AI Profit Pools and Shifting Sands

It is impossible to ignore the scale of capital being deployed into artificial intelligence. AI has dominated headlines for the past two years, not just for its technological advancements but also for the sheer magnitude of financial commitment being made to support its development. The most striking example of this to date is the announcement by OpenAI, Oracle, and SoftBank of their Stargate Project, a massive AI infrastructure initiative projected to cost up to $500 billion over the next four years. This staggering commitment was soon followed by multiple hyper-scalers announcing further increases in capital expenditure guidance, signalling an even broader wave of investment into AI infrastructure.

Such unprecedented levels of capital deployment is predicated on the expectation that AI will enable revolutionary new products and services, potentially reshaping entire industries. However, this rapid spending raises the fundamental question of who will ultimately capture the long-term economic value of these investments. Beyond that, the question of whether the magnitude of these investments is even justified given the uncertainties surrounding AI’s future capabilities also remains open. The hyper-clusters are being built, but we don’t know what the models they will train will be capable of. We also don’t know how likely it is that they will be adopted, nor the speed at which they will be adopted by the market, let alone how much they will cost to run. This level of uncertainty is fairly unprecedented in light of the magnitude of the investments. AI’s full potential remains unclear.

Despite this, real money is being spent (by the hyper-scalers like Microsoft, Alphabet and Amazon) —and real money is being made (by those providing the equipment for the super-clusters that the hyper-scalers are building). Companies involved in supplying the buildout of the AI infrastructure are already seeing massive revenue inflows. The speed at which capital is being deployed is astonishing, making this one of the largest capex booms in modern history. Against this backdrop, we ask ourselves whose problems these models are going to solve. Unlike previous technological shifts, where future applications were easier to predict, AI presents an unusually wide range of possible outcomes. Future models could be incrementally integrated into existing software products, acting as assistants or efficiency tools. Alternatively, some argue that AI super-agents could be so powerful that they ultimately replace the entire software ecosystems, making traditional applications obsolete.

The second scenario would certainly be highly disruptive, potentially bringing about the demise of the Software-as-a-Service (SaaS) business model. Instead of paying for multiple software subscriptions, users might interact with a single AI agent capable of handling everything from customer service to enterprise resource planning. Framing the question this way is too conceptual to be practical though at this point in time and given what we currently know. A more useful way to approach the problem is by asking ourselves whose problem AI is likely going to solve. Is it a CEO’s strategic tool? A revenue generator? A cost-cutting enabler? Will it simply enhance existing operations, or will it completely transform entire industries?

For now, these questions remain open. But what we do know is that, in its current state, AI is a long way from replacing SaaS models. While AI is still in its early commercialization phase, it is clear that the companies making money today are not the ones developing AI applications or the companies running the super-clusters. Instead, the primary beneficiaries are those providing the infrastructure needed to build and train the foundation models. Prominent examples of companies benefiting from the AI infrastructure buildout include Nvidia, which designs the powerful chips that drive AI systems; TSMC, which manufactures these chips for many of the world’s leading tech firms; ASML, which provides the advanced machines needed to produce the most sophisticated semiconductors; and SK Hynix and Micron, which supply the memory chips essential for storing and processing AI data. New growth opportunities are even emerging for energy and utility providers as the training of the future models and the running of these huge hyper-clusters will require enormous energy inputs, benefiting electricity producers and specialized data centre operators.

These firms are experiencing explosive revenue growth. However, history suggests that their dominance may not last forever. To understand why, it is useful to examine how profit pools have shifted in past technological revolutions. Profit pools refer to the distribution of profits within an industry, which shift from one part of the value chain to another as industries mature. Historically, early-stage profits in major technological booms have accrued to those companies that have taken part in building out the infrastructure. But over the long run, economic value has tended to migrate toward applications and services.

During the nineteenth-century railroad boom, initial profits flowed to railroad builders and locomotive manufacturers. As the infrastructure matured, dominant profit pools shifted instead to logistics and retail chains such as Sears, whose mail order business took advantage of the new national supply networks. In the 1990s and early 2000s, the telecom industry saw early winners in fibre-optic network builders like Nortel and Lucent. Over time, however, the value shifted to mobile ecosystems such as Apple and Google, as well as digital platforms like Facebook and AWS. A similar pattern unfolded in cloud computing, where the early boom benefited data centre operators and hardware manufacturers such as Intel. Yet today, the biggest winners are SaaS providers like Salesforce, Adobe, OpenAI, and Databricks. Ultimately, in all these cases, not all segments of the value chain experience the same profit dynamics. Infrastructure profits tended to be cyclical while those of the application layer were more stable and predictable. If AI infrastructure follows the example of these other technological transformations and is built too rapidly, future demand may not justify current levels of investment, leading to industry-wide overcapacity. Whether this happens will ultimately depend on the actual adoption rates of AI technology, which remain a major unknown.

While you can make a strong case for why it won’t be those supplying the infrastructure buildout who will ultimately be reaping the profits, it is harder to say with any degree of certainty who will. As things stand today, it is companies like Adobe and Microsoft that seem to be best positioned. The reason for this is well highlighted in my colleague Nina’s newsletter. While having a great product is important, it is the ability to get that product widely distributed in a seamless way that is crucial. Just as Google’s default status on browsers locked in users, Microsoft is integrating Copilot deeply into its Office suite. By controlling how users encounter AI capabilities, Microsoft could capture the lion’s share of AI-driven enterprise productivity markets.

AI is already being used by companies to drive both revenues and efficiencies, and this is something that we begin to see in our own investee companies, which are probably more likely to be early adopters of the technology given their large services exposure. Despite the hype, however, the magnitude of this contribution from AI to anyone but the infrastructure players is still small, and nowhere near enough to justify the scale of investment that we are seeing.

Nonetheless, ultimately these investments are being made because investors expect returns. AI infrastructure is not an end in itself—it exists to support future products that can be sold to customers. Here the range of possible outcomes is vast. Large Language Models may remain just another software feature, integrating into existing applications and complementing SaaS rather than replacing it. Alternatively, they could become the dominant software paradigm, rendering traditional software obsolete and reshaping the software industry along with multitudes of others. Nobody knows the answer to this question, but the money keeps flowing. Investors must ask themselves whether AI infrastructure investments will follow past tech booms, leading to an eventual migration of profits to software and services, or whether this time will be different, with infrastructure providers maintaining long-term dominance. While AI’s future remains uncertain, history suggests that today’s infrastructure winners may not be the long-term beneficiaries. The real challenge is predicting where AI’s ultimate profit pools will emerge and positioning accordingly.