Computing Power Leasing: China's New Core Asset in AI

April 21, Liu Xi, the chief analyst in the computer industry at Guohai Securities, released the industry report "Computing Power Leasing: China's Sharply Defined Core AI Asset." The report systematically reviews the investment logic of the computing power leasing track from three dimensions: business model, industry volume and price trends, and domestic enterprise financing progress. The report asserts that the AI inference turning point has arrived; computing power leasing has entered a high prosperity cycle where both volume and price are rising. It is the current AI industry chain's core track with strong certainty. This judgment is supported by a group of data: The one-year lease price for the H100 GPU rose sharply from its low point of $1.70/hour in October 2025 to $2.35/hour in March 2026, an increase of nearly 40%. Meanwhile, the available GPU leasing capacity for all GPU types has been sold out. Business Model: Long-term Contracts Lock In, Significant Profit Elasticity After Depreciation The business logic of computing power leasing is similar to "heavy asset investment first, stable rent income later." The report points out that the industry mainly relies on 2–5 year commitment-based long-term contracts. As of 2024, the weighted average duration of such contracts is about 4 years. Clients reserve computing power capacity and pay fixed fees within the contract period, typically paying 15–25% of the contract’s total value as a prepayment at signing. After delivery and acceptance, they enter the billing stage. "Performance realization rhythm is clear; cash flow predictability is strong; income is highly stable.” Taking Coreweave as an example, its data center leasing, power, and depreciation costs are recognized before revenue, but "after business normalization, it can achieve a long-term profit margin of 25%–30%." The report further notes that as computing power equipment ages and server depreciation is gradually completed, "core costs will drop sharply. Subsequent stable rental income will largely be converted directly into net profit, and companies are likely to see stepwise profit growth with significant long-term elasticity.” Additionally, the report mentions that demand in different application scenarios is strong, but solutions are not universal—“MoE and other inference workloads operate best on the latest large world-class systems (such as GB300 NVL72), while training workloads achieve the best cost-effectiveness on H100 series GPUs. Therefore, even older GPUs are in high demand.” This means the service life of computing power equipment is extended, and its residual value at the end of depreciation may be revalued. Token Demand Explosion: Overseas Expansion Opens Incremental Space for Chinese AI Changes on the demand side are key to understanding this round of computing power price increases. The report cites Jensen Huang’s judgment from Nvidia GTC 2026, who divides AI into three phases: Generative AI, Inference AI (o1/o3), and the Agent era represented by Claude Code. "AI needs to think, act, read, and infer. Every link involves inference. It has far surpassed the training stage and entered the realm of inference. Over the past two years, computational demand has increased by about 10,000 times, while usage has increased about 100 times.” China's market data is particularly impressive. The report quotes VolcEngine data: The daily average token calls of Doubao’s large model exceeded 120 trillion in March 2026, doubling in the past three months, and up 1,000 times compared to May 2024 when it was first released. For overseas, OpenRouter data show that from March 30 to April 5, 2026, the top six large models in global call volume were all Chinese AI large models, totaling 12.96 trillion tokens per week; from April 6 to April 12, Chinese AI large models topped the weekly call volumes for six consecutive weeks compared to the U.S. Kimi's data is more intuitive: As of February 2026, its overseas revenue has exceeded domestic revenue; from January end to February 23, nearly 20 days of revenue surpassed the whole year of 2025. MiniMax had about 73% of its revenue from international markets. The report suggests that "frequent scenarios like intelligent agents and multi-modal generation continuously increase, directly driving explosive token demand," and domestic large models are expected to form a positive cycle of “application proliferation—call volume increase—model optimization—ecosystem expansion.” Volume and Price Rise Together: Cloud Providers Raise Prices Collectively, Supply Tightening Logic Verified Explosive demand combined with tightened supply is driving computing power services into a price increase cycle. Domestically, the report outlines price moves by three major cloud providers: Tencent Cloud: From March 13, GLM 5, MiniMax 2.5, and Kimi 2.5 ended their free public testing and switched to commercial use; prices for the Hunyuan series models, Tencent HY2.0 Instruct and Think, rose over 400%. From May 9, AI computing power, container services, and Elastic MapReduce (EMR) products had list prices increased by 5%. Alibaba Cloud: From April 18, Pingtou Brother Zhenwu 810E and other computing card products increased 5%–34%, file storage product CPFS (Intelligent Computing Edition) increased 30%. From May 15, service prices for certain MU (Model Unit) model units increased by 2%–7%. Baidu Cloud: From April 18, AI computing power-related products and services rose by about 5%–30%, and parallel file storage rose about 30%. Overseas, Amazon announced on January 22 a 15% price increase on EC2 instances used for large model training; Google announced on January 27 price increases for AI and computing infrastructure services, with a maximum increase of up to 100%. The report notes, “This round of price hikes is not a single vendor action but a clear signal of changes in industry supply and demand, and the upward trend in volume and price has strong sustainability.” There is also a new variable on the supply end: DRAM and NAND memory prices are rising rapidly. The report quotes SemiAnalysis data, predicting that in Q1 2026, LPDDR5 and DDR5 contract prices will rise by about 4x and 5x year-on-year, respectively. OEM vendors have consequently sharply raised AI server prices. “The rising server procurement cost compresses expected project returns, forcing some operators to slow down or even abandon deployment plans,” which may further tighten supply in the leasing market. Domestic Enterprises: Financing Expansion Accelerates, Computing Power Cluster Layout Speeds Up With high industry prosperity, domestic leading enterprises are ramping up financing and procurement at the same time. The report reviews financing progress as of April 21, 2026: - Xiechuang Data: In 2025, applied for a total credit line of up to RMB 51.5 billion; in 2026, applied for up to RMB 50 billion. Filing for H-share listing has been accepted by the CSRC. - Hongjing Technology: Comprehensive credit line of RMB 20 billion in 2025; plans to apply for RMB 60 billion in 2026. Plans to raise RMB 1.29 billion, of which RMB 990 million is for intelligent computing power cluster construction and operations. - Shengshi Technology: In 2026, applied for a total credit line of up to RMB 23 billion. - Litong Electronics: Successfully issued RMB 200 million science and innovation bonds, with raised funds dedicated to computing power equipment and spare parts procurement. - Zhiwei Intelligence: In 2026, applied for a credit line of up to RMB 14 billion. The report concludes, "Stable procurement channels combined with substantial financing to support equipment investment can quickly land computing power capacity. With long-term contracts locking the industry, performance growth is highly predictable". Risk Warning and Disclaimer The market has risks, investment needs caution. This article does not constitute personal investment advice, nor does it take into account individual users’ special investment targets, financial status, or needs. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific situation. Investment based on this is at your own risk.