Goldman Sachs: Energy Bottlenecks Will Slow the AI Race, with $7.6 Trillion in Related Capex from 2026 to 2031
nashnova research
Goldman Sachs warns that global AI compute expansion is hitting a hard ceiling in power supply, projecting $7.6 trillion in related capital spending from 2026 to 2031 — but grid-permitting delays and ballooning debt could slow the spend and push returns further out.
Why could power become AI's "permanent bottleneck"?
Goldman projects global data-center power demand will grow 170% between 2025 and 2030, with over 60% of the increase coming from the U.S.
Jason Tofsky, the bank's head of global digital-infrastructure investment banking, put it bluntly: "Power could be a permanent bottleneck."
This means → no matter how fast AI technology advances, compute cannot physically deploy if the electricity isn't there. The chokepoint is the power socket, not the chip.
How congested is the grid queue?
Roughly 2,600 GW of generation and storage projects are waiting for grid interconnection in the U.S. — more than twice the existing grid's installed capacity.
In plain terms = the line of projects trying to plug into the grid is longer than the grid itself, and most may never reach the front.
The median time from interconnection application to commercial operation is about five years. Since January, 86 projects have been cancelled or significantly delayed; 76 of those were blocked by local community opposition.
This reflects a reality: the true bottleneck for AI infrastructure is not technical permitting — it is community acceptance.
$7.6 trillion — where does the money go, and who borrows it?
Goldman estimates total global capex on compute, data centers, and power infrastructure at $7.6 trillion from 2026 to 2031.
The build-out leans heavily on debt. Hyperscaler — the handful of giant cloud platforms such as Amazon AWS, Microsoft Azure, and Google Cloud — debt issuance is projected to hit $420 billion in 2027, up over 60% from the bank's 2026 estimate.
This means → if construction slips, interest costs keep accruing while revenue is delayed — financial pressure compounds.
Are the tech giants the only ones borrowing?
Hyperscalers account for only 40% of total credit financing across the AI ecosystem.
In plain terms = the market focuses on a few big names, but the entire supply chain — chip vendors, power contractors, everybody in between — is borrowing heavily. The risk exposure is far wider than it appears.
Why are utilities the surprise winner?
Goldman notes that the AI boom is turning traditionally slow-growth utilities into major beneficiaries, with utility capex plans projected at $1.3 trillion from 2026 to 2030.
Tech giants are also considering funding power-generation, transmission, and grid infrastructure directly for their data centers.
This reflects a shift: when electricity is the scarcest resource, whoever controls the power holds the ticket to the AI race.
Is a slower build necessarily bad news?
Goldman also flags a silver lining: a slower deployment pace could help avoid over-investment, giving actual demand time to catch up with spending.
But for companies that committed capital before locking in power supply, the energy-bottleneck risk is already real and present.
This means → moving a step slower is not necessarily losing — but charging ahead without power secured may be the fastest way to fall first.
市场有风险，内容仅供研究参考，不构成投资建议。