Can AI answer the trillion question?
Three years ago, David Cahn from Sequoia was among the early analysts to quantify the immense financial implications of Silicon Valley’s massive investment in AI infrastructure. In 2023, reacting to Nvidia’s reported annual GPU revenue of $50 billion, he extrapolated that, considering the operational costs of data centers and their operator margins, a total revenue of $200 billion would be necessary to recoup the initial investment. Viewing this as a challenge, he urged entrepreneurs to develop AI products and services that would effectively utilize and monetize the extensive infrastructure. Fast-forward to today, after three years of rapid expansion, Cahn has revised his estimate for AI infrastructure spending in 2026 to $1.5 trillion. By his calculations, the AI sector will need to generate about $3 trillion in revenue to justify the expenditures made on chips and data centers. He also suggests that this figure is likely conservative, pointing to rising memory costs and a growing reliance on specialized or inference-specific chips as factors that will push the needed revenue higher. He notes that recently, the revenue required per gigawatt of capital expenditure has surged due to bottleneck challenges and climbing construction costs.
On the income side, Anthropic is believed to have reached $60 billion in annual recurring revenue, while OpenAI reportedly generated $13 billion in 2025, with a later statement indicating $20 billion ARR as of November that year and likely higher figures this year. Despite these impressive numbers, a substantial revenue gap remains to be closed. Torsten Slok, chief economist at Apollo, a leading asset management firm, highlights that the major hyperscalers — Google, Meta, Microsoft, and Amazon — are forecasting significant increases in free cash flow by 2028, anticipating returns on their substantial chip investments. However, he raises concerns about potential risks if these financial goals are not met. Current trends show more organizations opting for less expensive open weight models, often developed in China, instead of those from the leading labs. Concurrently, token prices overall are decreasing. OpenAI’s latest model, as stated by CEO Sam Altman, boasts a 54% increase in token efficiency for coding tasks. While this benefits users worried about AI operational costs, it could negatively impact companies that depend on volume-based token usage if users do not substantially increase their overall consumption.
Slok warns that if the hyperscalers fail to achieve their cash flow projections, the market impact could be severe. Given the concentration of market value in just a few companies, a slower-than-expected financial return on AI investments might not only affect the sector but also threaten to push the wider economy into recession and trigger a correction in the S&P 500. This scenario serves as a reminder to carefully consider the economic implications as AI usage continues to shift toward more cost-effective token strategies.
