The $3 Trillion AI Question and the Only Version of It You Can Answer

Key takeaways

• Cahn's updated math (via TechCrunch): $1.5 trillion of 2026 AI infrastructure capex implies about $3 trillion in required revenue, and rising memory and chip costs make that a likely underestimate.
• The revenue side is growing but far short: Anthropic is reported around $60B ARR; OpenAI reportedly earned $13B in 2025 and claimed $20B ARR in late 2025.
• Apollo's Torsten Slok warns hyperscalers all project a free-cash-flow payoff by 2028, and a slower payoff would be an economy-level problem, not just a sector one.
• The pressure valves, cheaper open-weight models and better token efficiency (OpenAI claims 54% fewer tokens on coding tasks for its latest model), hurt infrastructure owners while helping application builders.
• The macro question decomposes into millions of micro questions: does each user of each product generate more revenue than the tokens they consume?

Where does the $3 trillion figure come from?

In 2023, Cahn reacted to Nvidia's roughly $50B GPU revenue by calculating the industry would need $200B in revenue to pay back the buildout. Three years of hyperscaling later, his 2026 estimate is $1.5 trillion of infrastructure spend, requiring roughly $3 trillion of AI revenue to justify it. He notes the required revenue per gigawatt has been rising with memory prices and inference-specific chips, so the true number is probably higher.

How big is the revenue gap?

Large. The figures cited alongside the analysis put Anthropic around $60B ARR and OpenAI at $13B earned in 2025 (with a claimed $20B ARR run-rate in November 2025). Even generous growth assumptions and every other AI revenue line leave the industry an order of magnitude away from $3 trillion.

Apollo's Torsten Slok adds the deadline: Google, Meta, Microsoft, and Amazon all project their free cash flow to inflect upward by 2028, which is the market's way of pricing in the payback. If it does not arrive on schedule, he argues, the correction would not stay contained to tech.

Why is falling token pricing both the problem and the answer?

Slok flags the risks to the payoff: organizations shifting to cheaper open-weight models, and overall token prices falling. Add efficiency: OpenAI's latest model reportedly uses 54% fewer tokens on coding tasks. Every one of those is bad news for whoever owns the depreciating GPUs, and good news for whoever pays the token bill.

That is the inversion worth writing down: efficiency gains and price wars transfer margin from the infrastructure layer to the application layer. As a builder, you are on the receiving end of one of the largest cost-deflation programs in tech history, funded by other people's capex.

What is the founder-sized version of the $3 trillion question?

You cannot answer whether the industry earns back its capex. You can answer whether your product does: does each user, each cohort, each feature generate more revenue than the tokens it consumes? That is the same question at a scale where the inputs are knowable: tokens per action, actions per user, price per token, revenue per user.

The industry's answer is just the sum of millions of those small answers. Which is oddly comforting: the most useful thing a founder can do about the $3 trillion question is make their own margin math boringly positive.

One-line takeaway: you cannot control AI capex, but you can control whether your users are margin-positive. Calcaas exists to make that per-user, per-tier math visible before the macro debate settles it for you.