Industry

Goldman Sachs Maps the Assumptions That Could Swing AI Build-Out CapEx by Hundreds of Billions

On May 1, 2026, the Goldman Sachs Global Institute published a detailed framework showing that the $4T–$8T AI infrastructure spend estimates over five years hinge on four core supply-side assumptions — silicon useful life, data center complexity, chip architecture mix, and physical bottlenecks — with a baseline of $7.6 trillion cumulative from 2026 to 2031.

Tech Insights Reporter May 1, 2026 7 min read New York

TLDR

Goldman Sachs Global Institute's May 1, 2026 analysis "Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out" provides a sensitivity framework rather than a single forecast. Anchored to NVIDIA forward data center revenue estimates as a proxy for accelerator deployment, the baseline projects ~$7.6 trillion in total AI CapEx (compute, data centers, power) between 2026 and 2031, starting at $765 billion annually in 2026 and rising to $1.6 trillion by 2031. Four assumptions dominate variance: the economic useful life of AI silicon, next-generation data center costs and density, chip and architecture mix, and elongation caused by power, labor, and equipment bottlenecks.

Baseline Scale and Methodology

The report frames the debate around supply-side uncertainty. Estimates of $4–8 trillion in five-year capital investment are common, but the actual number is highly conditional on infrastructure assumptions.

Baseline model inputs include:

NVIDIA data center revenue projections (used as 75% proxy for total compute spend).
VR200 (Rubin-era) chip reference: ~$80.5K per GPU package (incl. node costs), 3,000 W.
1.2 PUE, $15 million per MW for data centers, $2,500 per kW for new power capacity.
15% brownfield data center space in 2026, rising to 30% by 2031.
Straight-line depreciation with no terminal value for accelerators.

This produces the headline baseline of $7.6 trillion cumulative and the annual ramp cited above. The analysis explicitly states it is a scenario framework to test how changes in assumptions move the totals, not a demand forecast.

The Four High-Impact Assumptions

1. Economic useful life of AI silicon
Silicon turns over far faster than buildings or power infrastructure (typically 4–6 years). Small shifts in assumed replacement cadence produce hundreds-of-billions swings in cumulative spend because accelerators dominate the cost stack. Extending life from four to six years materially reduces replacement cycles. Tiered deployment — using trailing-edge chips for inference, edge, and synthetic data — could support longer effective lives than front-line training clusters.

2. Cost and complexity of next-generation data centers
AI workloads drive higher rack densities, liquid cooling, tighter power tolerances, and codesigned systems rather than layered components. These changes increase both upfront construction cost and integration risk compared with prior cloud generations.

3. Chip and architecture mix
Whether demand is elastic (lower prices expand total compute purchased) or inelastic (fixed workload needs) determines whether architecture shifts primarily affect margins or headline capital totals.

4. Elongation from bottlenecks
Power, labor, equipment lead times, and permitting can stretch timelines. In stress scenarios this feeds back into demand uncertainty.

The report notes that many popular discussion topics (returns, monetization timing, value distribution) matter for investors but do not change the fundamental quantity of capital that must be deployed to deliver a given level of compute.

Why this story matters

The analysis shifts the conversation from binary "too much or too little CapEx" to the specific engineering and accounting variables that actually determine required investment. With silicon life identified as the single most influential lever and data center specifications evolving rapidly, operators and capital allocators now have a clearer map of where small assumption changes create large outcome differences. The $765 billion 2026 starting point and $7.6 trillion five-year baseline provide a concrete reference against which future NVIDIA guidance, hyperscaler spend, and power build-out data can be tested.

Sources

Goldman Sachs Global Institute: “Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out” (published May 1, 2026; George Lee and Lucas Greenbaum). https://www.goldmansachs.com/insights/articles/tracking-trillions-the-assumptions-shaping-scale-of-the-ai-build-out (includes downloadable PDF with baseline tables and sensitivity analysis).
Cross-referenced NVIDIA GTC presentations (March 2025 & 2026) and Wall Street projections cited in the methodology.

Featured Image Alt Text

Chart showing cumulative AI infrastructure capital expenditure baseline of $7.6 trillion from 2026–2031, with sensitivity bands driven by silicon useful life and data center density assumptions