialized infrastructure companies to access public markets.

The key question is whether the company's growth rate can sustain as China's 'Hundred Model War' consolidates. As weaker AI startups fold and the market concentrates around a few major players, the total addressable market for cluster building services could either expand (as surviving companies scale up) or contract (as large players build internal capabilities).

Hu Xiaohe's bet is clear: AI compute demand will continue growing faster than companies' ability to manage it internally. If that thesis holds, being the 'general contractor' of AI infrastructure could prove far more durable — and profitable — than building the models themselves.