Goldman: AI CapEx Is a Sensitivity Model — Silicon Life Swings Hundreds of Billions
Goldman Sachs Global Institute’s ‘Tracking Trillions’ frames $4–8T five-year AI infrastructure estimates as assumption-driven: baseline ~$7.6T (2026–2031), from $765B in 2026 to $1.6T in 2031. Four levers — silicon useful life, data-center complexity, chip/architecture mix, physical bottlenecks — dominate variance. Scenario kit, not a demand forecast.
The useful question is not “too much CapEx?” — it is which assumptions move the total by hundreds of billions. Goldman’s piece is a sensitivity kit anchored to NVIDIA data-center revenue as a proxy, not a single prophetic number.
Goldman Sachs Global Institute (May 1) published “Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out.” Baseline: ~$7.6 trillion cumulative AI CapEx (compute, data centers, power) from 2026 to 2031 — $765 billion in 2026 rising to $1.6 trillion by 2031. Popular $4–8T five-year bands hinge on four supply-side assumptions.
Baseline method (GSGI)
- NVIDIA data-center revenue as ~75% proxy for total compute spend.
- VR200 (Rubin-era) reference: ~$80.5K per GPU package, 3,000 W.
- 1.2 PUE; $15M per MW data centers; $2,500/kW new power.
- 15% brownfield space in 2026 → 30% by 2031.
- Straight-line depreciation; no terminal value for accelerators.
Explicitly a scenario framework, not a demand forecast.
Four high-impact assumptions
- Economic useful life of AI silicon — typically 4–6 years; small cadence shifts swing hundreds of billions; tiered trailing-edge use for inference can extend effective life.
- Next-gen data-center cost/complexity — density, liquid cooling, codesign raise upfront cost and integration risk.
- Chip/architecture mix — elastic vs inelastic demand determines whether architecture shifts hit margins or headline CapEx.
- Bottleneck elongation — power, labor, equipment lead times, permitting stretch timelines and feed demand uncertainty.
Monetization/returns matter for investors but do not change the capital required to deliver a given compute level (GSGI framing).
Claims vs checks
Figures and assumption list are Goldman Sachs Global Institute primary (Lee, Greenbaum). NVIDIA-forward revenue proxy inherits NVIDIA guidance risk. Cross-check against future hyperscaler CapEx and power build-out data — the point of the framework.
Limits
- Proxy-based — not a bottoms-up census of every hyperscaler.
- Baseline sensitive to silicon-life and $/MW assumptions.
- Not a recommendation to buy/sell any security.
Sources
- Goldman Sachs Global Institute: “Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out” (May 1, 2026).
- NVIDIA GTC presentations and Wall Street projections cited in the methodology.