Goldman Sees $7.6T AI Build-Out; Power Is the Bottleneck

Goldman Sachs published a report on 19 October 2026, “Navigating a System of AI Bottlenecks”, in which its research team expects global data center power demand to rise 170% between 2025 and 2030, with more than 60% of that growth in the United States. The bank estimates roughly $7.6 trillion of capital spending from 2026 to 2031 across computing, data centers and power infrastructure, according to Axios, which received the report exclusively.

“Power is probably the forever bottleneck,” Jason Tofsky, Goldman’s global head of Digital Infrastructure Investment Banking, says in the report.

Business Pill · GRID COST ALLOCATION

A one-minute explainer of grid cost allocation: who pays for a new power line as electricity demand grows. It teaches the general idea only and says nothing about any company in this story.

What Goldman Found

The report says the US interconnection queue has swelled to roughly 2,600 GW of generation and storage projects, more than double the installed capacity of the existing grid, although the figure includes projects at varying stages of maturity and potential overlap. It puts the median time from an interconnection application to commercial operation at roughly five years.

It says US construction of new high-voltage transmission lines fell from an average of 1,700 miles a year between 2010 and 2014 to 350 miles a year from 2020 to 2023. Environmental reviews can take three to four years, and lead times for high-voltage substations and generation step-up transformers have stretched to roughly three to five years, according to Tyler Miller, global co-head of Power, Utilities, and Infrastructure Investment Banking.

Tofsky adds in the report that “to get any one of these projects right, you have to get dozens of things right over a six-month to two-year time frame, including financing.”

What Was Posted on X

Axios posted its exclusive on X at 09:18 UTC on 20 October 2026 under the line “Power is probably the forever bottleneck”, a quote its report attributes to Goldman banker Jason Tofsky, and linked its report on Goldman’s findings about grid delays.

The Local Gate

Goldman counts 86 projects canceled or extensively delayed since January 2026, “with 76 directly linked to local opposition.” The report lists concerns ranging from water and emissions to noise, traffic and potential pressure on residential electricity bills, and says the result is a growing interest in hybrid and behind-the-meter solutions.

The Debt Behind the Build

Axios reports that Goldman projects $420 billion of global hyperscaler debt issuance in 2027, up more than 60% from its 2026 estimate, and that debt is expected to finance roughly 35% of hyperscaler capital expenditure next year. Hyperscalers account for just 40% of AI ecosystem credit fundraising, according to the same account.

On the utility side, the report says utility capital plans are expected to reach an aggregate $1.3 trillion between 2026 and 2030, and that utilities and their large-load customers are considering structures in which hyperscalers contribute directly to the cost of generation, transmission and substation infrastructure.

Connecting the Coverage

Goldman’s count sits beside a different one we reported on 19 October, that at least 47 large US data-centre projects were delayed, withdrawn or blocked in 2026, citing RegLayer data reported by the Financial Times. The two counts use different methods and periods, and Goldman’s readable report does not say which countries its 86 projects are in.

On 18 October we reported that Microsoft dropped an Nscale UK deal to rent 23,000 Nvidia chips at Loughton; the reports did not say why Microsoft withdrew, and the Telegraph reported that grid bosses do not expect a power supply at the site until the 2030s. The structure Goldman describes, with the customer paying for the power, matches the plan we covered on 19 October, under which Amazon would cover the costs of AEP’s 960 MW nuclear plan.

The debt figures sit next to the credit-market data in our 20 October piece on record debt-insurance costs for Meta and Nvidia, and the equipment side shows in ABB’s record $12 billion of third-quarter orders, which ABB said included a triple-digit increase in data-centre orders.

The Structural Read

As we read it, the report’s list of constraints sits mostly outside the data centre itself: transmission lines, interconnection queues, permits, transformers, local approval and workers.

As we read it, local opposition is the factor Goldman names for most of its count: 76 of the 86 canceled or extensively delayed projects are, in the report’s words, directly linked to it.

As we read it, the financing figures Axios reports, $420 billion of 2027 hyperscaler debt and roughly 35% of capex funded by debt, put a timing risk on borrowed money: Goldman’s report says longer timelines delay returns and increase risk.

Three Implications

A FIVE-YEAR WAIT Goldman puts the median time from an interconnection application to commercial operation at roughly five years.

LOCAL OPPOSITION COUNTS Goldman says 76 of 86 projects canceled or extensively delayed since January 2026 are directly linked to local opposition.

CUSTOMERS PAYING FOR POWER The report says utilities and large-load customers are considering structures in which hyperscalers pay directly for generation, transmission and substations.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of AI Data Centers.

The framework says “The structural bottleneck isn’t computing, it’s electrons.”

As we read it, Goldman’s report now puts numbers on that bottleneck from the banker’s side: queue size, connection time and transmission miles. The framework was published in December 2025 and does not discuss Goldman’s report.

What Is Not Established

The $7.6 trillion, $420 billion, 35% and 40% figures come from Axios’s account of the report; we could not find them in the readable text of Goldman’s web version, which may carry them in charts. The report is published by Goldman’s investment banking division, which advises companies in the sectors it describes.

The Bottom Line

Goldman’s report puts power at the front of the AI build-out: a roughly five-year median from an interconnection application to commercial operation, transmission construction down to 350 miles a year, and 86 projects canceled or extensively delayed this year, 76 of them directly linked to local opposition. Its own conclusion is conditional: if power remains the binding constraint, AI growth will move at the speed of power delivery.