In today’s world where the AI wave is sweeping the globe, computing power has become the core engine driving industrial transformation. Yet compared with the chip giants or large model companies in the spotlight, one “less glamorous” segment has quietly become the first to cash in on profits — computing power leasing. In late May 2026, rumors about “prohibiting listed companies from disclosing overseas card deployments without invoices” hit the market, compounded by the shock of DeepSeek’s announcement of permanent API price reductions, sending computing power leasing concept stocks into consecutive declines. Pingzhi Information fell more than 12%, while Jiahua Technology, Alte, and multiple other stocks fell more than 9%. Behind this volatility, is this a fundamental reversal of industry logic, or a normal pullback in an overheated market?
A Rising Tide of Price Increases: What Is the Computing Power Leasing Market Going Through?
The price signals from the computing power leasing market are issuing the most direct supply-demand alerts.
After the 2026 Spring Festival, China’s domestic computing power leasing market welcomed a significant wave of price increases. According to industry expert interviews conducted by relevant research institutions in the industry, domestic computing power leasing prices rose approximately 30% overall after the Lunar New Year. Taking the H100 as an example, annual lease rates rose from approximately RMB 50,000 (≈ USD 6,944) during the holiday period to RMB 67,000 to 68,000 (≈ USD 9,306 to USD 9,444); the H200 rose from just over RMB 60,000 (≈ USD 8,333) to RMB 80,000 (≈ USD 11,111). This price increase trend is not an isolated phenomenon in the Chinese market. Global computing power leasing leader Nebius announced a comprehensive upward adjustment to GPU leasing prices effective June 1, with high-end models including the H100, H200, B200, and B300 all seeing increases of more than 30%.
It is not only leasing prices that are rising — the prices of the servers themselves are also climbing sharply. High-end B-series card prices have been driven up to approximately RMB 6.5 million (≈ USD 902,778), with the B300 series even showing premiums of 30% to 40%; the H200’s procurement price has risen all the way from a previous RMB 2.1 million to 2.3 million (≈ USD 291,667 to USD 319,444) to RMB 3.5 million (≈ USD 486,111).
Even more noteworthy is the severity of the supply-demand tension. According to market research, the current utilization rate of NVIDIA computing cards domestically is as high as 90% to 95%, with essentially no idle capacity, and supply rigidity is extremely tight. Some industry insiders have reported that even trying to lease as few as 32 mid-range H100 or H200 cards for one year is simply impossible, and even the lower-end A800 is incredibly hard to get.
Leasing contract terms have also been tightening. The current market mainstream is contracts with a minimum of one year, with lower unit prices for longer lease terms, but the minimum lease period is essentially locked at one year.
From Training to Inference: A Structural Turning Point in Computing Power Demand
The driving force behind the price increases comes from a profound shift in the structure of computing power demand.
Over the past two years, the market focused more on the large model training phase — who was building bigger models, who could stack more parameters. But now the center of gravity of the industry is shifting. As applications such as AI assistants, intelligent customer service, AI programming, financial investment research, and industrial quality inspection progressively come to life, inference computing power is becoming the dominant form of computing power consumption.
According to industry survey data, the current ratio of inference computing power demand to training computing power demand has leaped from the 1.5 to 2:1 ratio of last year to 3 to 4:1, with the inference proportion continuing to grow. The significance of this structural change lies in the fact that training is a project-based, one-time investment, while inference is an operations-based, ongoing consumption. Training is like building a factory; inference is like running the machines every day.
The drivers of demand come from multiple levels. Domestically, the explosive growth of consumer-facing agent applications and short-video generation has directly driven demand for inference computing power. Overseas markets are also showing a distinctive migration trend: China’s Token prices are only one-tenth of those in Europe and America, and global mid-to-low-end computing power is migrating toward China. At the same time, the H100 and H200 spot leasing prices from the U.S. computing power leasing leader Nebius have risen by approximately 70%, further corroborating the global scarcity of computing power.
Relevant research institutions in the industry have pointed out that the industry has already transitioned from the phase of computing power infrastructure construction to the phase of efficient computing power scheduling and operations. This means that whoever can efficiently schedule existing computing power in an environment of tight supply will occupy a key position in the supply chain.
The Commercial Map: Three Mainstream Models of Computing Power Leasing
Faced with the price increase environment, enterprises with different business models are exhibiting markedly different degrees of benefit elasticity.
The current computing power leasing market has broadly formed three categories of business models. The first is the heavy-asset intelligent computing center model — self-invested, self-built, self-held, and self-operated, covering almost all segments. A representative enterprise is Runze Technology, whose Langfang park is the largest single-site intelligent computing industry park in Asia, with deployments in the Beijing-Tianjin-Hebei region, the Yangtze River Delta, the Greater Bay Area, and the Chengdu-Chongqing economic zone. It has signed 10- to 15-year contracts with ByteDance, possessing strong customer barriers.
