In 2006 Amazon introduced Amazon Web Services (AWS) as the first general purpose cloud computing platform. Microsoft and Google followed with Azure and Google Cloud Platform (GCP), respectively. These three players have been able to build scale in what has developed into an astounding $300 billion high return on capital business. The foundation of these businesses is offering Infrastructure as a Service (IaaS), namely compute and storage, to millions of enterprise customers. Most of these customers that run workloads on cloud infrastructure also adopt platform services such as databases, analytics, security, and many others.
Once an enterprise customer adopts numerous proprietary platform services from a cloud provider like AWS or Azure, it becomes difficult for that customer to move its workload off the platform. Customers often build significant customization into these platform services, and the process of rebuilding the workload on another cloud and migrating all its data is expensive and time-consuming.
Customers also rely on large internal IT departments to manage these critical workloads. Most of the IT workforce is trained and certified on one of these few scale cloud platforms, making it difficult and risky to move workflows off the platform the IT staff already knows, or even to consider a newer general-purpose cloud.
As a result of these dynamics, the two or three scale general-purpose cloud vendors have generated consistent high returns on massive and growing levels of tangible capital, making them some of the world’s greatest businesses.
While the general-purpose cloud market matures, a new and potentially larger market has developed: AI Cloud, providing IaaS for Large Language Models (LLMs). The economic characteristics of this market are, so far, significantly different from those of the general-purpose cloud business.
The first difference is the customers. Unlike the general-purpose cloud market, AI Cloud demand is so far dominated by a few exceptionally large and sophisticated customers, such as Anthropic, OpenAI, and other LLM model providers.
The second difference is the services being offered. Most AI lab customers demand graphics processing unit infrastructure services that are quite similar across vendors, and therefore much more commoditized than the many proprietary platform services of general-purpose cloud. This makes it much easier to switch workloads between different AI cloud vendors.
This brings us to the third difference: competitive intensity. Given the sophistication of the customers and the commoditized nature of the services being offered, we see a much more competitive environment for providing AI cloud services. AWS, Azure, and Google Cloud are facing significant competition from Meta, Oracle, SpaceX, CoreWeave, Nebius, and many others.
To date, the amount of capital being invested into the AI cloud market dwarfs that invested in general purpose cloud. So we believe it is reasonable to assume that revenue from providing AI cloud services will be much larger. However, it also looks apparent that given the economic characteristics outlined above, returns on capital for AI cloud infrastructure providers will be significantly lower than they were for general purpose cloud. Furthermore, history suggests that whenever this many competitors invest aggressively into a growth market, there is a high likelihood that the industry will overinvest, potentially massively, and that the returns of many participants will not even cover their cost of capital for some time.
The hyperscalers have built some of the best business models the world has ever seen, Google Search, Microsoft Office and Azure, AWS, and Facebook and Instagram among them. These are high return on equity, incredibly profitable businesses that face very few competitors in their core markets. As outlined above, the economics of the AI Cloud business look far more competitive and far lower return than the hyperscalers’ existing businesses. Despite this, these companies are now committing capital to AI Cloud at a pace exceeding their own free cash flow generation.
The hyperscalers’ core, more mature businesses continue to thrive. And there is little doubt that these companies are finding opportunities to leverage AI to extend these franchises. But any investor trying to value a hyperscaler should recognize that a significant and growing portion of these companies’ capital expenditure and therefore of their revenue and earnings growth, is now tied to this competitive Cloud AI business, whose long-term sustainability and profitability remain unclear.