The Arithmetic of the AI Capex Boom

The four dominant US hyperscalers (Microsoft, Alphabet, Meta, and Amazon) and Oracle which provide the majority of the West’s AI infrastructure have made an aggregate three-year AI infrastructure investment commitment expected to exceed $1.4 trillion.  To borrow a metaphor from physics, the AI capex boom is starting to feel like a firestorm: a thermal column so intense that it pulls oxygen in from miles around, depriving the surrounding forest of the air it needs to keep burning.

At Ptarmigan Capital, we don’t try to make major macro-economic forecasts, but we do use market feedback to test our investment assumptions to help identify where we might be wrong.  Given the rally in semiconductor company share prices, we have asked ourselves two connected questions and used simple “back of the envelope” calculations to test our position:

1)  What would hyperscalers (and Oracle) need to charge for AI to earn a 15% return on the capex of the last three years and the next three years?

2) What portion of the labour force needs to be replaced to sustain current AI expectations?

The Required Revenue: A 15% Return on Investment Hurdle

Given the Big Four hyperscalers spent roughly $780 billion between 2023 and 2025 (approximately $150 billion in 2023, $230 billion in 2024, and $400 billion in 2025), and the assumption of $700 billion per year for the next three years implies a further $2.1 trillion in capex.  This assumption is, if anything, conservative: 2026 capex guidance from the same four companies has already been raised to roughly $725 billion, and Alphabet has explicitly said 2027 capex will be “significantly” higher.

Using the $700 billion-per-year assumption gives a cumulative invested capital base of approximately $2.9 trillion by the end of 2028.  A 15% return on that investment requires roughly $430 billion of net operating profit after tax, every year, in perpetuity.  (Our choice of 15% may be generous or conservative, but we chose it because it sits slightly above their estimated long-term blended cost-of-capital but well below the 25–30% returns on invested capital that their legacy businesses produce.)

Translating this profit requirement into a revenue line requires three further assumptions:

• Depreciation.  Microsoft has just extended useful lives on AI-server hardware to six years. A blended six-year straight-line schedule on a $2.9 trillion fleet at steady state implies annual depreciation of roughly $480 billion.

• Cash operating costs.  Power, real estate, networking, model licensing, and staff at hyperscale data centres typically run at 30–40% of revenue; this is consistent with mature Amazon Web Service-style economics.  We have assumed 35%.

• Tax.  Use 21% to convert pre-tax to after-tax.

Solving these inputs gives a required pre-tax operating profit of $430 billion divided by (1 – 0.21), or approximately $545 billion.  Adding back depreciation gives required gross profit of $545 billion plus $480 billion, or $1,025 billion, and grossing up for cash opex (35% of revenue) implies a required revenue line of roughly $1.6 trillion per year, in perpetuity, from AI compute and AI-tied cloud services.  Currently, we estimate the current annual run-rate of AI services revenue is about $50 billion.

The GDP Share: An Industrial-Era Footprint

To help size this, we believe current hyperscaler cloud revenue from all sources is about $450 billion, so our estimate means the AI industry needs to grow to around four times the size of the current cloud market.  Using the US Bureau of Economic Analysis Value Added data (gross profit + employee wages + net production taxes), this is about 20% larger than the whole US construction industry, six times larger than the US agriculture industry, and five times larger than the US oil & gas and mining industries combined.  Alternatively, it equates to about 5.0% of US 2025 nominal GDP of $30.8 trillion.

Put differently, customers (enterprises, governments, and consumers via subscription) need to be willing and able to redirect roughly 5% of US GDP per year, every year, into AI compute spending.  This is larger than collective corporate IT budgets, which means either labour must be displaced, or AI must deliver substantial revenue growth for corporates.  While this is not impossible, the dotcom-era internet buildout did eventually deliver revenue of this magnitude, it took 15+ years to arrive, rather than five.

The Wage Bill Ceiling: What Can Enterprises Actually Pay?

As hinted above, one company’s revenue is another customers’ cost. If we assume the AI industry generates $1.6 trillion in revenue, it must come from somewhere else. The obvious source of funding is labour, particularly the knowledge workers where AI expected to augment or replace.

Using US Bureau of Labor Statistics data (May 2024 Occupational Employment and Wage Statistics) for occupations most directly threatened by generative AI (software developers, accountants, lawyers, paralegals, financial analysts, management consultants, marketing analysts, customer service representatives, bookkeepers, secretaries, and so on) produces approximately 26.7 million workers earning a combined wage bill of roughly $2.2 trillion per year. Adding the standard 30% benefits and payroll-tax burden brings fully-loaded compensation to approximately $2.8 trillion.

This means that AI revenue of $1.6 trillion is approximately 56% of the US white collar wage bill.  For a company to switch from a human worker to AI which will come with the associated risks of switching any provider (supervision risk, integration risk, error monitoring, reputational risk), AI probably needs to be noticeably cheaper than the human worker for the same output, or much more productive for the same cost.  Given this, a 15% return on capital for the AI providers looks challenging.

So is the Market Pricing the AI Boom Correctly?

Our view is that the market is pricing the AI boom with confidence in three things, two of which are reasonable and one of which is rather optimistic.  Firstly, the technology clearly is transformational, and secondly the major hyperscalers have demonstrated execution capability with sufficient customer demand to build a meaningful revenue stream.  However, the assumption that AI revenue could grow from ~$50 billion currently to $1.6 trillion per year by 2030 feels stretched.

Of course, there is a more optimistic outcome where AI does not substitute labour on a one-for-one basis but instead develops entirely new categories of work creating markets which do not currently exist, as has been the case for other transformative technologies.  This could grow the addressable market well beyond the assumptions we have considered.  Nevertheless, the pace of growth remains an issue and delivering this fast enough to satisfy market expectations may not be possible.

As always at Ptarmigan Capital we are not predicting any of this, but we are observing that the market currently appears to be assigning a probability close to 100% that the bull case plays out, and a probability close to 0% to a path in which AI is real but slower than the capex schedule.  That asymmetry both concerns and excites us.

CDAJ, EFJR, TMNF

Risk Warning & Disclaimer

This article is provided for general information purposes only and is not intended to constitute investment advice, investment research, or a personal recommendation. It does not take into account the investment objectives, financial situation, or particular needs of any individual or entity.

The views expressed are those of Ptarmigan Capital as at the date of publication and are subject to change without notice. References to investment concepts, asset classes, or portfolio construction approaches are included for illustrative purposes only and should not be construed as a recommendation or a solicitation to buy or sell any security or financial instrument.

Drafting tools, including artificial intelligence and automated systems, may be used as part of our content production process, however all materials are subject to human review and approval.

The value of investments and any income derived from them may fall as well as rise, and investors may receive back less than they originally invested. Past performance is not a reliable indicator of future results. Where investments involve overseas assets, returns may be affected by movements in exchange rates.

Ptarmigan Capital Limited is an employee-owned investment management firm providing discretionary investment management services to private clients, trusts, charities, and family offices. This article is intended to explain the firm’s perspective on bespoke investment management and the factors some clients consider when choosing such an approach. It does not relate to any specific investment product or strategy.

Ptarmigan Capital Limited is authorised and regulated by the Financial Conduct Authority (FRN 940407). Registered in England and Wales (Company No. 12715470). Registered office: 17 Cavendish Square, London W1G 0PH.