Twenty meters. That was the distance between the nearest bedroom window and the perimeter fence of Project Mars, a hyperscale data center that Goodman, one of the world's largest industrial landlords, quietly abandoned on the outskirts of Sydney. No court order ended it. No regulator signed its death certificate. A community simply refused, and the refusal was enough. Half a world away, in the state of Victoria, legislators wrote three clauses into planning law that will reshape an entire industry: data centers must supply their own renewable power, they may not cool their servers with drinking water, and they must stand at least 150 meters from any home. Read those three lines slowly. They are not environmental policy. They are a verdict on a business model that assumed geography was a resource to be consumed rather than a relationship to be negotiated. The most advanced industry on earth has met its first real opponent, and that opponent has no IPO, no cap table, and no spokesperson. It has a mailbox, a town council, and a grievance.
For three years, the dominant story of artificial intelligence has been a story of scarcity: not enough chips, not enough power, not enough grid. The market has learned to price silicon and electricity with remarkable precision. It has not learned to price consent. That is the blind spot, and it is now the most expensive line item nobody is modeling.
For a decade, the operating assumption of the hyperscaler was frictionless expansion. Land was cheap, power was available, and communities were grateful for the tax base. That assumption held as long as data centers were modest, quiet, and water-frugal. AI broke it. The new generation of facilities is enormous, power-hungry, and โ critically โ visible. A single campus can draw as much electricity as a small city and consume water at a rate that competes with residential supply. The shift from a background utility to a foreground neighbor is what converted a zoning matter into a movement.
The scale is no longer anecdotal. According to tracking by Data Center Watch, roughly 120 projects worth close to $200 billion were blocked or delayed across the United States in the first half of the year alone. A parallel European monitor counts some 70 contested sites across the continent. The opposition is no longer a fringe of rural homeowners. Quinnipiac polling recorded disapproval rising from 65 percent in March to 72 percent by late September โ a seven-point drift in six months, in a political climate where almost nothing moves seven points.
The geography is the part that should unsettle anyone who has spent a decade watching infrastructure build out. The revolt is not American. It has surfaced in South Korea, where residents have picketed on weekday mornings for months. It has surfaced in Japan, where a lawsuit in Inzai names noise, hot exhaust air, and the loss of sunlight as injuries. It has surfaced in Malaysia, where more than fifty people gathered in the country's first such protest. It has surfaced in Australia and across Europe. Six regions, one pattern. When a phenomenon appears simultaneously in six regulatory cultures โ from common-law Australia to civil-law Japan โ it is not a local mood. It is a structural condition, and structural conditions do not resolve on their own.
I should flag my own epistemic caution here, because the numbers carry weight and I want to be honest about their limits. The $200 billion figure has no denominator. Nobody has published the total value of data centers planned in the same period, which means we cannot yet say whether the revolt is a rounding error or a systemic brake. The tracking organizations have not disclosed their methodologies. I treat the direction as reliable and the magnitude as provisional. That distinction matters, and I will return to it.
I want to be precise about what is actually being built here, because the debate is usually conducted in the wrong vocabulary. These are not software companies. They are power plants that happen to compute. The technology question that matters is not which model runs inside the racks. It is how the heat leaves the building and where the electrons come from.
The Victorian rules are a direct attack on evaporative cooling, the dominant thermal strategy for a generation of data centers, which consumes somewhere between 1.5 and 2.0 liters of water per kilowatt-hour. Ban drinking water for cooling and you have not asked operators to be nicer. You have invalidated a technical lineage. The forced migration runs toward air-cooled systems, closed-loop water recovery, and increasingly toward liquid and immersion cooling, which were already being pulled forward by a different force entirely: the power density of modern AI accelerators. A rack of GB200-class hardware can exceed 120 kilowatts, a thermal load that air simply cannot carry. So two unrelated pressures โ community anger and silicon physics โ are converging on the same conclusion. The revolt is, unintentionally, the best sales force liquid cooling ever had. When I ran workshops bridging AI developers and blockchain engineers last year, the recurring theme was that trust has to be engineered into the substrate. The same is true of heat. You can no longer bolt a cooling strategy on after the fact; it has to be designed in from the first line of the site plan.
If-then logic applies here, and it is worth stating plainly. If a project must self-supply renewable energy, then it stops being a tenant of the grid and becomes an operator of generation assets. That is not a small engineering footnote. It means on-site gas with carbon capture, or small modular reactors, or solar-plus-storage, all of which multiply both the capital cost and the timeline. The numerator of the return โ revenue from cloud and AI services โ does not change. The denominator โ front-loaded cost and schedule โ inflates. A single megawatt of capacity can carry 15 to 40 percent more upfront capex under these constraints. That is the arithmetic the industry is only beginning to write down.
