Goldman Sachs has published baseline aggregate AI CapEx estimates through 2031. The headline number — $7.6 trillion across compute, data centers and power from 2026 to 2031 — is striking. But the breakdown is where the real intelligence lies for anyone working in AI infrastructure deployment.

THE NUMBERS

Goldman Sachs Baseline AI CapEx Estimates — 2026–2031 2026: $765B total ($494B compute · $232B data centers · $39B power)
2027: $1,011B total ($661B compute · $300B data centers · $50B power)
2028: $1,220B total ($808B compute · $353B data centers · $59B power)
2029: $1,392B total ($934B compute · $393B data centers · $65B power)
2030: $1,579B total ($1,073B compute · $433B data centers · $72B power)
2031: $1,636B total ($1,127B compute · $436B data centers · $73B power)
─────────────────────────────────────────────────────────
TOTAL: ~$7.6 trillion · Compute $5.1T · Data Centers $2.1T · Power $358B

THE CRITICAL PATH INSIGHT

The power segment — $358 billion over six years — represents 4.7% of the total $7.6 trillion forecast. It is the smallest of the three categories by a significant margin.

It is also the only one that, if it fails to deploy, prevents the other 95.3% from deploying.

This is not a theoretical observation. It is the operational reality that the CEOs of NVIDIA, Microsoft, Amazon, Meta and OpenAI have all described publicly in 2026:

The $5.1 trillion in compute investment requires power infrastructure to function. The $2.1 trillion in data center investment requires power infrastructure to connect. Without the $358 billion power layer — grid connections, transformers, substations — none of the stack above it gets deployed.

WHY THE POWER SEGMENT IS UNDERSIZED

The $358 billion power forecast may itself be an underestimate of what is actually required. The constraint is not capital — it is physical manufacturing capacity and grid connection timelines:

Power Infrastructure Constraints — Current Reality HV transformer lead times (major US OEMs): 48–60 months
Grid connection queue, Northern Virginia: 7–10 years
US AI data center capacity blocked 2026: 7 GW (Sightline Climate)
EU second-tier transformer lead times: 20–32 months
France RTE connection (brownfield site): 18–36 months
French nuclear baseload: €50–70/MWh

The bottleneck is not a lack of willingness to spend the $358 billion. It is the physical impossibility of deploying that capital fast enough given current transformer manufacturing capacity and grid connection queue lengths.

THE GEOGRAPHIC IMPLICATION

The Goldman Sachs forecast is global but the constraint is US-specific. The US faces the most acute combination of grid saturation, transformer lead times and GOES (Grain-Oriented Electrical Steel) supply concentration that limits how fast domestic power infrastructure can be built.

Europe faces the same demand signal — Nebius's 4 GW contracted capacity, the Béthune 240MW project, hyperscaler EMEA expansion — but with structurally different supply conditions:

The $358 billion power investment that Goldman Sachs forecasts will flow disproportionately toward geographies where it can actually be deployed. Europe — specifically France — is structurally positioned to absorb a meaningful share of that capital faster than US markets can.

WHAT THIS MEANS FOR 2026–2028 DEPLOYMENT

The Goldman Sachs forecast covers 2026–2031. For projects targeting commissioning in 2027 or 2028 — the near-term deployment window — the decisions that determine success must be made now:

The $7.6 trillion forecast is the demand signal. The transformer lead times and grid connection queues are the supply constraint. GridReadiness exists at the intersection — tracking the power infrastructure reality that determines how much of Goldman's $7.6 trillion actually gets deployed on schedule.

"~$7.6tr of capital between 2026 and 2031 across compute, data centers, and power." — Goldman Sachs Baseline AI CapEx Estimates

$358 billion. 4.7% of the total. The critical path for 100% of the stack.