David Cahn, a partner at Sequoia, calculated three years ago that Nvidia’s $50 billion GPU revenue in 2023 implied a need for $200 billion in total industry revenue to cover infrastructure costs. By 2026, Cahn estimates that spending on AI infrastructure has reached $1.5 trillion, meaning the sector must generate $3 trillion in earnings to justify the investment.
The figure is likely an underestimate. Rising memory costs and the deployment of exotic inference-specific chips are driving up the required revenue per gigawatt of capital expenditure. Cahn notes that construction costs have also increased sharply.
Current revenue does not match these projections. Anthropic is believed to have hit $60 billion in annual recurring revenue. OpenAI reported $13 billion for 2025, though the company stated in November 2025 that its ARR had reached $20 billion. The gap between current earnings and the required $3 trillion remains vast.
The risk
Torsten Slok, chief economist at Apollo, warns that hyperscalers Google, Meta, Microsoft and Amazon predict massive free-cash flow accelerations by 2028. They expect to see returns from their chip purchases then.
Slok points to a risk where more organisations adopt cheaper open-weight models, often from Chinese providers, rather than those from frontier labs. Token prices are falling. OpenAI CEO Sam Altman states the latest model is 54% more token efficient on coding tasks. This helps users managing agent costs, but it may hurt companies building token factories if users do not increase overall usage.
Slok argues that if hyperscalers miss their cash flow goals, the market reaction could be severe. With so much riding on so few names, a slower payoff would not just be a sector problem. It risks tipping the economy into recession and the S&P 500 into a correction.
What it means
Businesses deploying AI agents must consider the broader economic implications of moving toward cheaper tokens. If the underlying infrastructure does not pay for itself, the stability of the wider market is at risk.