The economics of retreat are worth separating carefully, because the market tends to treat "blocked" as a single category. It is not. A cancelled project loses sunk cost โ painful but bounded, a write-down the accountants can absorb. A delayed project loses something harder to price: the net present value of revenue deferred, and worse, the market window it was built to serve. If demand for AI compute is front-loaded โ and every signal suggests it is โ then a two-year delay is not a two-year loss. It is a permanent forfeiture of the most valuable years of the asset's life. Goodman's quiet withdrawal from Project Mars is the visible tip of this: a leading industrial REIT, with every incentive to build, chose to walk. When the landlord walks, the model is broken, not the landlord.
The second-order effect is geographic rebalancing, and it will reshape the map of compute. Projects that cannot proceed in coastal Virginia or suburban Sydney will migrate toward places with surplus power, permissive regulators, and communities that have not yet learned to say no โ the American Midwest, parts of the Middle East, select Southeast Asian jurisdictions, and eventually offshore platforms. This is not a solution. It is a deferral. Every destination has a community, and every community eventually learns what the previous one learned. The revolt is not being defeated; it is being relocated, and relocation buys time without buying consent.
But the deeper shift is subtler. What the communities are demanding, and what Amazon has begun to pay, is something the industry has no clean name for. Call it social license. Amazon's $1 billion community commitment is the clearest signal yet that consent has become a purchasable input. And here is where my own decade in this industry gives me a frame that most analysts lack: I have watched a different technology sector spend years trying to solve exactly this problem, and fail in instructive ways.
Blockchain was, at its origin, a machine for generating consensus without a central authority. Bitcoin's proof-of-work was not merely a security mechanism; it was a mechanism for making agreement expensive to fake and therefore trustworthy. The lesson the AI infrastructure industry is now learning the hard way is that consensus cannot be purchased at the margin. You can buy goodwill, but you cannot buy legitimacy, because legitimacy is a belief held by people you do not control. Amazon's billion dollars is a community benefit agreement โ a CBA โ and a CBA is a transaction. Transactions buy compliance. They do not buy the durable, voluntary alignment that keeps a system running when the money stops flowing.
I watched this failure mode up close during the DeFi summer of 2020, when I spent three months interviewing users who had lost savings to oracle failures. The smart contracts were flawless. The humans were not. Twelve people told me the same story in twelve different ways: the code did exactly what it promised, and the promise was not the thing they had actually been given. That gap โ between the formal correctness of a system and the lived reality of the people inside it โ is the entire story of the data center revolt. The system is technically sound. The people have been handed the wrong promise.
The vocabulary of the revolt is itself revealing. In Inzai, the complaint is not about abstractions like privacy or alignment. It is about sunlight, noise, and hot air โ the physical residue of computation dumped onto people who receive none of its upside. That is a distributive justice problem, and it maps precisely onto a debate the crypto world has been having for years: who bears the cost of a public good, and who captures the benefit. When I audited the tokenomics of failed projects years ago, I found the same failure mode repeated: value accrued to distant holders while risk pooled in a place nobody was looking. The data center is the physical instantiation of that asymmetry. The cloud revenue flows to shareholders in Manhattan and Menlo Park. The water bill, the noise, the lost view โ those stay local. The ledger remembers the revenue. The community remembers the cost.
There is a dimension the coverage has largely skipped, and it is the one I find most ethically load-bearing. Environmental burdens are not distributed randomly. In the United States and Europe, polluting and disruptive infrastructure has historically been sited in lower-income and minority communities โ a pattern documented across decades of environmental justice research. If data centers follow that pattern, then the revolt is not merely a clash between capital and residents. It is a question of who gets chosen to host the externalities of someone else's prosperity. The industry has an opportunity here that it is squandering: by being transparent and by sharing value locally, it could break a historical pattern rather than repeat it. Transparency is the cheapest concession available and the one most conspicuously withheld. Nobody has published the water and power consumption of the contested sites. In the absence of data, communities assume the worst โ and they are not being paranoid. They are being rational under uncertainty.
