AI as an investment cycle: what needs to go right

The AI investment cycle can create lasting value only if rapidly rising compute demand converts into durable revenue and returns before falling prices and efficiency gains commoditize the infrastructure.

• Hyperscalers are expected to spend more than $5 trillion on capital expenditures between 2024 and 2030, while five companies—Alphabet, Amazon, Meta, Microsoft, and Oracle—control roughly two-thirds of global AI compute capacity.
• Boston Partners estimates that a one-gigawatt data center costing at least $50 billion would need $15.5 billion to $18.6 billion of annual revenue to deliver a 10% to 20% equity ROE, after more than $12 billion of yearly operating, financing, and depreciation costs.
• If total AI infrastructure spending reaches $3 trillion to $4 trillion annually by the end of the decade, recapturing 31% to 37% of deployed capital implies nearly $1.5 trillion of annual revenue for the energy, chip, and infrastructure layers alone.
• Usage and early revenues are growing rapidly: OpenAI and Anthropic have reported sharply higher annualized recurring revenue, but the wireless industry shows that a 17-fold increase in usage can coexist with a 44% decline in average revenue per user.
• Nvidia expects its Rubin generation to deliver 35 times the throughput of Blackwell, while open-source model adoption and rising competition could reduce pricing power at the chip, model, and application layers.

Takeaway: Investors should remain constructive on AI’s long-term potential but demand disciplined valuations and selective exposure, because compute demand may not translate into durable margins and value may accrue to companies that capture the changing economics rather than those closest to the technology.