AI Power Demand Collides With Grid Reality as $7.6 Trillion Investment Faces Energy Bottleneck

Goldman Sachs Identifies Power as the Forever Bottleneck

AI power demand is colliding with an infrastructure reality that threatens to derail the industry's ambitious expansion plans. Goldman Sachs has identified power as "the forever bottleneck" in a report projecting $7.6 trillion in capital spending between 2026 and 2031 across computing, AI data centers, and power infrastructure1

1

. The bank expects global data center power demand to surge 170% between 2025 and 2030, with more than 60% of that growth concentrated in the U.S.1

1

. Jason Tofsky, Goldman's global head of digital infrastructure investment banking, emphasized that power infrastructure bottleneck represents the most significant risk to trillions in AI spending, fundamentally reshaping what the AI race actually means.

Grid Connection Delays Create Multi-Year Backlogs

The power grid cannot keep pace with AI expansion demands. The U.S. currently has roughly 2,600 gigawatts of proposed power generation and storage projects awaiting grid connections—more than twice the grid's existing installed capacity1

1

. The median time from a grid connection application to commercial operation stretches to roughly five years1

1

. In Europe, the situation proves even more challenging, with grid connection delays ranging from two to 10 years across the EU, according to the International Energy Agency4

4

. Major European data center hubs including Frankfurt, London, Amsterdam, Paris, and Dublin face connection queues averaging seven to 10 years4

4

. Since January, 86 projects have been canceled or significantly delayed, with 76 directly linked to local opposition1

1

.

Microsoft Withdrawal Signals Physical Infrastructure Constraints

Microsoft recently withdrew from a planned agreement for capacity at Nscale's AI data center in Loughton, England, after grid officials indicated power for the site may not be available until the 2030s3

3

. The abandoned project was designed as the largest AI supercomputer in the United Kingdom, initially anticipating 23,040 Nvidia GB300 graphics processing units in the first quarter of 20273

3

. The planned campus was built for 50 megawatts of initial AI capacity, scalable to 90 megawatts3

3

. This incident highlights how location and electricity supply are becoming critical factors for hyperscalers, even when clients, processors, and funding are already arranged. Microsoft expects roughly $175 billion in calendar 2026 capital expenditures as it races to bring more AI compute infrastructure online3

3

.

FERC Scrambles to Write New Regulatory Framework

The Federal Energy Regulatory Commission is drafting proposals to expand its authority over surging electricity demand, but AI developers' ambitions are colliding with fragmented governance between state regulators, public utilities, and the federal government2

2

. FERC initially missed an April 30 deadline from the Department of Energy to establish national standards for linking AI data centers to the power grid2

2

. The agency released orders in mid-June obligating regional grid operators to justify within two months how their pricing structures can accommodate more data centers or alter them immediately2

2

. Laura Swett, who chairs FERC, noted tension between AI hyperscalers and government regulators, describing "a breakdown of communication in many instances" and observing that "the hyperscalers, when they do come speak to us, they don't speak FERC"2

2

.

Time to Power Becomes the Defining Metric

The concept of "time to power" now determines where AI compute infrastructure gets built, fundamentally altering site selection criteria. For AI infrastructure operators, the wait between requesting grid capacity and receiving it has become more critical than land costs or even hardware availability4

4

. Companies like Clichmont rank available power as the leading site-selection criterion, followed by the time required to deliver that capacity, with cooling conditions and fiber connectivity coming next and land cost following all four4

4

. A site that secures deployable power in 18 months decisively outranks the financial appeal of a cheaper parcel where energization trails construction by several years4

4

. Currently, AI data centers consume 4% of U.S. electricity, with projections showing that figure could triple to 12% by 20282

2

.

Debt-Fueled Buildout Raises Stakes for Investors

The massive AI expansion is being financed with debt, amplifying investor risk if construction timelines slip due to power infrastructure bottleneck issues. Goldman Sachs projects $420 billion in global hyperscaler debt issuance in 2027, up more than 60% from its 2026 estimate1

1

. Debt is expected to finance roughly 35% of hyperscaler capital expenditures next year1

1

. Hyperscalers account for just 40% of AI ecosystem credit fundraising, underscoring how widely the borrowing extends across the industry1

1

. This debt-heavy approach means longer, less predictable buildout timelines and delayed financial returns for companies investing heavily before securing power.

Utilities Transform Into Major AI Beneficiaries

The AI boom is transforming utilities from traditionally slow-growing businesses into major beneficiaries of AI investment. Utility capital plans are expected to reach $1.3 trillion between 2026 and 20301

1

. Tech giants are considering arrangements to directly fund the power generation, transmission lines, and electrical infrastructure their AI data centers require1

1

. The Trump administration has pushed a "speed to power" campaign encouraging tech companies to construct their own power plants to accommodate data centers2

2

. South Dakota Republican Sen. Mike Rounds supports this approach, stating: "If you're going to do data centers, plan on bringing your own power, and let's put some incentives in so that they can actually build the power with the data centers"2

2

. In the PJM Interconnection covering 13 states and Washington, D.C., data center power demand is expected to nearly triple to 29 gigawatts by 2030, while transmission upgrades are already delaying 37.7 gigawatts of planned new generation3

3

. Watch for increased partnerships between hyperscalers and utilities, potential regulatory changes accelerating grid modernization, and growing emphasis on site selection based on existing grid capacity rather than traditional factors like proximity to fiber networks or labor markets.