50% are losing money; computing centers are largely gathering dust.

Nearly 900 million yuan in computing power services were sold in half a year, with net profit increasing sevenfold.

This is Parallel Technology's "harvest" in the first half of 2026. Similarly, several AI giants also have "almost no idle GPUs."

On the other hand, a thousand-GPU-level intelligent computing center in a western city has an occupancy rate of less than 50%, and the actual utilization rate of the servers already on the platform is less than 30%, yet its annual operating costs exceed 30 million yuan. Industry insiders estimate that the utilization rate of intelligent computing centers nationwide is only about 20%-30%.

While scrambling for GPUs, they are running idle.

The high-performance, large-scale, and stable clusters needed by leading AI companies remain scarce; some regions have built computing power in advance, but have not found matching customers and applications.

Lingkong Intelligent founder Peng Yuanyang told LeadDao that many companies don't lack the desire to use AI, but rather don't know which models to use, how much computing power to buy, how to process the data, how to deploy the system, or whether this investment will ultimately bring revenue or cost reduction. There's a missing layer of "connection" between computing power and enterprise needs. - 01 – Who Makes Money First? Peng Yuanyang observed that some small computing power leasing companies in Shenzhen are actually doing quite well. They don't have massive clusters, but they're willing to take on small orders of a few or dozens of cards, getting closer to their customers and understanding better what models their customers are running and how long they need them. Parallel Technology, on the other hand, has taken this business to a much larger scale. In the first half of 2026, the company's computing power service revenue reached 895 million yuan, accounting for 94.62% of total revenue; enterprise customer computing power service revenue reached 705 million yuan, a year-on-year increase of 168.07%. Its core is not simply selling cards from a single data center, but rather integrating computing power resources from different regions and clusters through a computing network, and then scheduling and delivering them. This is a business that improves turnover rate. The recently launched "Computing Power Supermarket" in Jiangning, Nanjing provides another example. During the trial operation phase, only 20 companies connected, but the model was called 1.308 million times, consuming 51 billion tokens. China Electric Transformer Co., Ltd. previously wanted to use a large model for bidding document review and code review, but the team was consistently hampered by computing power pricing and model selection. After connecting to the "computing power supermarket," enterprises can directly select GPUs and mainstream large-scale models for on-demand use. Enterprises do have AI needs, but their past methods of purchasing computing power were too cumbersome. Previously, a short-term project might require renting an entire server and signing a long-term contract; now, more and more platforms charge by card time, core time, or even tokens. Enterprises only pay for actual use, while computing power companies can break down a batch of GPUs into more orders. This is what "computing power supermarkets," "computing power banks," GPU leasing, and computing power scheduling platforms are all doing: repackaging scattered, heterogeneous, and idle computing power into purchasable standard services. However, this matching process isn't extremely profitable. Parallel Technology's gross profit margin for computing power services in the first half of the year was 16.97%. This means that simply selling cards from A to B doesn't yield the profit margin one might imagine. What truly determines profits is whether the same batch of GPUs can be used by more customers for a longer period of time. - 02 – While scrambling for GPUs, many are idle. Why are so many intelligent computing centers "sleeping" despite such high computing power demand? Many places build intelligent computing centers following the infrastructure logic: first determine the scale, buy equipment, build the center, and then wait for demand to come in. But computing power isn't like a highway; once it's built, cars won't automatically come. A survey by Xinhua News Agency's *Banyue Tan* magazine found that a province in central China plans three intelligent computing centers with a total capacity of 1200 PFLOPS (petaflops), but the local digital economy enterprises number less than 100, and the actual demand is less than 200 PFLOPS. In an eastern city, two districts and counties are building intelligent computing centers for biomedicine and industrial internet respectively, but fewer than 10 enterprises have actually moved in. According to data released by the National Bureau of Data Science, as of June 2026, China had over 500 intelligent computing centers in operation, with nearly 400 more under construction or planned. If the industry average utilization rate of these 500 intelligent computing centers is only 20%-30%, and a single intelligent computing center typically needs to reach around 40% utilization to break even, then roughly 250-350 centers may still be operating at a loss or struggling to cover their full costs. Large-scale model training requires high-performance chips, high-speed interconnects, storage, cluster stability, and a mature software ecosystem; industrial and financial scenarios may have different requirements regarding latency, data security, and privatization. Having idle GPUs in a location doesn't mean it can directly handle the training tasks of leading model companies. This is why "giants lacking computing power" and "local computing power being idle" occur simultaneously. Many ordinary enterprises don't know how to find suitable computing power. A factory might only want to do visual quality inspection, a retail company might only want to integrate customer service, and a company might only need a private knowledge base. The project scale might only involve a few or dozens of GPUs, yet people still need to handle model selection, data, deployment, security, and cost. Peng Yuanyang said there's a lot of "dirty work" involved. Large companies prefer to build standardized platforms, making it difficult to enter fragmented scenarios one by one; smaller companies are willing to do so, but their service chains are not yet mature. What computing centers often lack is not the next batch of GPUs, but the next batch of orders. - 03 – What's truly valuable is the Token. Peng Yuanyang believes that for most ordinary enterprises, what they are buying in the long term is inference computing power, which means directly calling existing models and integrating AI into customer service, marketing, code development, industrial inspection, and knowledge bases. Training has a project cycle, but inference occurs continuously along with the business. This is changing the financial model of computing power companies. Qingyun Technology's AI computing power cloud service revenue in 2025 was 43.36 million yuan, a year-on-year decrease, but its gross profit margin increased from 18.99% to 31.62%. One of the reasons the company cited was improved utilization of computing resources and optimized operating costs. Another company showcased another side of the asset-heavy business. Runjian Technology's computing power network business revenue in the first half of 2026 was 519 million yuan, a year-on-year increase of 50.31%, and it claimed that its "Token Factory" business model was basically closed-loop; however, the gross profit margin of this business was only 6.97%. The company explained that one of the main reasons was the high depreciation in the early stages of fixed asset investment. Both sets of figures point to the same fact: buying GPUs is just the beginning. Depreciation still occurs as long as the cards are idle; once utilization increases, each additional task run can potentially reduce fixed costs. Therefore, the next wave of opportunities may not belong to those who build another intelligent computing center, but rather to those who can generate more tokens from the same batch of cards. GPU scheduling, elastic inference, inference acceleration, multi-model routing, and cost optimization represent the first layer of opportunities; further down are model deployment, privatization, enterprise knowledge bases, agents, and vertical industry applications. The latter are closer to the customer's money. When AI remains in the demo stage, computing power consumption is a one-time event; once it enters the production stage, with factories conducting daily quality inspections, customer service handling daily orders, and programmers writing code daily, computing power becomes a continuous source of revenue. Peng Yuanyang believes that the most important thing for intelligent computing centers is to provide stable and low-cost infrastructure, rather than handling all applications themselves. Companies that truly understand the industry should then integrate models and computing power into enterprise business. More than 500 intelligent computing centers nationwide have already been built there. In the next phase, the industry competition won't just be about "who has more GPUs." More importantly, it will be about who can minimize the downtime of a GPU and generate one more collectible token; and who can turn these tokens into a service that businesses are willing to renew for long-term subscriptions.