The numbers surged. The room felt empty.
Somewhere in the sprawl west of Sydney, a data center campus that was meant to anchor a multi-billion dollar AI buildout was abandoned before a single rack of servers ever arrived. It was not defeated by a chip shortage. It was not defeated by a failed power contract or a cold-eyed board. It was defeated by distance โ twenty meters, to be precise, between the proposed facility and the nearest family's back fence. The project, known internally as Mars, simply stopped. No bankruptcy filing, no dramatic regulatory ban. Just a quiet retreat, the way a tide pulls back from a shoreline it has decided not to claim.
I have spent my career watching infrastructure die for reasons that had nothing to do with the technology inside it. In 2017, I left a corporate security role to join Gitcoin as a lead contributor, and I spent months auditing prototype smart contracts for a quadratic voting mechanism that was supposed to make public goods funding fair. The code worked. The math was elegant. And still, projects failed โ not because the cryptography was wrong, but because the humans around it had never consented to the terms. When the graph spikes, the soul remains quiet. That lesson keeps returning to me, and it is returning again now, at the edge of a grid that AI is trying to swallow whole.
Because the Sydney site is not an anomaly. It is a data point in a pattern that has gone global with a speed that should unsettle anyone who builds infrastructure for a living.
Context: a revolt with no borders
In the first half of the year, roughly 120 data center projects in the United States โ carrying a combined value of nearly 200 billion dollars โ were blocked or delayed by local opposition. Across Europe, another 70 projects stalled under similar pressure. The geography is the story. This is not a New Jersey problem or a Northern Virginia problem. It is happening in Australia, South Korea, Japan, Malaysia, and across the European Union simultaneously. Six regions, one phenomenon.
The shape of the conflict is remarkably consistent. In Seoul, residents organized early-morning protests on weekdays and kept them going for months. In Inzai, a lawsuit argued that a proposed facility would bring noise, hot exhaust air, and the loss of sunlight to surrounding homes. In Malaysia, more than fifty people gathered at a single site โ a first for the country. In Australia's Victoria state, regulators responded with a package of rules that reads less like a permit condition and more like a verdict: no potable water for cooling, mandatory self-supplied renewable energy, and a 150-meter setback from residential boundaries.
The public opinion underneath all of this is hardening, not softening. Polling tracked by Quinnipiac moved from 65 percent opposition in March to 72 percent by the end of September. That is not a transient mood. That is a trend line bending away from the industry. And when opposition crosses the threshold from local grievance to majority sentiment, it stops being a permitting nuisance and becomes something closer to a political weather system.
The industry's response has split into two tracks. The first is money: Amazon committed a billion dollars to community promises, a defensive gesture dressed up as generosity. The second is escalation. Matt Garman, the CEO of AWS, framed local opposition as a threat to America's AI leadership โ a move that reframes a zoning dispute as a question of national survival.
Both tracks are revealing. Neither one is working.
Core: the bottleneck is moving, and it is moving somewhere capital cannot reach
Here is the insight I want to sit with, because I think it is the most important thing happening in AI infrastructure right now and almost no one is pricing it correctly: the binding constraint on AI compute is migrating from chips to something far harder to manufacture โ buildable land, grid capacity, water, and the consent of the people who live nearby. Three of those four can be bought. The fourth cannot.
The industry has spent two years obsessing over GPU supply. It has modeled transformer shortages, advanced packaging bottlenecks, and the power draw of a single rack climbing past 120 kilowatts. All of that is real. But every one of those constraints has a capital solution โ you build fabs, you sign power contracts, you buy cooling. Social license does not have a capital solution. You cannot buy consent at scale, because the moment you buy it in one place, you raise the price everywhere else.
I watched this exact dynamic play out in DeFi, and I have the scars to show for it. In 2020, during the chaos of DeFi Summer, I was a senior product manager on a liquidity protocol, and I refused to deploy incentive programs that rewarded speculation over genuine usage. The investors in the room wanted numbers. They wanted total value locked to spike. I spent three months negotiating with core developers to restructure reward distributions toward long-term stability rather than a short-term headline. My concerns were dismissed, more than once, as naive โ the soft objections of someone who did not understand that growth was the only metric that mattered.
