The grand narrative of artificial intelligence infrastructure expansion in the United States is confronting a severe reality check. According to the latest "2026 Data Center Outlook" report from Sightline Climate, an estimated 30% to 50% of the roughly 16 gigawatts of new data center capacity planned for this year in the US is expected to be delayed or canceled. Currently, only about 5 gigawatts have actually reached the construction phase.
This figure stands in stark contrast to the annual capital expenditure budgets exceeding $700 billion from hyperscale cloud computing companies, revealing deep-seated challenges in AI infrastructure build-out related to power supply, supply chain issues, and socio-political factors. Canaccord Genuity analyst George Gianarikas summarized the situation as "the US data center boom hitting a powerful wall of logistical resistance." As reported by Bloomberg, severe shortages of electrical equipment like transformers, switchgear, and batteries are a primary cause of delays. Domestic US manufacturing capacity is far from sufficient to meet demand, forcing developers to rely on imports. Concurrently, community opposition, permitting hurdles, and lagging grid integration are collectively shrinking the feasibility window for project completion.
The supply-demand gap is alarming, with prospects beyond 2027 looking even bleaker. The Sightline Climate report indicates that 140 data center projects are planned to come online across the US in 2026, representing a combined capacity of approximately 16 gigawatts. Of this, 53% is slated to connect to the grid, 3% will rely on self-contained power sources, and 25% have not yet disclosed their power supply plans. However, within this 16 GW of planned capacity, 11 GW remain in the "announced stage" with no signs of construction. Given that the typical construction cycle for a data center is 12 to 18 months, the likelihood of this portion of capacity being delivered on time is very low.
Looking ahead to 2027, the gap widens further. Announced planned capacity reaches 21.5 GW, but the actual capacity under construction is only about 6.3 GW. The situation from 2028 to 2032 is even more severe: the vast majority of planned projects have not broken ground, and an additional 37 GW of planned infrastructure lacks a clear completion date, with only 4.5 GW of that actually under construction.
The transformer bottleneck is constricting a critical artery, with the supply chain heavily dependent on imports. Shortages in power infrastructure are a core bottleneck restricting data center construction. The rapid scaling of data center size requires higher-capacity transformers to safely convert and deliver power from high-voltage grids to chips. Philippe Piron, CEO of GE Vernova's Electrification business, noted that before 2020, lead times for large-power transformers were typically 24 to 30 months, which was "perfectly acceptable" under the old paradigm. However, AI companies often demand delivery within 18 months. Surging demand has pushed transformer lead times out to as long as five years, with prices also rising sharply. Some companies are resorting to stopgap measures; for instance, Crusoe has begun refurbishing old transformers from decommissioned power plants for emergency use.
Pressure on demand for electrical equipment like transformers comes not only from data centers but also from the proliferation of electric vehicles and heat pumps, which is driving up the need for grid expansion. US domestic manufacturing capacity is fundamentally unable to keep pace, deepening reliance on imports. This predicament reflects a deep-seated structural issue stemming from decades of US manufacturing outsourcing. Despite recent policy calls for reshoring, substantive capacity increases have shown limited effectiveness so far.
Promises of nuclear power remain unfulfilled, with a funding gap reaching trillions. Challenges on the power supply side are equally concerning. Promises of a nuclear renaissance from the Trump administration remain largely verbal, with almost no new nuclear plants breaking ground. While small modular reactors are seen as a beacon of hope, they are still years away from practical, large-scale deployment.
A few very large self-powering projects are attempting to bypass grid limitations. These include New Era Energy & Digital's planned 7 GW project in Lea County, New Mexico; Homer City's 4.5 GW coal-to-gas conversion project in Pennsylvania; and Crusoe's 1.8 GW hybrid natural gas and renewable energy project in Cheyenne, Wyoming. The report notes that waiting for the grid to provide power capacity of this magnitude "could take a decade."
On the funding front, analysis from JPMorgan indicates that fully supporting the current AI cycle requires no less than $5 trillion. Even if hyperscale cloud companies continue to increase capital expenditures, the US government would still need to fill a funding gap of over $1 trillion.
Social resistance is heating up as public sentiment shifts rapidly. Beyond infrastructure challenges, socio-political resistance is also accumulating quickly. The Maine House of Representatives passed a moratorium on large data centers by a vote of 82 to 62, prohibiting such construction until 2027 and simultaneously establishing the Maine Data Center Coordination Committee to assess the impact of data centers on the state's resources, environment, and finances.
In a recent report, Goldman Sachs Executive Director Shreeti Kapa wrote that during dinner discussions with investors, she sensed a strong consensus: "There simply isn't enough compute. Every participant is facing serious compute constraints—from fabs to data center permitting, to power, memory, and labor. The bottlenecks are real and will persist for a considerable time."
A recent Quinnipiac University poll shows that public wariness about AI's deep integration into healthcare, education, and daily life is rising rapidly, further increasing the social friction costs for industry expansion.
Canaccord concluded: "Energy constraints are intensifying, and socio-political constraints are intensifying as well. Something has to give." In the face of a reality where capital is abundant but project execution is hindered, the path for US AI infrastructure expansion is far more arduous than market expectations suggest.
Comments