Summary of Key Points
The commercialization of large AI models and AI agents has turned computing power into a highly sought-after commodity. High-end GPUs are in short supply, and building custom computing clusters is costly and time-consuming. As a result, more and more companies are turning to computing power leasing, creating an era where sellers have the upper hand in the market. Various players from different backgrounds (research service providers, cross-industry entrants, and low-capital platforms) are competing for market share using their respective approaches, but most of the profits go to chip manufacturers, channel providers with access to quotas, and financial investors. The industry is expected to evolve towards推理 computing power, domestic alternatives, and the financialization of computing power. The traditional model of making money simply by having access to hardware will become obsolete, and only those who focus on providing quality services and tailored solutions for specific use cases will survive.
Why Has Computing Power Leasing Suddenly Become So Popular?
In simple terms, there is a surge in demand that outpaces supply, giving sellers the power to set prices. Training and inference for large AI models require a large number of high-end GPUs (such as NVIDIA's H200 and GB200), which are not only difficult to obtain due to import and export restrictions but also extremely expensive to acquire and maintain (for example, building a small cluster can cost tens of millions of yuan). According to data from the China Academy of Information and Communications Technology, domestic demand for AI computing power increased by 417% in 2026, while supply only increased by 128%. This situation creates a scenario where those with access to high-end GPUs can set prices at will. Service providers with these GPUs often see long queues of customers waiting for their services, and they can raise rents at will. Only customers who sign long-term leases (1-3 years) get priority in accessing resources, as providers prefer stable, long-term contracts over short-term, high-price deals for better cash flow and financing opportunities.
Who Is Engaging in Computing Power Leasing?
Players from different backgrounds have distinct strategies:
1. Research service providers transitioning to leasing: For example, Wang Yan's company, which originally provided computing power support for academic research, noticed an increase in demand from engineering fields (such as chemistry and materials) and started leasing in 2023. They retain customers through excellent technical services (e.g., responding to issues within 10 minutes) and offer additional services like cluster setup.
2. Cross-industry entrants: For instance, ian, who used to rent luxury apartments, stumbled upon the high-profit potential of computing power leasing. They utilize overseas resources (such as Melbourne and Kuala Lumpur nodes) and collaborate with domestic IDCs to acquire and deliver GPUs, earning profits from the trade difference and rental fees.
3. Low-capital platforms: Startups established in 2026 often start by aggregating computing resources and use leasing as a foundation, avoiding the need to purchase hardware themselves.
4. Major players: Public clouds (like Alibaba Cloud and Tencent Cloud), telecom operators, and professional IDCs dominate the market, but smaller players can still find opportunities in niche areas (e.g., research, cross-border transactions, and compliance requirements).
How Does Computing Power Leasing Generate Profit? Is the Payback Fast?
The methods and time frames for making a profit vary significantly depending on the business model:
- Pricing:
- For research purposes, CPUs are priced based on core hours (0.04-0.1 yuan per core hour), and GPUs based on card hours (2-4 yuan per card hour for 4090 GPUs, 3.5-5 yuan per card hour for 5090 GPUs).
- In cross-border transactions, domestic customers pay up to 105,000 yuan per month for H200 servers, while overseas customers pay $550-830 per hour for GB200 GPUs; mid-range PRO6000 GPUs cost around $1.5-2 per card hour.
- Payback periods:
- Research GPUs take about 4 years to break even due to rapid hardware price increases and limited rent increases.
- Mainstream cross-border servers take 2.5-3 years.
- High-end computing power with high utilization rates can generate a faster return on investment (e.g., through token sales).
- General-purpose computing power with low usage requires longer payback periods.
- Profit margins:
- Research GPU services have low margins (due to hardware price pressures).
- Providers with direct access to chips (like ian) earn higher profits.
- Pure resource resellers face lower margins and the risk of hardware price fluctuations.
Who Is Reaping the Biggest Profits?
The biggest beneficiaries are the upstream chip manufacturers and financial investors:
1. Chip manufacturers and channel providers: Companies like NVIDIA control the supply of high-end GPUs, giving them significant influence over market prices and competition among cloud providers.
2. Financial investors: They invest in purchasing servers for long-term leases (3-5 years), enjoying stable returns. For example, they bought high-end GPUs at low prices in 2024-2025 and can now profit from both rental income and asset appreciation as hardware prices rise.
3. Midstream service providers: Those who manage data centers, bandwidth, and directly interact with customers only get a small portion of the profits and face risks (such as customer loss and hardware price drops).
How Will Computing Power Leasing Evolve in the Future?
Several key trends are likely:
1. Increasing demand for inference computing power over training: Inference tasks (e.g., AI chat and image generation) can be handled by consumer-grade GPUs (4090, 5090), making it more accessible to smaller service providers and reducing the reliance on high-end GPUs.
2. Domestic chip substitution: Domestic chips (such as Huawei's Ascend 950) are gradually being adopted, though there is still a gap compared to NVIDIA. This process will be driven by policies, especially in academia and state-owned enterprises. In the future, a mix of overseas and domestic GPUs will be used, and smaller providers without diverse supply sources and technical capabilities will struggle.
3. Financialization of computing power: GPUs may be used as collateral for financing, allowing service providers to expand by buying more computing power through loans. However, this requires stable long-term leases and high utilization rates.
4. Industry consolidation: Pure resource resellers will face competition, and only those who provide specialized services and tailored solutions for specific use cases (e.g., research and gaming) will thrive.
In the short term, there may be another shortage of computing power in 2027, leading to further rent increases. In the long run, prices will stabilize after market consolidation, and only those who offer comprehensive services and deep expertise will survive.
In summary, the current period is a seller's market for computing power leasing, but the days of easy profits are coming to an end. The future will depend on providing quality services, leveraging resources effectively, and focusing on specific use cases.
(End of translation.)