Goldman Sachs Models a $7.6 Trillion AI Infrastructure Build-Out Through 2031
Goldman Sachs published a May 1 scenario model estimating roughly $7.6 trillion of cumulative AI infrastructure capital expenditure from 2026 through 2031, rising from $765 billion annually to $1.6 trillion. The firm stressed that this is a supply-side baseline, not a forecast of AI adoption or demand, and that chip life, data-center costs, architecture choices, and physical bottlenecks could move the total materially.
Goldman Sachs published a May 1 framework estimating roughly $7.6 trillion of cumulative capital expenditure on AI infrastructure between 2026 and 2031. The baseline rises from about $765 billion in annual spending in 2026 to $1.6 trillion in 2031 across compute, data centers, and power.
The distinction Goldman makes is important: the model is a supply-side reference point based on current chip-sales expectations and associated infrastructure requirements. It is not a forecast of AI adoption or end-market demand. CBS News later used the estimate in a June 26 analysis of investor concern over whether AI revenue will justify the build-out.
Four assumptions drive the range
Goldman's framework identifies four variables with the largest effect on aggregate capital requirements:
- •the economic useful life of AI accelerators;
- •the cost and complexity of next-generation data centers;
- •the mix of chips and system architectures; and
- •delays caused by power, labor, and equipment bottlenecks.
Chip replacement is especially consequential because accelerators are expensive and turn over faster than buildings or power infrastructure. Extending their useful life reduces replacement cycles; faster obsolescence increases cumulative spending and depreciation. Higher rack density also raises cooling, cabling, and power-delivery costs beyond the price of the processors themselves.
Spending and payback are different questions
CBS reported that investors were asking whether consumer and enterprise demand would produce enough revenue and profit to support the capital intensity. That concern does not invalidate Goldman's infrastructure model, but it addresses a different part of the economics: the report estimates what a build-out implied by current supply expectations could cost, while monetization determines who earns an acceptable return.
For operators, the useful takeaway is not to treat $7.6 trillion as a guaranteed spending total. Capacity plans should be tested against hardware-refresh cadence, utilization, power availability, and data-center delivery schedules. For investors and enterprise buyers, revenue growth and measurable workload economics remain the checks on whether the physical build-out creates durable value rather than excess capacity.
Key Points
- 1Goldman Sachs' baseline model estimates roughly $7.6 trillion of AI infrastructure capital expenditure from 2026 through 2031.
- 2The model rises from $765 billion in annual spending in 2026 to $1.6 trillion in 2031, but it is a supply-side scenario rather than an adoption or demand forecast.
- 3Chip life, data-center costs, architecture choices, and power, labor, or equipment bottlenecks are the variables most likely to move the total.
Scoring Rationale
The scenario quantifies the physical scale of the AI build-out and makes its assumptions explicit. Its value is highest when read as a sensitivity framework for compute, facilities, and power planning rather than as a guaranteed spending or demand forecast.
Sources
Primary source and supporting public references used for this report.
Practice with real Banking data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Banking problems
