How does the AI capex cycle pay for itself?

Dominating investment headlines these days is the price tag of the AI capital investment cycle. The attention is understandable given that, as soon as next year, the annual dollar amount allocated to the buildout is expected to be measured in the trillions.

A question at the forefront of investors’ minds is: Is it worth it? Put in more precise financial parlance: Will companies ultimately generate returns sufficient to justify this level of investment?

To address this, the Janus Henderson Global Technology and Innovation team conducted an analysis by inserting the highly publicized consensus levels of capex into models to reverse engineer the amount of revenue needed to generate the returns we believe tech companies – and by extension, the investment community – would find acceptable to justify this capital commitment. Assumptions were made based on management commentary, our industry expertise, and historical precedent.

A top-down approach

With much of the discussion involving aggregate capex figures, we first applied a top-down analysis based on the following assumptions:

• Much of the investment is being undertaken by megacap hyperscalers and AI-centric neocloud
• We believe return on invested capital (ROIC) is a powerful metric, as it holds management accountable with respect to its ability to identify worthwhile projects and allocate capital at appropriate levels.
• For the purposes of this exercise, we assumed management teams would seek to earn 15% ROIC on AI capex – a level that we believe is largely being achieved at present.
• Importantly, ROIC at 15%, is less than capex returns generated prior to the AI investment cycle, as companies may be willing to accept lower returns – at least for a period of time – given the view that these initiatives are existential to the technology ecosystem.
• We also assume that longer-term operating margins of 30%, roughly in line with software-as-as-service (Saas) companies, would be acceptable. We assume AI will have a higher cost of goods sold (COGS) but lower operating expenses given AI’s automation and scale.
• We recognize that such a high-level model will incorporate some double counting, as much of the hyperscalers’ revenue is also the COGS of frontier labs – the other major players in this ecosystem. Our models compensates for that.

Based on these assumptions, the table below indicates the annual revenue that must be generated to justify this capex cycle.

A few takeaways:

• After reaching over $700 billion this year, AI capex is expected to top $1 trillion in 2027 and climb toward $2 trillion over the next decade. Note, we assume roughly Wall Street consensus capex estimates through 2030. Beyond 2030, we assume capex grows roughly 3% to 4% per year.
• Under current assumptions, AI-related revenue would need to rise to well over $2 trillion by 2030 and to roughly $4 trillion over the next decade.
• After a massive downpayment, 2026 marks the last year that AI-capex growth should rise faster than operating cash flow growth.
• As capex reaches the trillion-dollar level, 2026 represents the last year that hyperscalers, in aggregate, can fund capex growth solely through cash flows, with the upshot being a greater reliance on debt capital markets to make up the difference – something already occurring with companies not possessing massive Internet, software, or advertising cash engines.
• Relatedly, the accumulated level of capex means that AI-related revenue will have to sustain a higher level of growth than capex growth.
• Given the view that getting AI right is not a choice but possibly existential, companies may be willing to accept a lower ROIC. At 10% for example, required annual revenues could be 33% less than those projected in our baseline scenario. These are choices managers – and again, by extension, investors – will need to make as the scenario unfolds.
• The same assumptions hold true (less needed revenue) should companies be able to wring higher than 30% operating margins out of AI businesses.

And from the bottom up

Our company-level analysis is based on a combination of internal and consensus forecasts. Many of the assumptions from the above analysis hold, especially with respect to the intertwined relationships of – and thus double counting between – hyperscalers and frontier labs. We want to emphasize these are rough estimates as we are still very early in the adoption cycle and have limited visibility into many revenue categories.

Assumptions and findings here include:

• In addition to hyperscalers, neoclouds, and frontier labs, our bottom-up analysis also includes open-weight models, as well as sovereign, enterprise, and physical AI.
• Our calculations are partly based on expected AI-capex revenue yield, which has been depressed thus far but is widely anticipated to improve over time.
• As seen below, in each of the 12 years covered, bottom-up AI-derived revenue would fall short of what is required in the top-down, 15% ROIC/30% operating margin analysis.
• If companies accept ROIC of 10%, or operating margins rise to 40%, the bottom-up revenue would be sufficient to justify current capex expectations.

Reconciling the difference

The difference between the models and the sensitivity of the analysis to small changes in assumptions illustrates why the market tends to react in such a volatile way to reports of frontier lab revenue, changes in interest rates, and financing conditions for neoclouds. Such a large difference between the two models may seem concerning, but there are a number of factors that could bridge the gap, aside from accepting lower returns:

• We believe sovereign governments view AI as essential to their economic health and national security. As a result, this spending is not ROIC driven.
• Much of the AI investment is frontloaded, while capacity – and thus revenue – lags as data centers are built and come online. Once that compute capacity is added and scale and operating leverage materialize, returns could exceed expectations.
• Company-level analysis, in our opinion, will be critical, as there will likely be a profound difference between cloud providers that have a frontier model (or near-frontier capabilities) and those that don’t.
• The latter category could potentially command a lower multiple as these companies may be figuratively consigned to the status of an unregulated utility.
• Company-level cash flows and balance sheets are also likely to be a key differentiator, as those reliant upon debt financing will have less latitude to accept a wide variance in returns.

A fixed income perspective

In addition to generating acceptable ROIC, corporate managers must also think about how they will fund the massive capex requirements. The Janus Henderson fixed income team has analyzed and commented on the potential impact of financing the AI buildout within the public debt markets. While our team believes the investment-grade public markets will be able to absorb the expected capex-related funding needs, and cost of debt financing is likely to remain below ROIC levels discussed earlier, we continue to think spreads must move wider to allow the market to clear the amount of expected debt issuance.

–Mike Talaga, CFA, Global Head of Credit Research

What to watch for next

Another variable – and a major one at that – that investors must monitor is whether an even greater level of capex on the part of hyperscalers, frontier labs, and neocloud providers will be required over the remainder of decade. Given current macro and financing conditions, this could very well be the case. Should that materialize, even more AI-related revenue would be required to meet ROIC and margin hurdles.

With capex levels of this magnitude and commensurate revenue and return requirements, investors would likely welcome some form of incremental validation of the AI infrastructure thesis. Over the next 12 months, they may get two, in the form of the public listings of frontier labs OpenAI and Anthropic. Greater visibility into these companies’ investments, business models, and end markets may assuage the sticker shock felt by investors when dealing with dollar amounts typically associated with national economies. Not only would successful listings represent a vote of confidence, they could also inject additional capital into the AI ecosystem, thus introducing perhaps a wider band of tolerance around return hurdles.

It remains early days. As AI is deployed across a larger part of the global economy, we expect greater visibility into the revenues, margins, and returns generated with each successive earnings period. As stated, the assumptions and conclusions in this iteration of our revenue model are only a starting point. A company’s mix of AI offerings, investment discipline, and balance sheet management will all be key factors in identifying which businesses are most capable of delivering on their commitments and which ones may fall short.

IMPORTANT INFORMATION

Artificial intelligence (“AI”) focused companies, including those that develop or utilize AI technologies, may face rapid product obsolescence, intense competition, and increased regulatory scrutiny. These companies often rely heavily on intellectual property, invest significantly in research and development, and depend on maintaining and growing consumer demand. Their securities may be more volatile than those of companies offering more established technologies and may be affected by risks tied to the use of AI in business operations, including legal liability or reputational harm.

Technology industries can be significantly affected by obsolescence of existing technology, short product cycles, falling prices and profits, competition from new market entrants, and general economic conditions. A concentrated investment in a single industry could be more volatile than the performance of less concentrated investments and the market as a whole.