They were wrong, and the market eventually proved it. When the incentives stopped, the liquidity vanished, because it had never been real. Liquidity mining was never adoption. It was a subsidy wearing adoption's clothes. When the graph spikes, the soul remains quiet. The same thing is about to happen to the billion-dollar community check. It will buy a headline. It will not buy a neighbor.
Let me be concrete about why the Victoria rules matter, because they are the clearest window into where this is all heading.
When a jurisdiction bans potable water for cooling, it is not making a symbolic gesture. It is outlawing the dominant cooling technology outright. Traditional evaporative cooling consumes somewhere between 1.5 and 2 liters of water per kilowatt-hour โ a figure that becomes absurd when you are trying to cool racks drawing more than 120 kilowatts each. The rule forces a migration to air cooling, liquid cooling, or closed-loop water recovery. That migration is expensive, and it is mandatory. Regulation has just become the single most powerful accelerator of liquid cooling adoption in the industry's history, and it was written by a state government responding to its constituents, not by a vendor responding to a roadmap.
The self-supplied renewable energy requirement is even more consequential, because it changes what a data center is. A facility that must generate its own power is no longer a tenant of the grid. It is a power operator that happens to run servers. The economics invert. The red tape multiplies. The location calculus changes entirely, because now you are not just siting a building โ you are siting a generation portfolio. The freedom to plug into someone else's infrastructure, which every operator has taken for granted for two decades, quietly disappears.
And the 150-meter setback deletes entire categories of land from consideration. Suburban parcels that were the bread and butter of data center siting become unusable overnight. The industry built its expansion model on the assumption that cheap land near population centers would always be available. Victoria just told that assumption it was wrong.
Stack these three rules together and you get something the industry has never had to model before: a jurisdiction that has effectively decided certain kinds of projects cannot be built there at all, regardless of how much money is on the table.
Now multiply it. Because the pattern is global.
The economics of this shift are brutal and underappreciated. A blocked project does not simply disappear from a spreadsheet. Delayed projects are, in many ways, worse than cancelled ones โ they occupy land and grid capacity while producing nothing, they tie up capital in a holding pattern, and they miss the market window they were built for. The industry keeps treating delay as a scheduling problem. It is a value destruction problem. When you hold a parcel and a power allocation for a facility that cannot break ground, you are paying the carrying cost of a promise you cannot keep.
And then there is the cost that almost nobody has modeled yet: the community benefits agreement. What Amazon did with its billion dollars is not charity. It is a template. The moment one hyperscaler sets a community payment expectation, every subsequent project in that region inherits the new baseline. The cost of entry rises for everyone. What was once a case-by-case public relations exercise is becoming a standardized line item โ a community tax that belongs in every project's internal rate of return calculation, whether the finance team has admitted it yet or not.
I have seen this movie before, and I know how it ends. In the NFT world, I once refused to sign off on a royalty enforcement mechanism because the implementation would have penalized the very secondary-market creators it claimed to protect. I spent two weeks drafting alternatives that balanced platform revenue against creator rights, and it cost me goodwill with leadership. But the lesson was clean: when you optimize a system purely for extraction, the people the system depends on eventually revolt, leave, or both. You cannot build durable infrastructure on a foundation of resentment. The code does not care, but the community always does.
There is a technical irony buried in all of this that I find genuinely fascinating. The community pressure and the engineering pressure are converging on the same answer. High-density AI racks need liquid cooling anyway โ the physics demands it. And communities are demanding that data centers stop consuming drinking water and stop dumping heat into the sky. Two completely different forces, one technological destination. The opposition movement, which the industry regards as an enemy, is accidentally accelerating the exact cooling transition the industry needed to make on its own. The crowd is sometimes a better roadmap than the vendor.
But I want to be honest about the limits of what we can know here. The 200 billion dollar figure and the 120-project count come from sources whose methodology has not been disclosed โ data center opposition trackers whose institutional character and potential advocacy stance are opaque. We are told that projects were blocked or delayed, but not how many were cancelled outright versus relocated versus simply paused. We are given a dollar figure without the denominator: how many projects were planned in total, and what fraction do these 120 represent? Without that baseline, we cannot tell whether we are looking at a rounding error or a systemic crisis. The trend is legible. The magnitude is not.
That uncertainty cuts both ways, and I want to name it plainly. If the blocked projects represent a small slice of total planned capacity, then the revolt narrative is being oversold by advocates who benefit from the drama. But if the megawatt-level losses are significant โ and dollar figures always flatter the story more than megawatts do โ then the market is underestimating one of the most stubborn physical constraints in the entire AI supply chain. The number that would settle it is capacity in megawatts, not value in dollars, and that number is conspicuously absent.
