The GPU market has great investment potential, but it also faces price competition and supply challenges.
Morgan Stanley analyst Joseph Moore and his team report that the GPU rental market is currently in short supply and has an extremely high return on investment (ROI). Although the rental price of the H100 has declined in the past six months, its absolute price still shows that the return on investment of the hardware is less than one year, which is undoubtedly a safe choice for investors in data centers, such as Nvidia/Haiwuyun.
The report notes that conservative estimates suggest that investing in data centers can generate an internal rate of return of 30% to 50%. As more companies build AI applications through cloud computing, the market return on market for communication service providers (CSPs) and teri-2/3 cloud providers is gaining traction.
GPU computing power leasing refers to a way to perform tasks such as high-performance computing, machine learning, and deep learning by leasing GPU resources provided by cloud computing service providers. This method allows users to flexibly use GPU computing power according to actual needs without purchasing and maintaining expensive GPU hardware.
GPU computing power leasing has the following key differences compared to traditional methods of purchasing GPUs:
Cost flexibility
Leasing: With a pay-as-you-go model, users can pay based on actual usage, avoiding a large one-time investment. This method is particularly suitable for projects with limited budgets or unstable demand.
Purchase: You need to invest a large amount of money in a lump sum to purchase GPU hardware, and you need to bear the costs of maintenance, upgrades, and depreciation of the equipment.
Maintenance and updates
Leasing: The service provider is responsible for the maintenance and update of the hardware, and users do not need to worry about hardware failures, technical updates, etc., and can focus on the development and operation of the project.
Purchase: Users are responsible for maintenance, upgrades, and troubleshooting of the hardware, which may add additional operational burdens.
Elastic and scalable
Leasing: You can flex and adjust computing resources flexibly based on actual needs. This elastic scaling feature is very beneficial for handling changing workloads.
Purchase: Once the GPU hardware is purchased, its computing resources are fixed and difficult to adjust flexibly according to actual needs.
Application scenarios
Rental: Suitable for various computing and processing tasks that require GPU computing power, such as graphics processing, machine learning, artificial intelligence, etc. Particularly suitable for short-term projects or scenarios with large changes in demand.
Purchase: While it can also be used for various GPU-related tasks, it has a higher purchase cost and is more suitable for long-term or stable projects.
Currently, GPU leasing is mainly divided into On-Demand instance pricing and spot instance pricing. Morgan Stanley conducted an in-depth study of GPU spot instance pricing on the market and examined some GPU supply sources, such as third-party marketplace platforms GPUlist.ai.
GPUs are in short supply, and the return on investment is extremely high
According to the data, Morgan Stanley believes that the supply of GPUs remains tight, and demand continues to exceed supply. Although H100 rental prices have dropped from unrealistically high prices due to improved supply, their absolute level still shows a high return on investment. The report pointed out that the rental price of the H100 spot is about $4.50 per hour, which means that the internal rate of return on building a 100MW data center can reach 30% to 50%.
Additionally, the Infrastructure-as-a-Service (IaaS) gross margin for GPUs could reach over 50% in 2025, which is a fairly attractive return for investors. However, investors have also expressed concerns about the sustainability of returns, especially as companies focused on GPU cloud computing (such as Coreweave, Lambda Labs, etc.) continue to scale up.
Challenges under return on investment
While data centers have high IRRs, Morgan Stanley cautions investors not to be overly optimistic because GPU leasing is relatively cheap outside of major cloud service providers.
Nvidia has said that for every $1 invested in their GPUs, it will generate up to $5 in revenue over 4 years. But Morgan Stanley believes this estimate is too idealistic because it does not take into account the decline in rental prices and the lack of 100% GPU usage.
Morgan Stanley assumes that GPU leasing prices will drop by 20% per year, with an initial lease price of $4 per hour for H100 and an 80% utilization rate within 6 years. Based on the above three assumptions, Morgan Stanley estimates that a $3.2 billion investment in a new 100MW data center could generate $7.8 billion in revenue over six years, bringing an unlevered pre-tax internal return of 38%.
According to GPUlist.ai, the median hourly rental price for the H100 is $2.27, which is more than 80% lower than the price of AWS. This is a very attractive discount for consumers who need to lease a lot of computing resources.
Additionally, GPU leasing is often associated with less software, making it easier for consumers to switch between different cloud services in search of the best price. However, whether GPU spot instance pricing is sufficient to meet consumer demand remains a question, as these GPUs have a high turnover rate and may not be suitable for large-scale model training with stable computing resources over long periods of time.
Nvidia's market position
In Morgan Stanley's theoretical data center model, to achieve the same return, cloud service providers would need to price AMD's Mi300X 25% to 30% lower than the H100. Although AMD's pricing is reasonable to help it gain a foothold in the market, NVIDIA/Haiwu Group's Haiwu Cloud is still a safe choice in the rental market for cloud service providers, as the opportunity cost and risk of other options are relatively high.
In summary, the GPU market has great investment potential, but it also faces price competition and supply challenges. With the development of technology and changes in market demand, investors need to carefully evaluate various factors to formulate the best investment strategy.
