Why Data Centers Are Being Built In The Wrong Places

[00:00:00] Speaker 1: In 2025, over $156 billion worth of data center projects were blocked or stalled by local communities across the United States. Cancellations quadrupled in a single year. Residents in Virginia, Indiana, and Arizona, places that have historically welcomed large industrial development, are filing lawsuits and winning. And yet the demand for AI compute is only accelerating. The industry has committed somewhere in the range of a trillion dollars to building this infrastructure over the next five years. So where should they actually be built? The answer to that question is more complicated than it looks, because not all data centers are the same thing. And most of the debate about where to build them is conflating two very different problems. When people talk about AI data centers, they're usually picturing the same thing: a massive warehouse full of servers, drawing enormous amounts of power. That image isn't wrong, but it describes two completely different types of facilities that have almost nothing in common when it comes to where they should be built. The first type is a training data center. Training is how an AI model learns. You feed it enormous data sets, run billions of computations, and tune hundreds of billions of parameters until the model can do something useful. This process runs continuously for weeks and sometimes months. It consumes enormous amounts of power. Current GPU racks for training draw around 40 to 160 kilowatts each, and next generation hardware is expected to push past 600. Here's what a training data center doesn't need: proximity to anyone. A training run doesn't interact with users in real time. It doesn't matter whether the data center is 50 miles from a city or 5,000. What matters is that the power is cheap, abundant, and reliable. Training clusters go wherever those conditions are met. The second type is an inference data center. Inference is the model in action: answering your query, generating an image, or running a recommendation engine. Every time you use an AI product, you're hitting an inference facility somewhere. And inference has one constraint that training doesn't: it has to be fast. Consumer-facing AI targets response times under 50 milliseconds. To hit that, the infrastructure has to be close to users, in or near major population centers. Places like Dallas, Phoenix, London, Frankfurt, or Tokyo. These are the markets inference facilities cluster around, because the physics of data transmission don't leave much room for negotiation. By 2030, inference is projected to become the dominant AI workload, representing more than half of all AI compute globally. The industry is building toward a future where most of its infrastructure needs to be near people. But that is creating the fundamental tension. The opposition to data centers has become one of the more surprising political stories of the past two years. In 2023, two U.S. data center projects were canceled due to local resistance. In 2024, it was six. In 2025, that became 25. And in just the first quarter of 2026, more than 20 additional projects were killed, representing over $41 billion in planned investment. At least 188 organized opposition groups are now active across 40 states. Twelve states have filed moratorium bills on new data center construction permits. Maine's legislature passed a statewide moratorium, the first in the country, through 2027. The grievances vary by community, but water comes up most often. More than 40 percent of contested projects cite water consumption as the primary concern. A facility in Tucson, Arizona, Project Blue, a $3.6 billion campus backed by Amazon Web Services, was unanimously rejected by the city council in August 2025 after residents calculated it would consume 283 million gallons of water per year. That's equivalent to the annual usage of more than 3,000 households. The council voted against it unanimously. In Memphis, Tennessee, the controversy took a different shape. XAI's Colossus Data Center campus was found to be operating 33 unpermitted methane gas turbines. Turbines with the potential to emit over 1,700 tons of nitrogen oxide annually in a neighborhood already exceeding federal smog standards. These cases are different in character, but they share a pattern. What's counterintuitive about the geography of the backlash is where it's concentrated. Most cancellations are in red counties in states that voted for Trump in 2024, communities that have historically supported economic development and industrial investment. Texas, which has a massive data center pipeline and a different regulatory structure, saw zero cancellations due to local opposition in 2025. Part of what's driving the resistance is that inference facilities, which need to be near people, are creating friction with exactly the communities they need to be adjacent to. There's no easy resolution to that. But for training facilities, the constraint is different as they don't need to be there. For training data centers, power has replaced location as the primary site selection criterion. That's a relatively recent shift, and its implications are still playing out across the industry. Grid interconnection wait times in major US markets now run between five and seven years. The bottleneck isn't capital. It's transformers with lead times of three to five years and switchgear that's sold out through 2028. The areas that are best for this are the ones that don't have those constraints. Texas leads the consensus. Texas's Electric Council's independent grid allows faster interconnection than the major interstate networks. Land is cheap, power is abundant, and the state's regulatory posture has kept community opposition minimal. JLL suggests that Texas could overtake Northern Virginia as the world's largest data center market by 2030. The second best considerations are the Nordics, Sweden, Norway, Finland, and Denmark. They offer cheap, clean power from hydro and wind, cool ambient temperatures that cut cooling costs, and governments that have been more permissive about large industrial development. Microsoft has committed over 12 billion dollars to Norway and Sweden alone. Then we have the Gulf states, particularly Saudi Arabia and UAE, that rank near the top of industry scoring models on energy availability and permitting speed. These are the places where training infrastructure is migrating. The question of where training data centers should be built for the near term has a fairly clear answer: away from population centers and toward areas of power. The harder question is inference data centers, and that's where the frontier solutions, the ones that are still experimental, start to really matter. One of the more direct responses to the data center constraints is to go underwater. Microsoft spent