#532: Why Are Market Fears About Spending On AI Infrastructure Misplaced?, & More

ARK argues that the market is underestimating the revenue potential of frontier AI and therefore overstating the risk from hyperscalers’ and AI labs’ compute commitments.

• Anthropic reported 85% gross margins and nearly $900 billion of committed compute expenditures, while OpenAI and Anthropic together had penetrated less than 0.4% of a knowledge-work market that ARK estimates could support $40 trillion of software revenue.
• Compute rental rates remain elevated, and ARK’s model finds that greater normalized compute capacity is associated with higher frontier-lab revenue; the estimated hardware-equivalent compute needed to serve the full market has fallen from 7.5 terawatts in 2023 to below 4 terawatts.
• ARK estimates that each cumulative doubling of effective compute devoted to research and development lowers the cost of delivering a 1,000% productivity gain to knowledge workers by 13%.
• The model projects combined OpenAI and Anthropic run-rate revenue of $678 billion in 2027, $2.146 trillion in 2028 and $14.872 trillion in 2031, although demand, utilization and monetization are assumed rather than observed.
• Alphabet’s Search revenue growth slowed from 17% year over year in the second quarter to 12% in the third, while paid clicks rose 8% and cost per click only 4%; Reddit’s revenue grew 53% even as search referrals fell by more than 1 million per week.

Takeaway: ARK believes compute capacity, rather than demand, is the main constraint on frontier AI revenue and that the infrastructure bill could be justified if its adoption and monetization assumptions hold; investors should also account for near-term advertising pressure and the risk that AI search reduces traffic to external websites.