A significant early-stage funding round for an AI infrastructure startup signals a market shift from the brute-force expense of training models to the economics of running them. On May 20, AI infrastructure provider Qijing Technology announced it closed a Pre-A round of several hundred million yuan, co-led by StarLink Capital and Huakong Technology, with existing shareholder Hillhouse Capital also participating.
The large financing in a cautious venture market shows that investor focus is moving from the "bigger is better" model race to the practical challenges of commercial deployment. The market is now rewarding companies that can solve the high and often inefficient costs associated with using AI models at scale, a process known as inference.
Qijing’s core thesis is that the industry’s focus on Model-as-a-Service (MaaS) is misplaced. Instead, it offers Token-as-a-Service (TaaS), directly linking the unit of AI output to cost. The company targets a critical inefficiency in AI deployment: hardware utilization. Traditional inference methods heavily rely on expensive GPU memory, leaving much of the system’s CPU and standard RAM idle, with overall hardware utilization often below 20 percent. Qijing's "Liuhe" architecture and "Yuebing" technology aim to solve this by redesigning how KV Cache, a key component in AI processing, is managed, thereby reducing dependence on costly GPUs. The company already provides inference services for major models like Zhipu's GLM, handling nearly a trillion tokens daily.
This focus on efficiency arrives as the AI industry confronts the hidden costs of scaling. While large language models (LLMs) trained on trillions of parameters are powerful, they are expensive to run and can be inefficient for many tasks. Some businesses are finding that smaller, more compact models can offer a better balance of performance and cost, responding five to 10 times faster than larger models for certain applications. This has created a clear commercial opportunity for infrastructure providers who can optimize the total cost of ownership, moving beyond upfront model costs to include ongoing expenses like maintenance, monitoring, and energy consumption.
The New Competitive Battlefield
The need for inference optimization is intensified by a supply-demand imbalance for computing power. In March 2026, major providers like Tencent Cloud, Alibaba Cloud, and Baidu Smart Cloud raised prices for AI computing services, with some model costs increasing by over 460 percent. This environment creates a significant opening for specialized firms like Qijing.
However, the company faces a crowded field. The AI infrastructure space includes not only other venture-backed startups like Silicon Flow and MoreThanAI but also the formidable cloud incumbents. Tech giants like Alibaba Cloud, Huawei Cloud, and ByteDance's Volcano Engine are all building out their own full-stack AI infrastructure capabilities. To succeed, Qijing must use its funding to build a deep enough technological moat and secure customer loyalty before the giants can close the gap or offer bundled services that are "good enough" for the mass market. The company's success will depend on its ability to turn a technical advantage into sustained commercial growth in a rapidly evolving industry.
This article is for informational purposes only and does not constitute investment advice.
