In mid-September, cloud computing service provider Nebius issued a price adjustment notice to its customers, planning to raise the on-demand rental rates for major NVIDIA GPUs starting October 1: the H100 from $3.85 to $4.50 per hour, the H200 from $4.50 to $5.40, the B200 from $7.15 to $8.50, and the B300 from $7.85 to $9.50, representing an increase of approximately 17% to 21% in this round.
This marks the company's second comprehensive increase in on-demand pricing this year. Calculated based on the two rounds of public quotations, starting from the pre-adjustment rate of $6.10 in the first round, the cumulative annual increase for the B300 hourly rental has reached approximately 56%.
On the flip side of the continuous price surge is the widening gap between computing power supply and demand. Data from the China Academy of Information and Communications Technology (CAICT) shows that in the first quarter of 2026, domestic AI computing power demand grew by 417% year-on-year, while the growth rate of effective supply was only 128% during the same period. From overseas on-demand quotations to domestic contract values, the chain of price increases is being validated segment by segment.
01. A Price Curve Elevated Twice
This round of price adjustments is not an isolated event. In a price hike announced in May and effective June 1, Nebius had already raised the H100 from $2.95/hour to $3.85, the H200 from $3.50 to $4.50, the B200 from $5.50 to $7.15, and the B300 from $6.10 to $7.85, with an overall increase of about 29%. With the two rounds compounded, the absolute level of Nebius's own quotation sequence has thus been elevated round by round.
On-demand pricing deserves separate attention because it serves as the most direct pricing anchor for newly signed contracts and small-to-medium clients: an upward shift in on-demand prices means a simultaneous increase in the marginal retail pricing system. Zheshang Securities believes this round of increases confirms the upward shift across the board for long-term contracts, short-term contracts, and on-demand prices.
What better illustrates the tightness is that the scope of price increases has spilled over beyond accelerators themselves. In the same batch of adjustments, pure CPU computing for AMD EPYC Genoa was raised by 25% from $0.012/vCPU hour to $0.015, and Genoa memory was increased by about 41% from $0.0032/GiB hour to $0.0045. The supply tightness is not confined to GPUs alone but covers multiple segments of the computing infrastructure.
The long-standing convention in the semiconductor industry is that chips begin to depreciate upon leaving the factory; today, this rule has reversed in the computing power rental sector. Official statements from NVIDIA corroborate this trend: according to Colette Kress, NVIDIA's Chief Financial Officer, during the second-quarter fiscal 2026 earnings call, rental prices for the Hopper series GPUs have cumulatively increased by 20% this year, while cloud prices for the previous-generation A100 have risen by nearly 15%.
02. 417% vs. 128%: Where Does the Threefold Gap Come From
Data from CAICT provides the most direct footnote for this price hike: in the first quarter of 2026, domestic AI computing power demand grew by 417% year-on-year, while the effective supply growth rate was only 128%. Dividing the two figures, the demand growth rate is approximately 3.3 times that of the supply growth rate.
The gap is difficult to bridge in the short term, which is related to the pace of computing infrastructure construction. Kaiyuan Securities mentioned in a research report that the construction cycle for traditional data centers spans over three years; from site selection, approval, and civil engineering to equipment installation, debugging, and racking, every step is hard to compress, whereas the iteration of AI models and applications is measured in quarters. The inconsistency in time scales between supply and demand directly fuels the strong demand for asset-light models such as pay-as-you-go and elastic scaling.
Scale data corroborates this point. CAICT data shows that in the first quarter of 2026, the domestic computing power rental market size reached CNY 68 billion, a year-on-year increase of 62%, and is expected to exceed CNY 260 billion for the full year. IDC's forecast is more macroscopic: the global computing power rental market size will surpass $80 billion in 2026, with a compound annual growth rate exceeding 25%. Rental is transitioning from a temporary stopgap measure to a regular component of the computing power supply system.
03. 140 Trillion Calls: The Demand-Side Engine Has Been Upgraded
The driving force on the demand side can be clearly seen from a single figure: data from the National Data Administration shows that in March 2026, China's average daily Token invocation volume exceeded 140 trillion, while at the beginning of 2024, this figure was about 100 billion, representing a growth of over 1,000 times in two years; even compared to the 100 trillion at the end of 2025, it grew by more than 40% within three months.
This acceleration stems from the replacement of AI application paradigms. In the past, users asked a question to a large model and received an answer, consuming a limited number of Tokens; today, agents handle multi-step tasks—planning paths independently, invoking tools, and repeatedly verifying, with a single task consuming several or even dozens of times the Tokens of traditional Q&A. A forecast released by Gartner in August 2026 shows that global inference spending will reach $23.3 billion in 2026, surpassing the $19 billion for training, with 55% of AI-optimized Infrastructure as a Service (IaaS) spending dedicated to supporting inference scenarios. The focus of computing power consumption is shifting from training to inference.
Industry judgments on future potential are equally clear. Huawei's Intelligent World 2035 released on September 16 predicts that by 2035, global annual Token consumption will grow 100,000 times, with agent traffic accounting for over 90%. The Research Report on AI Infrastructure Development in the Agent Era (2026) released by the China Telecom Research Institute estimates that China's annual Token consumption will reach 100 quadrillion in 2026 and exceed 35 quintillion by 2030, with a compound annual growth rate of nearly 12 times.
The demand curve is not linear but rises in a multiplicative manner, which also explains why prices still have upward momentum even as supply continues to expand.
