Within my own network, I was early in flagging the growing local opposition to AI data centers. Since then, it has moved fast, from scattered zoning fights to outright moratoriums and statewide political battles, with data centers becoming a poster child for many of the broader anxieties around AI. A lot of that opposition is justified. These projects strain power systems, raise real questions about who ends up paying for them, and impose costs on the communities hosting them, costs that do not always show up in the press releases. But there is another side to the argument. If AI keeps spreading through software, science, manufacturing, medicine, finance, and government, we will need more compute.
Subscribe to our weekly newsletter
This short guide is my attempt to lay out that bull case as clearly as I can. Not because every proposed data center deserves to be built, and not because the industry has earned a free pass, but because the arguments for building more are stronger than the public debate usually gives them credit for. What surprises me is how rarely the AI industry makes that case in a way that is concrete, credible, and responsive to the people being asked to host these facilities.
Where the Local Economic Value Comes From
The Economic Case for the AI Buildout
Why the Power Case Beats the Compute Case
The Geopolitics of Compute
The Policy Levers That Actually Work
Where the Local Economic Value Comes From
The local jobs argument is easy to overstate. Data centers are enormous capital projects, but once they are running they do not employ enormous numbers of people. I think the stronger case starts elsewhere. Construction can support hundreds or even more than a thousand workers for several years, while the finished facility can add substantial recurring tax revenue to a local government. Data centers can also pull new fiber, power, and other infrastructure into an area. For many communities, the impact on the tax base may ultimately matter more than the permanent payroll.
The case gets more interesting when a data center gives an old site a second life. Consider an abandoned factory, a failed industrial development, or a contaminated property that other buyers have avoided for years. A deep-pocketed data center developer may be willing to clean it up, reuse infrastructure that is already there, and put otherwise idle land back on the tax rolls. In situations like these, the alternative may not be a data center versus a factory employing thousands of people. It may be a data center versus another decade of nothing happening. Add property taxes and funding for schools, fire departments, or other local priorities, and I can see why some communities would take that deal. The important part is making sure they keep enough of the upside rather than giving it away through subsidies.
The Economic Case for the AI Buildout
Zoom out from individual communities, and the bull case gets broader. Compute is starting to look less like another technology product and more like a basic economic input (“the new oil”). If AI keeps working its way into software, research, medicine, manufacturing, finance, and other large parts of the economy, access to abundant computing capacity will matter in much the same way that access to power, networks, and cloud infrastructure matters today. That argument does not require any particular AI company or model to win. It requires only that organizations keep finding valuable things to do with more compute.
There are also reasons to think the current buildout has more underneath it than pure speculation. Much of the spending comes from profitable hyperscalers, some new capacity already has customers attached to it, and enterprise AI usage continues to expand. The numbers are enormous, with US AI investment projected at roughly $581 billion in 2026, but at about 1.8 percent of GDP they are not yet outside the range of past technology buildouts. The US also has unusually deep capital markets capable of financing projects this large. None of that means every data center will earn an attractive return. Efficiency gains could reduce demand, customers could pull back, and leverage can make mistakes expensive. But I would separate those risks from the stronger claim that the entire investment cycle must be irrational simply because the numbers are big.
[The financing of AI data centers is genuinely worrisome, and I’ve written about this topic in a recent post.]
Tweet text appears here…
— Account Name (@account)
August 23, 2026
Why the Power Case Beats the Compute Case
One part of the buildout may have more durable value than the data centers themselves. The US already needs substantial investment in electricity generation and the grid as older power plants retire and demand grows elsewhere. AI is arriving with customers willing to spend enormous amounts of money on power, which could accelerate investment in generation, transmission, substations, and storage. It can also provide anchor customers for technologies such as advanced nuclear, geothermal, and batteries that are expensive to get off the ground. If AI demand comes in below today’s forecasts, much of that infrastructure can still serve factories, electric vehicles, growing cities, and other industries. A transmission line ages very differently from a GPU.
There is an important caveat. The near-term scramble for electricity could keep gas and coal plants running longer, and “America needs more power” does not mean every proposed data center is a good idea. Still, the bottleneck increasingly looks physical. Power plants, transformers, transmission lines, and grid connections take much longer to build than servers. Data centers may also become more useful grid customers than their reputation suggests. Some computing jobs can be shifted to another time or location, while batteries and onsite generation can reduce demand when the grid is stressed. I would not assume all of this works smoothly, but it creates an interesting possibility: the AI boom could leave behind a larger and more capable power system even if some of the compute investments turn out to have been too aggressive.
The Geopolitics of Compute
AI data centers become harder to dismiss when national security enters the picture. Frontier data centers can hold valuable model weights, sensitive government and corporate data, and workloads used for intelligence, cybersecurity, and military applications. Some of that capacity probably should remain inside the US, under US law and protected against foreign surveillance, coercion, or disruption. A small number of facilities may even look more like defense infrastructure than ordinary commercial data centers, especially when they are designed specifically to protect sensitive models and workloads from nation-state attackers.
But I think the bigger geopolitical argument is about the full stack. AI leadership increasingly depends on chips, electricity, data centers, networks, financing, models, and deployment expertise working together. China can package many of those pieces when competing for influence abroad, while countries that host scarce compute can use that infrastructure as bargaining power. The US does not need to build everything at home. Canada, Australia, Europe, and other allies can contribute energy, capital, semiconductor capacity, minerals, and talent to a larger trusted compute base. The mistake would be to make domestic construction so difficult that strategically important infrastructure simply moves to countries whose interests are less aligned with ours. That still does not make every proposed hyperscale campus a national security project. It does mean that where large-scale compute gets built is becoming a geopolitical choice, not just a real estate decision.
The Policy Levers That Actually Work
There is also a policy lesson buried in the economics. Governments have spent years competing for large technology projects with tax incentives, but for AI data centers the scarcer resource may simply be time. In one modeled 100 MW project, a one-year delay destroyed roughly $550 million of lifecycle value, more than either doubling the price of electricity or losing typical state tax incentives. I would not put too much weight on the exact number, but the ranking makes sense. Expensive GPUs generate no return while they sit waiting for a grid connection. That suggests states may get more leverage from faster permitting, shorter interconnection queues, and more transmission than from writing ever larger subsidy checks.
But faster cannot just mean pushing costs onto everyone else. If a new data center requires a substation, transmission line, or generating plant, the data center should pay for it rather than existing ratepayers. And I think the same logic applies to the politics. A project stuck for years in lawsuits, referenda, and local opposition is not actually moving fast. Transparent negotiations, enforceable ratepayer protections, sensible siting, and credible community benefits can make projects easier to approve and harder to derail. In that sense, public consent is not necessarily the enemy of speed. Done well, it can be one of the things that produces it.
Related Content