Goldman Sachs Maps $7.6 Trillion AI Infrastructure Spending Through 2031

TLDR:

• Goldman Sachs projects $7.6 trillion in cumulative AI infrastructure capex between 2026 and 2031.
• Nvidia is forecast to capture 75% of the $5.1 trillion compute layer over the six-year period.
• Power is the smallest budget segment at $358 billion but the critical bottleneck for full deployment.
• Vistra locked in a 20-year, 2,600 MW nuclear deal with Meta to meet surging AI power demand. 

Goldman Sachs has released a detailed projection of artificial intelligence infrastructure capital expenditure, forecasting $7.6 trillion in cumulative spending from 2026 to 2031.

The breakdown covers three core layers: compute at $5.1 trillion, data centers at $2.1 trillion, and power at $358 billion.

The report identifies specific companies positioned to absorb capital across each segment. At $765 billion in 2026 alone, annual AI capex is expected to reach $1.6 trillion by 2031.

Compute and Data Center Demands Drive Infrastructure Buildout

Goldman Sachs AI infrastructure projections place Nvidia at the center of the compute layer. The firm assumes Nvidia will capture 75% of all compute spend over the forecast period, translating to roughly $3.8 trillion in cumulative revenue through its products. The baseline unit for this projection is the Rubin VR200 chip, priced at $80,500 per GPU.

Nvidia’s data center GPU gross margins sit at 75%, which is why major hyperscalers are developing custom silicon.

However, performance gaps mean those same companies continue purchasing Nvidia hardware in parallel. No current alternative matches its output at scale.

The data center layer reflects a sharp escalation in physical requirements. Standard cloud infrastructure runs between 5 and 15 kilowatts per rack.

Transitional Blackwell-era AI facilities operate at 130 to 200 kilowatts per rack. Next-generation AI factories running Rubin and Feynman chips require over 500 kilowatts per rack, with liquid cooling as the only viable thermal option.

Construction costs are rising alongside density. Traditional hyperscale data centers cost approximately $10 million per megawatt to build.

Next-generation AI data centers are being planned at $15 to $20 million per megawatt, a sharp increase driven by cooling and power infrastructure requirements.

Power Constraints and Key Companies Shape the Capital Cycle

Goldman Sachs AI infrastructure analysis identifies silicon useful life as the single largest variable in the model. At a three-year replacement cycle, cumulative compute depreciation reaches $3.99 trillion. A seven-year cycle drops that figure to $2.23 trillion, a difference of $1.76 trillion on one assumption.

Vertiv is positioned directly within the data center upgrade cycle. Every rack transitioning from 40 kilowatts to 500-plus kilowatts requires new liquid cooling systems and power distribution equipment. The liquid cooling market is projected to grow from $5.5 billion today to $15.75 billion by 2030.

Power, at $358 billion, is the smallest budget segment but the most operationally critical. Amazon CEO Andy Jassy captured the constraint plainly: “Our single biggest constraint is power.”

Grid connection timelines for large data centers extend into years, making early contracting essential for deployment.

Vistra has responded to that constraint by locking in long-term nuclear power purchase agreements. The company secured a 20-year deal with Meta covering over 2,600 megawatts of nuclear energy, along with a separate agreement with AWS.

Goldman Sachs and Jefferies both upgraded Vistra following the Meta announcement, according to Milk Road AI’s breakdown of the report.