Debating AI capex trajectories
Sahil Mahtani – Head of Macro Research at the Investment Institute
This article is based on research presented by equity specialists William Nott, Paul Vincent, Anton du Plooy and Niall Hartnett at a recent Ninety One thematic meeting.
With hundreds of billions in capital expenditure flowing into data centres, investors naturally want to know if the spending will pay off. Venture capital firm Sequoia recently called this a US$600 billion question, referring to the gap between fast-growing AI capex and the less certain AI revenue stream that will be generated from it. But at the pace that AI capex is rising, this could easily be framed as a US$1 trillion problem. The key questions are: Will investments in AI models pay off? What valuations are the current AI leaders discounting? What scenarios might accommodate different implied valuations.
The evolution of technology investment
To assess whether the surge in capital expenditure is justified, we should consider that this isn’t the first rodeo for large capex cycles driven by innovative technology. New technology tipping points have frequently been preceded by a multi-year capex buildout. In the late 1990s and early 2000s, a significant capex cycle focused on building fibre optic networks, creating the capacity that later fuelled the explosive growth in internet services, smartphones, and cloud computing from the late 2000s onwards.
Similarly, the mainframe computing era that began in the 1960s required significant capital expenditure, not just on centralised computer systems to control and process data, but also on supporting infrastructure like data centres, air conditioning, generators, and power systems. In each of these cases, early adopters thrived, while others lagged, unwilling to cannibalise their existing businesses - a concept famously captured in The Innovator’s Dilemma1 .
1 The Innovator’s Dilemma, introduced by Clayton Christensen in his 1997 book of the same name, explains how companies can fail by focusing on existing customers and products, allowing disruptive innovations to overtake them in the long run.
Nvidia: is the growth sustainable?
Nvidia embodies the questions around AI most clearly, having achieved remarkable growth over the past decade. Its growth, divided into two clear phases, is illustrative:
From 2015 to 2020, Nvidia rode the first major AI wave, capturing 90% of the AI accelerator market in data centres as deep learning took off. During this time, its revenue jumped nearly tenfold, driven by tech giants investing heavily in data centres. By 2020, Nvidia’s AI accelerator revenue hit US$3 billion—impressive, but still a small part of the data centre market.
Since 2020, Nvidia’s data centre revenue has exploded, growing more than 35 times as demand for AI infrastructure surged post-pandemic. This shift has transformed Nvidia into the biggest player in the US$100 billion data centre space.
Given the incredible pace of Nvidia’s growth, the question arises: is this momentum sustainable, or a sign of market exuberance.
Different capex estimates — who is right?
Nvidia’s place in the server market has grown as AI-related capital expenditures increased. By fiscal 2025, the company’s revenue is projected to reach US$110 billion - a figure driven largely by explosive AI adoption. Once a small component of data centre investments, graphics processing units (GPUs) have become crucial with the rise of AI workloads, tripling in data centre share since the pre-ChatGPT era.
Consensus projections2 imply that Nvidia’s data centre revenue could outstrip the total addressable market (TAM) for data centre infrastructure. Based on Nvidia’s consensus numbers, AI infrastructure spending is projected to reach US$516 billion by 2027, while hyperscalers like Microsoft, Google, Meta, and AWS are expected to generate only US$245 billion by the same year.
There are therefore three scenarios: a) both those numbers are right, and there is a good deal of capex to come from non-hyperscaler firms, implying a major broadening out of capex from beyond the tech giants and much greater fragmentation; or b) the Nvidia consensus numbers are wrong, with implications for Nvidia’s share price or; c) the hyperscaler consensus capex numbers are too low, with implications for capex/sales ratios, and potentially earnings as well, for the hyperscalers.
For the capex figures implied by Nvidia’s projections to be realised, hyperscalers will need to generate revenue or savings of US$600 billion to US$700 billion. To achieve this, their customers must see tangible benefits from AI. A recent example is Swedish payments company Klarna, one of the most progressive adopters of AI. AI assistants now handle two-thirds of Klarna’s customer service chats, performing the work equivalent to 700 full-time agents, with customer satisfaction scores matching those of human agents. If other companies can harness AI as effectively as Klarna, the US$600 billion challenge may soon resolve itself.
Infrastructure is key
The AI stack provides a framework to understand where capital is flowing within the technology sector. This stack includes three primary layers:
- Infrastructure layer: This attracts the most investment and includes the physical and cloud infrastructure that powers AI. Companies like Nvidia, with strong AI hardware positions, are experiencing outsized gains as they meet the computational demands of AI. Companies focused on data centres, GPUs, and other infrastructure components could capture significant investment interest.
- Software layer: Perhaps the most variable and dynamic area, software is where much of the alpha will be won or lost. Companies focused on AI software, like natural language processing, autonomous systems, and predictive analytics, will face intense competition and opportunity. Success here depends on their ability to deliver solutions that drive efficiencies for end-users across sectors.
- Application layer: This layer includes the diverse AI-driven applications transforming industries from healthcare and finance to logistics and retail, with potential winners and losers emerging as AI adoption continues. As this layer matures, investors could see more specialised plays rather than broad technology bets.
Judging by the numbers implied in the capex estimates earlier, this technology boom is likely to have higher capital spending and revenues in the infrastructure layer. In other words, the share of the pie taken by companies in the AI boom is likely to be larger than in the software boom of the 2010s.
Ripple effects across industries
The transformative potential of AI is not confined to technology firms alone. As AI expands, so do its implications for other industries, especially those linked to the growing need for data centres and the associated energy demands. Here are a few areas outside of tech that may offer unique investment opportunities:
Energy: The energy-guzzling requirements of AI-focused data centres are creating an urgent need for clean energy solutions. This has implications for utility providers, renewable energy companies, and even traditional energy firms working on efficiency technologies. As big tech firms strive for sustainable data centre operations, energy companies that offer scalable and renewable solutions are likely to benefit.
Industrials: Companies involved in the construction and maintenance of data centres, from real estate to specialised materials, could benefit. The expansion of data centres is fueling demand for industrial equipment, energy-efficient heating, ventilation, and air conditioning (HVAC) systems, and other infrastructure-related products.
Utilities: With big tech firms rapidly building new data centres, utility providers face increased demand for reliable energy and water sources. Companies that can meet these needs sustainably will be particularly attractive, as they align with the green mandates of tech firms and regulatory requirements.
Conclusion
The outlook for AI capital expenditure varies widely, reflecting the challenges of forecasting AI-related spending. What is clear, however, is that the infrastructure layer will remain central to investment and revenue generation. This marks a shift from past tech cycles, with a greater share of economic value likely to be captured by companies focused on infrastructure rather than purely on software.
Moreover, AI investment is expected to extend beyond the technology sector. Industries such as energy, industrials, and utilities are set to see increased spending as AI adoption spreads, creating a broader and more diverse impact. This expansion signals the deeper integration of AI technologies across the global economy.
2 www.visiblealpha.com.