There is a second, more troubling parallel. I have written before about Tornado Cash, and I will not rehearse it here except to note the shape of the precedent: writing code, publishing it, and letting others use it was treated as a crime. The data center revolt is the mirror image. Here, the act of building physical infrastructure โ the most legible, taxable, regulated activity imaginable โ is being treated, in effect, as a trespass. In one case, the law reaches too far into the abstract. In the other, communities reach into the physical to stop what capital wants. Both are symptoms of the same underlying condition: the institutions that are supposed to mediate between builders and the built-upon have stopped working. The bottleneck of the AI era is not compute. It is the absence of a legitimate mechanism for distributing the costs of compute. That is a governance problem wearing an engineering costume.
Consider the escalation to national security, because it is the most cynical move on the board and the most revealing. Matt Garman, the chief executive of AWS, has framed local opposition as a threat to American AI leadership. Read that sentence again. A dispute over a substation in a Virginia county and a setback rule in Victoria has been reframed as a question of geopolitical survival. This is textbook issue escalation: when you cannot win at the local level, you relocate the argument to a venue where your interests and the national interest are indistinguishable. The carrot is the billion dollars. The stick is the fear of losing to China. The prize, if the framing holds, is federal preemption โ a legal tool that would let Washington override the zoning decisions of towns that do not want to be overridden.
I find this move both clever and dangerous. Clever, because it is not entirely wrong: China does build data centers faster, and one reason is that its land and power allocation runs through administrative coordination rather than community negotiation. The absence of a veto point is a genuine competitive advantage. Dangerous, because the same framing tells every affected community that their objection is unpatriotic, and voters do not respond well to being told their concerns are a national security liability. You can escalate an issue into the Capitol. You cannot escalate away the people who live next to the substation.
Here is the angle that I think the market is systematically getting wrong, and it is worth being blunt about it. Everyone is modeling the wrong constraint. The consensus view holds that AI's limits are chips and electricity, and that both will yield to capital and time. I do not think that is the binding constraint at all. The binding constraint is that the industry has no mechanism for manufacturing consent, and consent โ unlike silicon โ does not scale with capex. You cannot fab legitimacy. You cannot order it in a backlog. It is produced, if at all, through years of slow, local, unglamorous political work, and that work has a timeline measured in election cycles, not quarters.
There is a specific mechanism most analysts overlook: the interconnection queue. In the United States, a new large load can wait three to seven years simply to be studied for grid connection, before a single permit is contested. Layer community opposition on top of that queue and the effective delay compounds multiplicatively rather than additively. A project blocked for two years may sit in a seven-year queue behind it, meaning the true time-to-energization stretches toward a decade. No capex line captures that. No earnings model prices it. And unlike a chip shortage, which a fab can relieve in eighteen months, a legitimacy deficit is resolved only by the slow accretion of trust โ or not at all.
This is why I keep returning to a lesson the crypto world learned painfully and the AI world is about to learn expensively: decentralization is not a technical property; it is a social achievement, and it can be revoked. A network is only as decentralized as the willingness of its participants to keep participating. A data center is only as buildable as the willingness of a community to keep tolerating it. The AI industry has spent a decade assuming that because it can compute anything, it can build anywhere. That assumption is now being tested by the oldest technology in existence: the town meeting.
And there is a counterintuitive corollary. The projects that will survive are not the ones with the most capital. They are the ones that understood, early, that a CBA is not a cost but an asset โ a locked-in social license that functions as a moat. I have spent enough time in public goods funding debates to know that the crude, transactional version of community compensation โ write a check, get a permit โ has a short half-life. Expectations ratchet upward. The first community that accepts a billion sets a floor for the next hundred. What compounds instead is the slow accumulation of genuine reciprocity, the kind that Optimism's RetroPGF was designed to reward by paying for public goods after the fact, based on demonstrated contribution rather than promised intent. The data center industry needs its own version of that mechanism, and it does not have one. It has a checkbook, and checkbooks run out of goodwill long before they run out of money.
There is also a quiet irony worth naming. The AI industry is being asked to prove, in the physical world, a claim the crypto industry has been asked to prove for fifteen years: that a powerful system can serve the many without extracting from the few. Faith in the protocol is not faith in the people. The protocol can be perfect and the people can still walk away. That is the lesson of every abandoned data center parcel from Sydney to Inzai, and it is the lesson the industry keeps refusing to read.
So watch the next earnings cycle, not for the capex numbers, but for the silence around them. Watch whether the megawatt figures of the blocked projects are ever disclosed โ because dollars can be spun and megawatts cannot. Watch the polling, and ask what happens if opposition crosses 75 percent. The revolt is not a public relations problem to be managed. It is a consensus problem to be solved, and no treasury is large enough to buy the answer. We built the temple, but forgot who the god is. The question, in the end, is whether the builders will learn to ask the people who have to live beside the altar.