Sequoia Raises the AI Payback to $3 Trillion — Three Years In, the Revenue Gap Is Still Widening
In 2023, Sequoia partner David Cahn published what became one of the more-cited memos of the AI boom: a back-of-envelope calculation showing that $200 billion in annual revenue would be needed to justify the AI infrastructure build-out then underway. The memo was not a bear thesis. It was a challenge to founders to build the products that would close the gap.
Three years later, Cahn has published an updated figure: $3 trillion.
The Arithmetic
The method is unchanged. Start with AI infrastructure capital expenditure — data centres, chips, cooling, networking. Add operating costs: power, staffing, maintenance. Apply the margin structure needed by the operators and investors who funded the build. The result is the revenue that must flow through the stack to make the capital productive.
In 2023, the annual AI infrastructure spend was roughly $100 billion. The $200B payback requirement followed from that base with standard opex and margin assumptions.
In 2026, Cahn puts total AI infrastructure spending at $1.5 trillion for the year. Under the same methodology, the payback requirement doubles the spend: $3 trillion.
Why the Number Is an Underestimate
Two cost pressures are not fully captured in the top-line figure.
The first is memory. HBM prices have risen sharply throughout 2026, with SK Hynix, Samsung, and Micron all reporting record margins as AI demand outpaces manufacturing capacity. HBM now accounts for 63% of AI chip component costs, up from 52% eighteen months ago. The model that was priced when 2026 infrastructure commitments were made has already been revised upward.
The second is the shift toward inference-specific silicon: custom accelerators, ASICs, and photonic interconnects. These carry different cost structures than standard GPU clusters — higher per-unit costs offset by efficiency gains that take time to materialise in production workloads. Capex committed to these platforms today will generate returns on a longer payback horizon than the headline numbers imply.
Cahn notes that adding these factors raises the number beyond $3 trillion without altering the underlying logic.
What the Gap Looks Like
The AI industry’s actual revenue position is not public in aggregate, but partial data points suggest the gap remains large. Anthropic is tracking toward roughly $30 billion annualised run rate. OpenAI is running at a similar scale. Google Cloud — which is not purely AI revenue but is the most AI-dependent segment among hyperscalers — reported $20 billion in Q1 2026.
Scale these across the full market, including enterprise deployments, API resellers, and embedded AI in enterprise software, and you get an industry-wide figure that is likely in the hundreds of billions annually. That is meaningful but well short of the $3 trillion payback requirement.
The Counter-Argument
Cahn is not making a crash prediction. He closes with the same point he made in 2023: cloud infrastructure had a comparable financing gap in 2000 and 2001. That gap closed — slowly, then quickly — as application revenue caught up. The bet is that AI follows the same curve.
The tension is timing. Cloud compute’s payback played out over a decade during which infrastructure costs fell and usage grew. AI infrastructure costs are not falling at the same rate — memory shortages and power constraints are pushing them higher. If application revenue does not accelerate materially in the next two to three years, the debt-funded portion of the build begins compounding against the thesis rather than with it.
The $3 trillion number is not a verdict. It is a clock.