04. From On-Demand Prices to Contract Prices: Segment-by-Segment Validation of the Price Hike Chain
If overseas on-demand quotations reflect the immediate supply and demand at the retail end, then the contract values of domestic listed companies serve as evidence that price increases have made it into financial statements.
On the evening of September 16, Saiyi Information issued an announcement stating that the company signed a supplementary agreement with a customer regarding a high-performance computing power service contract, adjusting the tax-inclusive amount of one contract from CNY 3.6 billion to CNY 3.87 billion, and the combined tax-inclusive total amount of the two contracts from CNY 6.45 billion to CNY 6.72 billion, an increase of CNY 270 million. The announcement showed that the service period of the contract is 60 months, and the original tax-inclusive total amount accounted for 311.11% of the company's audited operating revenue for 2025; apart from the amount adjustment, core terms such as the service period, delivery schedule, and monthly settlement model remain unchanged.
This is not an isolated case. Public information shows that in June 2026, Tianyang Technology signed a CNY 303 million computing power cloud service agreement with Yunli Wisdom, and less than a month later, signed a supplementary agreement to raise the total amount to CNY 402 million. In July 2026, Xingyun Technology's wholly-owned subsidiary Yueyunshu signed supplementary agreements with customers for two server rental agreements, increasing the total rent by CNY 2.879 billion, a 79% increase over the original signed amount; during the same period, after signing a supplementary agreement with a customer, its other wholly-owned subsidiary Shenzhen Xingyun increased the number of computing power service units from 128 to 256, and the total contract amount rose from CNY 1.014 billion to CNY 3.053 billion, of which the monthly computing power service fee for the 8 already delivered units increased by 21.21% per unit compared to the original contract, and the subsequent 248 units increased by 36.36%. Jingang Photovoltaic, on the other hand, terminated the original computing power service contract of approximately CNY 399 million and signed a new contract of approximately CNY 614 million with a new counterparty, with both contracts having a service period of 5 years.
Zheshang Securities summarizes this change as a shift in pricing logic: in the first half of 2026, the pricing core of the computing power rental sector was the supply-side capability to acquire cards, and the market was trading on who can get GPUs; entering the current stage, pricing power is shifting to the demand side, and the market is beginning to validate whether customers are willing to pay more for computing power. The increase in domestic computing power rental prices has entered the stage of announcement validation.
Official price adjustment announcements from cloud vendors have added the middle link to this chain. On March 18, Alibaba Cloud and Baidu AI Cloud released price adjustment announcements on the same day: Alibaba Cloud raised prices for computing cards such as T-Head Zhenwu 810E by 5% to 34%, and increased prices for the file storage product CPFS (Smart Computing Edition) by 30%; Baidu AI Cloud raised prices for AI computing-related products and services by about 5% to 30%, and parallel file storage by about 30%. On April 9, Tencent Cloud announced a uniform 5% price increase for three product categories: AI computing, container service TKE-native nodes, and elastic MapReduce. Kingsoft Cloud raised prices for AI computing-related products by 15% to 50% in July. The overseas market adjusted synchronously: Amazon implemented a price increase of about 15% for EC2 instances used for large model training in January this year, and Google Cloud adjusted prices for multiple services including data transfer, AI, and computing infrastructure in the same month.
Beyond prices, business models are also iterating. Kaiyuan Securities points out that the business models of computing power rental vendors are upgrading from simple bare computing power leasing to model services or Token sharing, shifting from selling computing power to selling Tokens.
05. 2185 EFLOPS: Catching Up and Structural Optimization on the Supply Side
A widening gap does not mean supply has stagnated. The Ministry of Industry and Information Technology (MIIT) introduced at a State Council Information Office press conference on August 26 that by the end of June 2026, China's intelligent computing power scale reached 2185 EFLOPS, a year-on-year increase of 177%; over 70 major computing power corridors have been built around computing hubs, and the overall racking rate of national computing facilities has reached 71.4%.
Structural changes are equally worthy of attention. According to the China Data Industry Development Report (2026) by the National Data Development Research Institute, by the end of June 2026, the intelligent computing scale of China's eight major computing hubs reached 2.14 million PFLOPS (FP16), a 54.2% increase from the end of 2025, accounting for nearly 90% of the national total. Computing resources are concentrating towards hub nodes, and cross-regional scheduling capabilities are accordingly improving. The National Data Administration disclosed that by August 2026, over 126,000 high-quality datasets had been built nationwide, with a total data volume exceeding 1815 PB—computing power, data, and model support are advancing in tandem.
However, constraints on the supply side remain real. According to public reports, TSMC's advanced process capacity has been locked in by major clients such as NVIDIA, and there are no signs of relief expected until the first half of 2027; the supply of HBM (High Bandwidth Memory) remains tight, thereby lengthening the delivery cycle for complete high-end AI servers. Kaiyuan Securities judges that, constrained by advanced packaging capacity, the tight supply-demand balance of high-end GPUs may persist throughout 2026. In other words, the support for computing power prices is not a short-term disturbance but is caused by the physical rhythm of capacity construction.
The key variables facing the computing power industry are now very clear: demand-side growth remains in a high-slope range driven by agents, and whether the supply side can simultaneously accelerate in chips, advanced packaging, power support, and data center delivery will determine whether the gap continues to widen or converges.
Computing power is transforming from a one-time capital expenditure into a continuous operational expenditure—this shift in identity is perhaps the most noteworthy footnote for this round of price increases.