36Kr reports that the AI sector faces a critical shortage of exit buyers.

Recently, ByteDance secured a USD 29.6 billion syndicated loan, and Alibaba completed a HKD 80 billion share placement. A few months earlier, Google and Amazon raised over USD 800 billion in total financing, and NVIDIA announced a "grand plan" worth USD 500 billion. Major technology corporations around the world are scrambling to raise capital, and they are directing the funds in one main direction: AI infrastructure.

From a macro perspective, every underlying opportunity that reshapes the industrial landscape in modern history has essentially been a top-tier infrastructure upgrade. The railway mania and the Gold Rush in the western United States in the 19th century, China's real estate urbanization drive, and the current global AI infrastructure boom all fit this pattern. Massive capital injections are likely to be marred by market disputes and bubbles, but the infrastructure they leave behind will firmly underpin industrial development for decades.

Compared with previous infrastructure waves, however, this round of AI infrastructure has broader coverage, an unprecedented scale of capital investment, much faster industrial iteration, and more disputes.

In July, AI infrastructure experienced a sharp fluctuation as the capital market began to question the gap between huge investment in AI infrastructure and the returns it can generate. The era of mindlessly stacking GPUs and frantically expanding clusters has not ended, but the market is starting to ask whether, after trillions of dollars have been invested in computing power, companies can actually make a profit.

This signals change on both the capital side and the industrial side. What distinguishes this year from last is that ordinary users have begun to use desktop Agents regularly. AI is no longer a simple chat tool; it has become an agent that can run automatically and carry out tasks in cycles. This year is the true first year of the Agent.

The widespread adoption of consumer-facing (C-end) agents has also pushed Token consumption beyond expectations. In March this year, the average daily Token call volume in China exceeded 140 trillion, more than 1,000 times the level at the beginning of 2024.

There is an awkward problem, however. Everyone is talking about a shortage of GPUs and competing for computing power, yet a large number of GPUs sit idle and unused every day. Carnegie Mellon University ran 756 GPUs continuously for 31 days and found that nearly 20% of the cluster's time was spent "waiting."

For OpenAI's inference business, more than half of energy consumption is wasted on idling. The situation at domestic intelligent computing centers is even more extreme, with average GPU utilization below 30%. Over the past few years, AI infrastructure grew wildly, and rapid expansion could mask efficiency problems. Now that the time for returns is approaching, the competition is over who can reduce Token costs and raise computing power utilization.

Efficiency has become the new key to success for AI Infra. The urgent questions are how to revitalize idle computing power and how to reduce inference costs. As the AI Infra market enters deep water, industry noise is growing.

The AI track is expanding too broadly, with chips, power, hardware, software, models, and applications stacked layer upon layer. New financings, stories, and concepts emerge every day, and each party tells a self-consistent version that serves its own interests. As the book *The Signal and the Noise* observes, in an era of information explosion, noise is always more abundant and faster than signal.

The AI industry is in exactly this state. Everyone tells fervent long-term stories, but no one dares to be certain about the short term. For the hundreds of billions of dollars invested in infrastructure, what is the ratio of long-term dividends to short-term bubbles? Does AI Infra have real barriers? Will cooling market sentiment affect financing and exits in the primary market? What are the differences and gaps between China and the United States in AI Infra, and what opportunities exist to overtake on the curve? Answering these questions requires filtering out the "noise" to find the real "signals." With these trends and questions in mind, ChinaVenture, together with Zhongguancun Science City Corporation and Zhongguancun Venture Street, co-hosted an in-depth closed-door salon titled "The First Year of Agent: Reconstructing New Forces of AI Infra."

**The Seven Sisters Stick Together, While China's Market Is Fragmented**

At this AI Infra salon, Jiuzhang Yunji, Terminus, and Convergent Intelligence were all described as core direct Token suppliers in China. Using their self-developed scheduling, heterogeneous integration, and full-link optimization capabilities, they each have their own approaches to GPU idling, low computing power utilization, and high inference costs.

As Agents break out at scale and Token demand surges across the network, enterprises that can improve efficiency, reduce costs, and ensure stable computing power supply have become the core underlying support for the refined upgrading of AI Infra.

Jiuzhang Yunji has deployed 12 intelligent computing center nodes across China and expanded into the Southeast Asian market. Its business covers three main sectors: new intelligent computing cloud, training factory, and Token factory.

Shang Mingdong, Co-founder and COO of Jiuzhang Yunji, observed a trend: in the US market, computing power supply is highly concentrated, with leading cloud vendors and AI infrastructure companies forming close ecosystem collaborations. The Chinese market, by contrast, is driven more by actual demand and shows more scattered, industry-proximate characteristics.

Convergent Intelligence's core team members have scientific research and engineering backgrounds from the Institute of High Performance Computing, Department of Computer Science, Tsinghua University. The company inherits more than a decade of technical accumulation from that team, and most of its core R&D members have Tsinghua backgrounds. When it was founded at the end of 2023, at the height of the "Hundred Model Battle," Convergent Intelligence chose to focus clearly on large model inference, a rare choice among startups at the time.

Wu Wenjie, President and CFO of Convergent Intelligence, was an investor and researcher in the industry three years ago and is now an entrepreneur. She observed: "Real entrepreneurs and practitioners are deeply aware that there is still a big gap between domestic GPUs and overseas high-performance GPUs, and it is becoming increasingly difficult to obtain those overseas GPUs."

Liu Yue, Head of AI Infra Business at Terminus, focuses on the scenario side. Terminus was founded in Chongqing in 2015. Starting from AIoT scenarios, it has grown into an AI-IoT unicorn. Its shareholders include state-owned central enterprises, well-known investors such as IDG Capital and AL Capital, and leading industrial investors such as JD, iFLYTEK, and SenseTime. Its products are now deployed in more than 170 cities around the world and serve more than 900 customers.

Liu's core view is that Terminus uses vertical scenarios to drive product iteration in AI Infra and Agents. This differs from the mainstream market path, which focuses on general Token services and computing power leasing. "US vendors are more concentrated in infrastructure and upper-layer general models. The domestic market is different, with high scenario density, complete industrial categories, and a flourishing application layer."

The gap at the chip level takes another form. Liu Chen, Managing Director of Haisong Capital, gave an example. Some new companies specialize in inference chips and even write the architecture of a particular model company directly into the chip. This amounts to a dedicated chip for a specific model, with operating efficiency 100 times higher than existing solutions. The obvious cost is that once the customer switches to another model, or the model is iterated, the chip can no longer be used.

Liu Chen explained that there are many intermediate zones between absolute versatility and absolute efficiency, and each chip solves only a specific scenario within one of those zones. "It is difficult to solve all problems through one company or one product." Large companies cannot acquire an entire startup, so they acquire talent