Following the release of hit applications by AI giants such as Anthropic and ByteDance, along with widespread use of open-source models, the rental market price for NVIDIA's H100 chips has seen a notable rebound. Data from SemiAnalysis indicates the hourly rental rate has climbed from $1.7 in October 2023 to $2.35 in March of this year, a surge of nearly 40%. Demand for GPU computing power remains robust, with on-demand resources sold out and clients even willing to bid higher prices for instances. Some H100 contracts have been renewed through 2028.
Although the next-generation Blackwell chip delivery is delayed, demand for the H100 has increased instead of decreased, primarily driven by media generation and multi-agent workloads. While the market anticipated computing power might become commoditized, the reality is tight supply and high prices. Future attention should be paid to factors such as new supply additions, chip shortages, and the revenue expansion of major AI companies, which will influence prices.
The global AI computing power market is experiencing unprecedented supply constraints and price surges. The nearly 40% jump in NVIDIA H100 chip rental prices highlights a severe imbalance between supply and demand for high-end GPUs. The core driver of this phenomenon is the explosive growth of multi-agent workloads, particularly the soaring global popularity of AI Agent applications, epitomized by "Crayfish," leading to a geometric increase in token consumption. Concurrently, the proliferation of native AI media generation applications has further intensified computing power consumption, pushing inference-side demand to historic highs.
Supply-side constraints are also significant. The delivery cycle for Blackwell chips has extended to 6-7 months, reflecting ongoing tightness in the global AI chip supply chain. This shortage stems not only from limitations in wafer manufacturing capacity but also from the high requirements for AI computing infrastructure development, including needs for power supply upgrades and liquid cooling systems, all of which elevate overall operational costs. Although consumer-grade DDR5 memory prices have seen a correction, DRAM chip prices remain firm, further confirming the supply-demand imbalance for AI computing-related hardware.
Market analysts believe this wave of computing power price increases is fundamentally a result of the combined effect of exploding AI demand and supply bottlenecks. As AI applications transition from being training-driven to a dual-driver model of both training and inference, the computing power gap is expected to persist, catalyzed by more modalities and broader scenarios. If supply chain costs continue to rise, there is a possibility that cloud computing providers may implement further price increases.
The AI computing power rental market is undergoing a price surge driven by supply-demand imbalance. The near 40% increase in H100 chip rental prices over six months reflects the extreme scarcity of computing resources. This tightness originates from the explosive growth of multi-agent workloads and the adoption of native media generation applications, causing token consumption to rise parabolically and sustaining strong demand even for the relatively older H100 chips. The intensified supply-demand矛盾 has led to extended Blackwell chip delivery cycles of 6-7 months, with clients renewing H100 contracts as far out as 2028, underscoring the long-term nature of the computing power shortage.
The computing power rental business model is primarily based on 2-5 year commitment contracts, supplemented by pay-as-you-go models. Cloud service providers have successively initiated a wave of price hikes, indicating the industry's持续 high景气度. Although narratives about potential oversupply exist, the reality under aggressive supply constraints is that all types of computing resources maintain vigorous demand, creating a disconnect from market sentiment. Future focus should be on the progress of GB300 cluster volume production, the evolution of chip shortages, and the expansion of recurring revenue for AI giants. These factors will determine whether GPU rental prices can remain high.
AI computing power demand is experiencing explosive growth, with NVIDIA H100 chip rental prices surging 40%, presenting a market where demand outstrips supply. As giants like Anthropic and ByteDance launch hit AI applications, multi-agent workloads and media generation needs are激增, driving computing power rental prices continuously higher. Some H100 contracts have been locked in until 2028.
The computing power rental/GPUaaS industry is entering a high-growth cycle. H200 monthly rentals have increased 25%-30% month-over-month, and cloud service providers domestically and internationally are beginning a wave of price increases. The storage segment is also benefiting, with AI accelerator chips equipped with HBM4 nearing mass production. The Rubin platform's memory usage is expected to double compared to Blackwell. Domestic computing power supply is expected to improve, with NVIDIA's H20 resuming shipments to China and leading manufacturers accelerating procurement to supplement high-end computing capacity. Jensen Huang predicts that the Blackwell + Vera Rubin product line revenue will reach $1 trillion by 2027, indicating clear long-term growth space for the computing power industry.
The ChiNext Artificial Intelligence ETF Huabao (159363) invests in hot sectors like communications, computers, and media. Its top ten holdings are Sunwave Communication, Zhongji Innolight, TFC Optical Communication, Rizen Technology, BlueFocus Communication, Xcreate Data, Kunlun Tech, Ingenic Semiconductor, Wangsu Science & Technology, and Tonghuashun, with a combined weighting of 59.98%.
The STAR Market Artificial Intelligence ETF Huabao (589520) invests in hot sectors like electronics, computers, and home appliances. Its top ten holdings are VeriSilicon Microelectronics, Cambricon Technologies, Montage Technology, Kingsoft Office, Allwinner Technology, Fudan Microelectronics, Roborock, CAS Star, Bestechnic, and CloudMinds Technology, with a combined weighting of 67.65%.
The Big Data ETF Huabao (516700) invests in hot sectors like computers, communications, and media. Its top ten holdings are iFlytek, Sugon, Rizen Technology, Inspur Information, Topscomm Communication, Unisplendour, Wangsu Science & Technology, Hundsun Technologies, CAS Star, and China Great Wall, with a combined weighting of 51.43%.
MACD golden cross signals have formed, indicating positive momentum for these stocks.