The $5 Trillion Bet

Samuel Ziff
Samuel Ziff
Partner | Portfolio Manager

Samuel Ziff

Partner | Portfolio Manager

Sam Ziff joined OP in April 2013. He was previously employed by J.P. Morgan Cazenove working in the UK Industrials Corporate Finance team for a total of 4 years. He graduated from Oxford University. He is CIO, manages the global equity portfolios, and contributes to the overall investment selection.

Samuel Ziff

PDF

Summary

AI is a general-purpose technology that will lead to extraordinary productivity gains and reshape the economic system. This we do not doubt — it helped research and write this piece, and even served as a test bed for some of the logical reasoning. However, we do doubt the economics of some of the investments.

The risk of capital misallocation looks as high as anything we have seen since 2008, and the scale is large enough that spillover into the broader economy is a material concern. The bet is not whether AI matters; it is whether $5trn of infrastructure investment can earn an acceptable return for those making an investment. The evidence is looking thinner than valuations and expectations suggest.

The Spending Boom?

Hyperscaler capital expenditure has gone vertical. The six big spenders collectively spent $233bn in 2024. 2026 estimates have this more than tripling to $722bn, rising to roughly $1trn per year from 2027 onward, or ~$5trn cumulatively over the next five years.

To justify $5trn of capital expenditure at a 10% return1, at a 35% EBITDA margin, the same companies need to generate roughly $2.8trn of annual revenue by 2030. That is a 7x increase over today’s c.$400bn cloud revenue base, or ~48% CAGR sustained for five years. Google Cloud grew 63% last quarter, and Amazon’s AI run-rate is growing in triple digits. They are executing, but blended growth across the three majors is currently closer to 35%.

For context, $2.8trn of annual revenue is nearly twice China’s best-ever year of organic GDP creation ($1.6trn in 2018) and comparable to Italy’s entire annual GDP today (~$2.5trn). Hyperscalers are essentially betting they can manufacture, in five years, a revenue stream the size of a major G7 economy.

1Assume assets are depreciated over 8 years is optimistic compared to the current assumptions of 5-6 years for semiconductors.

What Does This All Mean?

Hyperscaler capex is now roughly 30% of all US business investment. Even if AI delivers on its GPT promise over 20 years, the capex being spent between 2026 and 2030 needs heroic revenue and margin assumptions to make sense. This mismatch has the risk of producing a boom and then a bust, even when the underlying thesis turns out to be right.

We continue to believe that AI is a game-changer, but the benefits will play out over decades, not years. Past general-purpose technologies tell a clear story: the infrastructure layer rarely captures most of the value it enables. Railways, electricity, and fibre optic networks all reshaped the economy but bankrupted most of the companies that built them and created booms and busts in the process.

Application companies (most of which are morphing into hyperscalers) captured the long-term economics from the web; Claude may be doing something similar today. Picks and shovels (think memory) are also capturing value. The infrastructure layer in the middle, where the hyperscalers sit, is where capital is most likely to be destroyed. The fibre operators of the late 1990s were right about internet demand, and Victorian Britain got the railways. The shareholders got wiped out.

 

Important information

OP