The second is the light-asset computing power leasing model, where companies procure or lease GPU cards and servers to schedule and sublease computing power. This model offers the greatest flexibility, especially for companies already holding large quantities of GPU cards and servers, who are able to directly enjoy the price increase dividend in the short term. Litong Electronics is a typical representative of this model. As NVIDIA’s only “Preferred”-level AI cloud partner in mainland China, its card procurement costs are 10% to 15% lower than peers, with quotas 30% to 50% higher than ordinary partners. From a performance standpoint, this model has already begun delivering profits. Litong Electronics’ 2025 net profit attributable to parent company shareholders reached RMB 293 million (≈ USD 40.69 million), up 1,088.59% year-on-year. The company has explicitly stated that this is primarily attributable to the continued increase in AI computing power demand and the ongoing execution of previously established computing power leasing projects.
The third is the Token factory model. Enterprises self-build intelligent computing centers and deploy open-source large models, charging users based on Token invocation volume and splitting revenue with the model provider according to an agreed ratio. This model also benefits when computing power is scarce — the scarcer the computing power, the more expensive the models, and Token prices rise accordingly.
It is worth noting that the industry’s center of gravity is migrating from pure computing power leasing toward “Token distribution platforms.” By building platforms that connect to open-source large models and providing Token services to global users to earn the price differential, profit margins can reach 20% to 30%. Xiechuang Data’s performance confirms this trend — its 2025 intelligent computing power products and services revenue reached RMB 2.761 billion (≈ USD 383.47 million), up 1,727.17% year-on-year.
Profit Realization and Competitive Landscape: Who Is Making Money, Who Is Catching Up?
The reason computing power leasing is “unglamorous yet first to deliver profits” is fundamentally because it directly absorbs the real gap created by supply-demand mismatch.
From a profitability standpoint, the profit margin of computing power leasing has risen from a previous 10% to 20% to approximately 30%. Based on a five-year payback period calculation, older equipment has relatively lower depreciation pressure due to its higher residual value. Relevant industry research institutions have also noted that some computing power leasing companies, relying on large-scale computing clusters deployed early, have completed delivery and acceptance and entered the substantive billing phase, with the commercial closed loop gradually being established and profit release accelerating.
However, the competitive landscape is far from smooth sailing. Domestically produced computing power still faces significant challenges in market competition. According to industry surveys, the hardware price of domestically produced computing power is 1.5 to 2 times higher than comparable NVIDIA products, with actual performance below 60% of NVIDIA’s, lagging comprehensively in cost-performance ratio, software ecosystem, and adaptation cycle. It is currently primarily used for government-driven “domestic substitution” projects and has limited market competitiveness. More critically, the process technology lags behind overseas peers by more than 1.5 generations, and the technology gap continues to widen.
On the other hand, the possibility of a relaxation of H200 policy adds another variable to the market. There are rumors that 10 leading enterprises may receive approximately 750,000 H200 units of quota, corresponding to 30% to 40% of the domestic existing market. If this policy materializes in substance, computing power prices could experience a short-term decline of 10% to 15%. It is worth noting, however, that since the proportion of memory and hard drives in total machine costs has skyrocketed from 10% to 30% to 40%, the center of total machine costs has been permanently elevated, and even if the H200 restrictions are lifted, computing power prices are unlikely to return to their previous historic lows.
New Directions in a Changing Landscape
Despite the current heat of the computing power leasing market, investors and industry participants still need to maintain clear-headed judgment. The collective adjustment in computing power leasing concept stocks in late May was the market’s immediate reaction to rumors and price war expectations. DeepSeek’s announcement of a permanent API price reduction to 25% of the original price also signals that competition at the Token service level is intensifying.
Relevant industry research institutions believe that computing power leasing prices still have an upward trend on the pricing side. But looking at a longer cycle, the essence of computing power leasing is a service layer connecting supply and demand; its prosperity is rooted in the mismatch between high-quality computing power supply and real demand. As domestically produced computing power gradually catches up and AIDC construction continues to advance, this mismatch will ultimately tend toward convergence.
The long-term direction of the industry may lie in “going overseas.” Overseas, apart from large companies and research institutions, a large number of startups and independent developers are using Chinese large models, and the market space for Token overseas expansion is continuously growing. Token aggregation and distribution platforms — a business model oriented toward global monetization — may become the direction with greater profitability advantages within the computing power leasing track.
For market participants, understanding the short-term dividend window and the long-term industrial logic of computing power leasing is more important than chasing every round of concept speculation. In the long race of computing power, whoever can continuously create value in the chain connecting supply and demand will be the one who transcends market cycles and becomes the true winner.
[Disclaimer]: The above content reflects analysis of publicly available information, expert insights, and BCC research. It does not constitute investment advice. BCC is not responsible for any losses resulting from reliance on the views expressed herein. Investors should exercise caution.