Paul Krugman argues that the planned rapid expansion of artificial intelligence infrastructure in the United States may run into practical constraints sooner than many optimistic market participants expect. The Nobel Prize–winning economist highlights two main obstacles: more expensive and selective financing, and growing local and national resistance to new data centers.
How large are the investment needs?
A Brookings/BPEA estimate by Stijn Van Nieuwerburgh of Columbia Business School projects that building US AI infrastructure could require a total of $10.3 trillion in investment between 2025 and 2032. That would equal an average of about 3.63 percent of US GDP per year, a share higher than previous large US investment waves such as rail or highway programs. The estimate is forward-looking and based on announced plans and capacity needs, so it outlines a potential pathway rather than a point forecast.
Krugman stresses that a multi-percent-of-GDP investment requirement would be a severe stress test for capital markets, the energy system and the broader economy, and that markets have started to price in the difficulty of executing such a cycle.
Financing: winners and losers
Financing risk is uneven across companies. Microsoft, Alphabet, Amazon and Meta have diversified businesses and strong cash flows, whereas so-called “neocloud” infrastructure providers—firms that rely heavily on leasing GPU compute and use leverage—are more sensitive to future paying demand for AI capacity. Credit markets are beginning to reflect that split: S&P Global Market Intelligence and LSEG data cited in reporting show Oracle’s five-year CDS spread rose to about 288 basis points in mid-October, while CoreWeave’s spread exceeded 1,000 basis points. Those insurance premia signal rising perceived risk on corporate debt.
Krugman and the article note that the market’s response to date points toward a pricier, more selective and risk-sensitive financing environment rather than a blanket shortage of capital.
Will the technology pay off?
Historical analogies to the dot-com and telecom booms show that infrastructure built during such waves can later prove highly useful, but ill-timed or overleveraged investment can also inflict large losses on companies and investors. For AI, additional uncertainties include cheaper models, open-weight systems, competition from Chinese firms and rapid technological obsolescence—all factors that could compress prices for AI services and put the returns on expensive, highly financed data-center capacity under pressure.
Energy and grid constraints
Beyond capital and chips, data centers need concrete physical inputs: large amounts of electricity, network connections, water for cooling and land. The International Energy Agency (IEA) in 2026 noted rapid growth in AI-optimized data-center capacity: global data-center electricity use reached about 485 TWh in 2025 (up from under 300 TWh in 2020) and, in the IEA’s central scenario, could reach roughly 950 TWh by 2030, although the projection includes a wide uncertainty band through 2035.
The US Energy Information Administration (EIA) flags regional risks: under a faster data-center demand path, US gas-fired generation could rise by 7.3 percent between 2025 and 2027 (rather than an expected 1.7 percent), putting pressure on regions such as Texas’s ERCOT and parts of the PJM grid. The EIA’s short-term outlook also expects the commercial sector, which includes data centers, to raise electricity consumption by about 2.8 percent year-on-year in 2027; wholesale power prices could be on average 11 percent higher in 2026, with much larger increases in some regions such as PJM.
Successful deployment therefore depends not only on capital and chips but on network connections, transformers, water supply, long-term power contracts and local political consent.
Local opposition and permitting
Public resistance to new data centers is high. An ABC News / Washington Post / Ipsos survey found 66 percent of Americans would oppose new data-center construction in the United States, rising to 72 percent for projects in their own communities. An Annenberg Public Policy Center summer survey showed 61 percent of US adults would oppose a new data center in their area (up from 49 percent in spring). Key concerns include rising energy demand, water use, noise, land use and increased grid costs.
Local opposition frequently translates into permitting and regulatory barriers. In July 2026, New York State Governor Kathy Hochul imposed a one-year moratorium on new hyperscaler (large cloud and platform company) data-center construction and permitting while the state develops new rules for large energy-intensive developments. A concrete example from New Mexico — the Project Jupiter debate — shows how a dispute over a roughly 0.6-mile (less than 1 km) section of pipeline crossing state land led Stephanie Garcia Richard, the state Commissioner of Public Lands, to deny a right-of-way for an Energy Transfer gas pipeline. That denial stalled the project and highlights how data-center timelines can be derailed by land-use, water and energy access disputes and local decision-making.
Slowdown more likely than collapse
Krugman suggests that a slowdown in AI investment could be a manageable macroeconomic outcome: postponing or canceling some planned projects would hurt involved firms and investors, but could be less damaging than an abrupt, disorderly market correction. The AI story is far from over, he notes; the core question is the pace, cost and political feasibility of executing very large investment plans.
His message is a pragmatic one: the rise of AI is not only a technology competition but also a test of infrastructure finance, energy systems and permitting processes. Organizations that can manage capital, power, grid access and local political tolerance simultaneously are likeliest to succeed.
(Article synthesized from the figures, surveys and analyses cited in the source material.)
