Goldman Sachs Maps $7.6 Trillion AI Infrastructure Bet Through 2031 — Nvidia Takes 75%

The numbers are staggering, and Goldman Sachs wants you to sit with them. In its new infrastructure spending analysis, the bank’s equity research team maps out $7.6 trillion in cumulative AI capital expenditure between 2026 and 2031 — a sum equivalent to roughly one-quarter of annual U.S. GDP, deployed into compute silicon, data centers, and power infrastructure over six years.

The report arrives at a moment when even the largest hyperscalers are having difficulty explaining where their capital is going. Amazon, Microsoft, Google, and Meta collectively committed roughly $320 billion in AI infrastructure spending in 2025 alone. Goldman’s projection suggests that pace not only continues but accelerates: annual AI capex rises from $765 billion in 2026 to $1.6 trillion by 2031.

The Three-Layer Breakdown

Goldman structures the build-out across three components.

Compute: $5.1 trillion. This is the dominant category — GPUs, ASICs, and the associated networking fabric that connects them. Nvidia, the report projects, captures 75% of this layer. At that share, Nvidia would accumulate approximately $3.8 trillion in compute revenue over the six-year horizon, a projection that makes its current ~$2.3 trillion market cap look, depending on your discount rate, either wildly overvalued or astonishingly undervalued.

The baseline unit of analysis in the report is the Rubin VR200 GPU, priced at $80,500 per chip. Despite the aggressive pace at which Amazon (Trainium), Google (TPU), Meta (MTIA), and Microsoft (Maia) are developing custom silicon, Goldman’s analysts conclude that performance gaps in training workloads force hyperscalers to continue purchasing Nvidia hardware in parallel with their own designs — not as a replacement strategy, but as a hedge.

Data Centers: $2.1 trillion. The physical infrastructure housing AI compute is undergoing what Goldman calls a fundamental architectural reinvention. Standard cloud infrastructure requires 5–15 kilowatts per rack. Current Blackwell-era AI facilities run at 130–200 kilowatts per rack. Next-generation AI factories — built to house 2027–2028 generation chips — will exceed 500 kilowatts per rack, requiring liquid cooling infrastructure that simply doesn’t exist at scale today.

Construction costs reflect this shift. Traditional hyperscale data centers cost approximately $10 million per megawatt to build. AI-optimized facilities are now running $15–20 million per megawatt — a 50–100% cost premium before the power contracts are signed. Vertiv, which manufactures liquid cooling systems and power distribution equipment, emerges as a critical downstream beneficiary: the liquid cooling market is projected to grow from $5.5 billion to $15.75 billion by 2030.

Power: $358 billion. The smallest category by dollar value is, paradoxically, the most operationally constrained. “Our single biggest constraint is power,” Amazon CEO Andy Jassy stated publicly earlier this year — a sentiment that Satya Nadella and Sundar Pichai have echoed in their own earnings calls and investor days.

Grid connection timelines for large-scale data centers now extend three to five years in most Western markets, with permitting, interconnect queues, and utility upgrade cycles all contributing to the delay. This is driving hyperscalers toward unconventional power sources: Vistra secured a 20-year agreement with Meta covering 2,600+ megawatts of nuclear capacity, with a separate deal with AWS; Microsoft committed to Three Mile Island’s restart; Google struck deals with geothermal and advanced nuclear providers. Goldman upgraded Vistra and several other power names following these announcements.

The Critical Variable: Chip Depreciation

The most technically important section of Goldman’s report is its sensitivity analysis on silicon replacement cycles. AI chips are capitalized as fixed assets with defined useful lives, and that accounting assumption drives enormous variance in the total build-out cost.

Under a three-year replacement cycle — aggressive, reflecting the pace at which new GPU architectures render predecessors obsolete — cumulative compute depreciation reaches $3.99 trillion. Extend that to seven years — conservative, reflecting traditional enterprise IT cycles — and depreciation falls to $2.23 trillion. The difference: $1.76 trillion on a single input assumption.

This matters for investors reading the report as a demand forecast. If Nvidia releases Rubin in 2025, Feynman in 2027, and another generation in 2029, hyperscalers face pressure to refresh every 24–36 months to maintain training efficiency. The shorter the cycle, the higher the recurring capital expenditure — and the higher Goldman’s $7.6 trillion base case becomes.

The Demand Thesis

Goldman’s bull case rests on three assertions. First, AI inference demand is growing faster than deployment capacity. The gap between what enterprises want to run and what the existing infrastructure can serve creates a sustained pull on new compute. Second, model capability gains are creating new use cases faster than existing use cases are monetized — expanding the addressable market before prior investments are recovered. Third, the geopolitical dimension (U.S.-China chip restrictions, European sovereignty concerns, India’s national AI missions, South Korea’s compute commitments) creates parallel demand loops rather than a single consolidated market.

The bear risks are real, and Goldman acknowledges them. Hyperscaler returns on AI infrastructure investment remain unproven at scale. The ad revenue gains at Meta and the Azure AI premium at Microsoft are real but small relative to capital commitments. If large language models hit a capability ceiling — if the scaling law bends — the demand rationale collapses before the infrastructure is paid for.

What It Means for the Market Map

The Goldman report is partly a research document and partly a portfolio map. Nvidia’s position in compute is the central node, but the surrounding ecosystem benefits substantially from the build-out regardless of which LLM architectures dominate. Custom memory from SK Hynix and Samsung (HBM4E), networking from Arista and Cisco, liquid cooling from Vertiv, power solutions from Vistra and Constellation, and fiber backbone providers all sit downstream of the capex.

The $7.6 trillion number is not a guarantee. It is a scenario that becomes true if AI capabilities continue advancing, if enterprise adoption of agentic systems reaches the penetration rates that hyperscalers are projecting, and if geopolitical disruptions don’t fragment the global chip supply chain further than they already have.

Given that Goldman’s prior 2024 AI capex forecast was substantially revised upward in 2025, which was again revised upward to produce this 2026 report, the most likely thing the next version will show is a number larger than $7.6 trillion.

The build-out has its own momentum now. The question is who captures the margin.