The gap everyone quotes
The AI infrastructure debate is stuck on the wrong question. It isn’t “is AI a bubble?” — it’s whether the people paying for the buildout and the people who capture the revenue are still the same peop
Sequoia’s David Cahn puts 2026 hyperscaler data-center capex at ~$750B. Apply his 2x rule (to cover non-chip costs and ~50% margins) and a single year’s build needs ~$1.5T in end-customer revenue to earn a fair return; cumulatively since ChatGPT, ~$3T. This is revenue required, not cash spent — the two get conflated constantly.
On the other side, OpenAI and Anthropic together run at roughly $100B ARR (SemiAnalysis estimate; frontier labs are still burning cash overall). $3T of cumulative required revenue against ~$100B of current run-rate — different units, but the distance is the point.
The real risk is distribution, not total
Even if AI ends up creating far more than $3T in value, the money splits four ways:
• Spending it: a handful of hyperscalers, data centers, frontier labs
• Compressing it: open-weight models, inference optimization, cheaper silicon
• Enjoying the productivity: enterprises and consumers using AI
• Actually earning it: apps that own the user, the workflow, the transaction, and proprietary data
Here’s the view from our own data. Across the 5,718 AI companies in the SVTR AI VC Database, the model and infrastructure layers are only ~22% of firms but hold ~67% of disclosed funding — model-layer firms average ~$1B raised each, about 13x the application layer. Capital sits upstream; revenue has to come from the ~80% of firms downstream.
Efficiency and open weights change who gets paid
The “token factory” assumes usage grows to fill capacity. Two forces in 2026 move the revenue, not the volume.
Efficiency: OpenAI’s GPT-5.6 Sol is 54% more token-efficient on agentic coding than rival models (Altman, July 9). Good for users; for token-metered vendors, the same job may bring in less — unless usage scales up to compensate.
Open weights: on OpenRouter, Chinese models overtook US models in the week of Feb 9–15 and by June ran ~18T vs ~5.5T weekly tokens; the US share fell from ~70% to ~30%. They’re 10–20x cheaper.
Jevons’ paradox all but guarantees total usage keeps exploding. The open question is who gets paid: when the price of intelligence falls faster than demand rises, incremental usage leaks toward models that don’t pay for that $750B of capex.
China isn’t answering the question — it’s making it harder
China runs the opposite playbook: capex-light, open-weight-heavy, priced to win developers. But a nuance that gets lost: the weights are Chinese, the calls often aren’t Chinese-hosted — they still run on AWS, Azure, Nvidia. What open weights compress first is model-layer margin and high-margin API revenue, not necessarily all of infrastructure revenue. Value is moving from model ownership to compute hosting, distribution, and workflow.
Why second-order investors should care
Torsten Slok’s worry isn’t AI itself — it’s transmission. Consensus has the four hyperscalers’ combined free cash flow growing 4x+ from 2026 to 2030. But with the Magnificent 7 so heavy in the index, a slower payoff wouldn’t stay contained — it spreads to chips, power, data centers, and the S&P 500.
The bottom line: AI will likely create far more than $3T in value, but the people who put up the $3T in capital risk won’t necessarily capture a proportional share of the revenue. For the next 12 months, the bet isn’t whose model is best — it’s who can own a high-frequency workflow, accumulate data a model can’t copy, and hold the customer relationship. When tokens are cheap, the call isn’t the moat; embedding it in a business outcome is.
Read the full piece with the AI VC Database charts →