Charted: Goldman Sachs Puts 2027 AI Capex at $961B — Chips and Data Centers

The direct answer from Goldman

Goldman Sachs Global Institute published the most detailed public breakdown of global AI infrastructure spending in Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out (April 2026, authors George Lee and Lucas Greenbaum).

The interactive chart above maps their baseline scenario model. For the years most investors ask about:

"Compute" is accelerators and systems (GPUs and ASICs including node costs) — the closest line to chips. "Data centers" is shell, cooling, and fit-out at a baseline $15M per MW. Power generation is reported separately and remains small relative to the other two layers.

Not a forecast — read the disclaimer first

Goldman is explicit that this is not a house forecast. The report describes itself as a scenario-based framework to explore how infrastructure assumptions affect aggregate capital requirements. The disclaimer states it was prepared by the Goldman Sachs Global Institute and is not a product of Goldman Sachs Global Investment Research.

That distinction matters because a different Goldman number circulates constantly: ~$1.14 trillion in hyperscaler capex for 2027 (Investment Research base case, strategist Ryan Hammond, June 2026). That figure is narrower — hyperscaler spending only, not all-in global AI infrastructure. Street consensus on hyperscaler capex sits near $920B. Do not mix the scopes.

What drives the data center line

The data center capex figure is the most assumption-sensitive piece. Goldman's baseline uses $15M/MW. Legacy hyperscale cloud was built around $10M/MW; next-generation AI facilities are running $15–20M/MW with upside as density and redundancy rise.

At $11M/MW, 2027 data center capex drops to $220B and 2028 to $259B. At $19M/MW, those rise to $380B and $447B. The compute line is unchanged — only the facility cost assumption moves.

Cumulative build-out: $7.6 trillion

Across 2026–2031, the baseline model totals:

• $5.1T compute
• $2.1T data centers
• $358B power
• $7.6T all-in

Compute alone crosses $1T annually by 2030 ($1,073B) in the baseline path.

Chip stack: top-down model vs bottom-up checks

Tracking Trillions does not roll compute into a single semiconductor industry revenue forecast. Global Institute builds the $661B / $808B compute lines top-down from accelerator shipments; equity research sizes suppliers separately. Those bottom-up figures — Broadcom AI silicon, HBM, MediaTek ASICs, TSMC capex — are in the chip & component table in the chart above, not in the main scenario model.

Worth noting outside that table: Goldman raised its 2027 HBM market estimate from $75B to $116B, a signal that memory is tightening faster than earlier assumptions.

What actually moves the total

Goldman ranks four assumptions as decisive:

1. Economic useful life of AI silicondoes not change annual compute capex but swings implied depreciation by hundreds of billions
2. Data center cost per MWthe biggest lever on the facility line
3. Chip and architecture mix
4. Elongation from power, labor, and equipment bottlenecks

It argues three widely debated factors — training vs inference mix, per-chip memory growth, and behind-the-meter vs grid power — affect returns and value distribution but do not materially change aggregate capital required.

Methodology

All figures from Goldman Sachs Global Institute, Tracking Trillions (April 2026). Baseline assumptions: NVIDIA forward data center revenue estimates (March 3, 2026), 75% NVIDIA share of compute, VR200 at $80.5K/GPU and 3,000W, $15M/MW data center cost, $2,500/kW new power, PUE 1.2, 15–30% brownfield space exclusion rising through 2031. Hyperscaler cross-check figures from GS Investment Research via public reporting, June 2026. Chip/component rows in the dashboard from GS equity research (Broadcom, MediaTek, TSMC) and Lee/Schneider memory research, 2026.

Primary source: Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out