seven years investigating that concept. Project Natick began in 2015 with a small pod submerged off the California coast. Phase 2 in 2018 deployed 855 servers in a sealed container off Scotland's Orkney Islands, powered by the island's wind and solar grid, sitting 35 meters underwater for two years. When they retrieved it in 2020, the failure rate was roughly one-eighth what it would have been on land. The sealed nitrogen environment eliminated corrosion and the absence of humans eliminated physical disturbance. With this, Microsoft declared it feasible. But then they stopped. In 2024, Microsoft's head of cloud operations said simply it worked, but they were applying the learnings elsewhere. China then came in and took it to commercial scale. Shanghai Heilanyuan technology, HiCloud, in partnership with state-owned China communications construction company, deployed a 24 megawatt facility off the coast of Shanghai's Lingang area, submerged about 10 meters below the surface. It draws more than 95% of its power from a nearby offshore wind farm. It uses direct seawater cooling, eliminating freshwater entirely, and cutting land use by more than 90% compared to an equivalent land-based facility. The facility went live in May 2026, just a few weeks ago. The efficiency numbers are significant here. HiCloud claims a power usage effectiveness below 1.15. The industry average for terrestrial facilities is around 1.5. That gap represents a meaningful reduction in wasted energy at scale and makes this truly viable. Some of the limitations are real too. Maintenance requires lifting entire modules to the surface. Environmental effects, particularly localized ocean warming, are poorly understood at commercial scale. The concept solves the power and water problems elegantly. But it doesn't solve the latency problem for inference. A submerged facility is still a fixed point in space, and the distance to users still determines response time. What underwater infrastructure may eventually offer, though, is being closer to coastal population centers without the land constraints those cities impose. This may be the best outcome for both sides. But the most ambitious response to the data center constraint is to leave Earth entirely. Three credible programs are now in development. Google's Project Suncatcher, announced in late 2025, envisions constellations of solar-powered satellites carrying Google's own TPU chips linked by laser communication in low Earth orbit. Google's own preprint found that its chips survived radiation exposure significantly beyond the expected five-year mission dose without failure. Two prototype satellites are targeted for early 2027. StarCloud, founded in early 2024, launched a satellite carrying an NVIDIA H100 in November 2025 and claimed to have trained a small language model in orbit. It raised $170 million at a $1.1 billion valuation in March 2026, making it the fastest unicorn in Y Combinator history. And three days ago, SpaceX unveiled AI-1, its first-generation orbital data center satellite, timed to the week of the company's IPO. The craft has a 70-meter wingspan, wider than a Boeing 747. It's designed to carry 120 kilowatts of sustained compute in low Earth orbit, powered by solar arrays and cooled by a 110 square meter deployable liquid radiator. SpaceX has filed with the FCC for a constellation of up to 1 million satellites. Musk described AI-1 as "much simpler than a Starlink satellite." There are a lot of real advantages to a data center in orbit. A solar panel in the right orbit is exposed to sunlight far more consistently than any ground installation. There's no land, no fresh water, and no community opposition. And heat can radiate directly into the vacuum of space, which sits near absolute zero. But that is easier said than done. The physics of that last point is where the vision runs into hard constraints. In a vacuum, heat can only be removed through radiation. There's no air, no water, and nothing to carry it away. Starcloud's own white paper calculates that a two-sided radiator at operational temperature emits roughly 633 watts per square meter. Water cooling on Earth moves heat over a thousand times faster. A single megawatt of orbital compute would require roughly 1,600 square meters of radiator. At scale, the radiators start to outweigh the servers. The other constraint is economics. Current launch costs run between $1,500 and $3,600 per kilogram. Google's own analysis projects that orbital compute becomes cost-competitive with terrestrial infrastructure only if launch costs fall to around $200 per kilogram. A threshold that depends on Starship-class launch economics, and one Google doesn't expect to arrive until the mid-2030s at the earliest. Space is a serious engineering effort. The capital is real, but the hardware is still being built. But gigawatt-scale orbital compute is a 2030s proposition, contingent on launch economics that don't yet exist. So where does this land for today's problems? For training data centers, the answer is increasingly clear, and the industry is arriving at it. Power-rich, permitting-friendly, sparsely populated markets like Texas, the Nordics, and the Gulf states are where large training clusters should go, and are starting to go. The backlash in suburban Virginia and rural Indiana is partly a consequence of building facilities that had no reason to be there in the first place. For inference data centers, the answer is harder and the timeline is longer. Inference facilities need to be near people, and the markets near people are running out of power, land and community tolerance simultaneously. In the near term, the industry's best options are better grid infrastructure, faster interconnection reform, and, where possible, smaller, denser facilities that create less friction with the communities they operate in. The medium-term answer, and from my perspective, the actual answer, may be underwater. Coastal proximity without land constraints, passive cooling, and no fresh water. The concept works, and China has now proved it at commercial scale. The question is just whether it can work at the scale inference requires. The long-term wildcard is space. If launch economics follow the trajectory, SpaceX and Google are betting on. Orbital constellations could eventually deliver low-latency compute globally, without land, without water, and without the community opposition that is currently costing the industry billions of dollars a year. That outcome is plausible, but it's not immediate. What's clear is that the current approach, building the same type of facility everywhere, regardless of what it does, is failing. The $156 billion blocked or stalled in 2025 is the cost of that confusion. The industry knows where it needs to go. Getting there is the problem. And the people in control of that may be the problem too.