Executive Summary
Disclaimer: This piece was written before the recent news involving NVIDIA CEO Jensen Huang, which was disclosed on November 8, 2026. Huang resigned from his role as CEO on November 13, leaving the future of Nvidia, and much of the AI sector, in turmoil.
While the developments at Nvidia likely have broader geopolitical implications, they also underscore two themes in our essay: first, that this cycle remains highly unpredictable and may turn on idiosyncratic events, requiring investors to remain nimble; and second, that the key players in the AI infrastructure buildout (semiconductor vendors, memory makers, power utilities, hyperscalers, neoclouds, and model labs) are all exceptionally interconnected so that a break in any link in the value chain is likely to have serious ramifications for all.
The largest technology companies are making a bet that increasingly requires the rest of the economy to cooperate. Their data centers need electricity, equipment, construction workers, and financing now. The productivity gains that would justify the expense will take longer. Meanwhile, customers are learning to buy intelligence more cheaply.
The five largest American hyperscalers, the large cloud-computing companies, are expected to spend roughly $800 billion on capital equipment this year and $1.2 trillion next year. Historical comparisons depend on what is counted, but the announced AI buildout belongs in the same discussion as the railroads. One difference is that a railway could remain useful for generations. Much of the equipment in a data center may need replacing before the customers have finished figuring out what to do with it.
The change since our papers last year is the growing reliance on debt. Borrowing allows construction to proceed faster than internally generated cash would permit. It also introduces creditors, refinancing dates, and losses that can spread beyond shareholders. The overlapping investments and contracts among chipmakers, cloud providers, and model laboratories make those exposures harder to separate.
The revenue needed to justify this spending is formidable. One influential calculation puts annual US compute-rental revenue needed to justify the planned infrastructure at roughly $3.7 trillion by 2032, or 9% of projected GDP. That is a return-on-investment calculation, rather than a forecast of debt payments, but even more accommodating assumptions leave a revenue requirement measured in trillions.
Our argument is that:
- Significant productivity gains are needed to support both the debt-funded capex binge and the broader economy’s ability to absorb it. The evidence is encouraging in some tasks and sectors, much less so across the typical firm.
- The gains may reach customers faster than they reach the investors financing the buildout. Competition, smaller models, and open alternatives can expand AI use while limiting what providers can charge.
- Borrowing and construction add demand before productivity adds supply. That timing helps explain why a promising technology can raise the cost of capital and contribute to inflationary pressure during its installation.
- The likely constraints are financial, physical, and commercial: more expensive funding, difficult power connections, public resistance, and customers unwilling to pay frontier prices for routine tasks.
- Cheaper, smaller and open-weight models offer customers a cheaper route. They favor users of intelligence over sellers of capacity and penalize valuations that extrapolate the current slope of spending.
AI need not disappoint as a technology for some of its investments to disappoint. The question for investors is who can earn an adequate return, at what price, and before which bill comes due.
Read the full whitepaper here.










