Three years ago, in 2023, Sequoia partner David Cahn was among the first to quantify the implications of Silicon Valley’s large-scale AI infrastructure spending. Reacting to Nvidia’s reported $50 billion annual GPU revenue, he argued that once operating costs for data centers and operators’ margins are included, roughly $200 billion in revenue would be required to pay back the upfront investment.
Cahn framed that as a challenge to entrepreneurs to build AI products and services that would use and monetize the infrastructure. Counting three years of hyperscaling since then, he now gives a new estimate for AI infrastructure spending in 2026: $1.5 trillion.
Overall, Cahn calculates the AI industry will need to earn about $3 trillion to justify those chips and other data-center expenditures. He also warned this is likely a conservative estimate: rising memory prices and growing use of exotic or inference-specific chips could push the required revenue higher. “Recently,” he wrote, “the required revenue per GW of CapEx has sharply increased due to these bottleneck dynamics and rising costs of construction.”
On the revenue side, some large companies have reported sizable numbers. Anthropic is thought to have reached $60 billion in annual recurring revenue (ARR), while OpenAI was reported to have earned $13 billion in 2025 — although in November 2025 OpenAI said it was at $20 billion ARR — and is presumably earning more this year. Still, a substantial gap remains relative to the roughly $3 trillion target.
Torsten Slok, chief economist at Apollo, is another analyst monitoring the mismatch. In a recent note he observed that the hyperscalers — Google, Meta, Microsoft and Amazon — are forecasting large accelerations in free cash flow in 2028. In other words, they expect the chip purchases to pay off around that time.
Slok also highlighted a current risk in AI usage: more organizations are adopting cheaper open-weight models, often developed in China rather than by frontier labs, and token prices are generally falling. OpenAI’s latest model, according to CEO Sam Altman, is 54% more token efficient on coding tasks. That efficiency helps users concerned about agent costs, but it could hurt companies that were counting on token consumption to generate revenue if overall token use doesn’t rise sharply.
Slok warned that if hyperscalers fail to meet their cash-flow targets the market reaction could be severe. “With so much riding on so few names,” he wrote, “a slower payoff wouldn’t just be a sector problem, it would risk tipping the economy into recession and the S&P 500 into a correction.”
That is a notable risk to keep in mind as organizations shift their AI agents toward cheaper tokens.