There is a variable in this equation that almost nobody mentions, and it may be the most important one of all. Even when a data center finds land, wins its permits, and secures community support, it still has to connect to the grid โ and in the United States, the interconnection queue routinely runs three to seven years deep. That means the delay caused by community opposition does not happen in isolation. It stacks on top of a delay that was already baked into the system. A project that loses a year to a zoning fight does not lose a year. It loses a year plus whatever position it forfeits in the queue behind it.
This is where the physical constraint becomes genuinely unmanageable. You can compress a construction schedule. You can accelerate a chip fab. You cannot compress a queue governed by transmission upgrades measured in half-decades. The revolt is not creating a new bottleneck. It is pouring concrete into one that was already there, and the concrete sets fast.
Capital, meanwhile, does not sit still when it meets a wall. It moves. And that movement is already reshaping the map.
The projects that cannot be built in Victoria, in New York, in the suburbs of Sydney, are not disappearing. They are looking for ground where the grid is abundant, the water is cheap, and the neighbors are quiet. The American Midwest. Parts of the Middle East, where land and power are allocated by decree and opposition has no venue. Southeast Asia. And increasingly, offshore โ platforms and coastal sites beyond any municipality's jurisdiction.
This is a rebalancing, and it has a cost that the industry has not begun to price. Every project that relocates away from a population center moves further from the users it serves, which raises latency and transmission cost. Every project that relocates toward a permissive jurisdiction trades community consent for regulatory risk. The map of AI compute is being redrawn not by engineering, but by politics โ and the new map is less efficient than the old one, because efficiency was never the variable that mattered most.
For those of us who watch markets, the transmission channel is what matters. And the channel runs in a specific direction: delay in construction becomes delay in compute supply, which becomes delay in GPU orders, which becomes delay in the equipment that supports them โ transformers, backup generators, switchgear, the unglamorous plumbing of the entire AI economy.
In the short term, paradoxically, the constraint may help existing operators. If new supply is delayed, the compute that already exists becomes scarcer and more valuable. Scarcity premiums accrue to whoever locked in capacity before the revolt. But in the medium term, the same delay undermines the expansion expectations that the entire AI application layer is priced against. The market is currently modeling AI as a supply story โ as much compute as we can build. It is actually becoming a consent story โ as much compute as the neighbors will allow.
The asymmetry is where the opportunity hides. Operators with locked-in land, secured power, and genuine community relationships suddenly hold a moat that capital cannot replicate quickly. Liquid cooling vendors and on-site generation providers โ gas, small modular reactors, storage โ move from optional to mandatory. And the data center REITs that built their models on rapid suburban expansion, the ones with concentrated exposure to the jurisdictions now raising barriers, carry a regional discount that the market has not yet fully applied.
I learned something in 2022 that I have carried into every analysis since. When Terra and Luna collapsed, I retreated from public speaking for months. I sat with the wreckage and asked myself a question that felt almost unbearable: was the entire premise flawed, or just the execution? The answer, painfully, was both. Algorithmic stability had been sold as a law of nature. It was a bet โ a bet that confidence would hold, that the incentives would align, that the reflexive loop would not invert. When the confidence broke, the law broke with it.
AI infrastructure is running the same kind of bet. The industry has convinced itself that compute demand is a law of nature โ that if you build it, the users will come, and if the users come, the neighbors will accept. But that is not a law. It is a bet on consent, and consent is reflexive. Build enough resentment and the bet inverts. The 72 percent number is the market telling you the bet is starting to wobble.
Which brings me to the political escalation, the part of this story that worries me most.
When the CEO of AWS frames local opposition as a threat to American AI leadership, he is not describing a problem. He is executing a strategy. It is called issue escalation, and it works by relocating a fight from a venue where you are losing to a venue where you might win. The industry is losing at the town council level, where neighbors show up and speak, and it would very much like to move the battle to the federal level, where lobbyists show up and legislate. The national security frame is the vehicle. It converts a private interest into a public one, and it sets the stage for a preemption fight โ a bid to let federal or state authority override local zoning decisions.
I have spent enough time in policy rooms to recognize the tell. When an industry starts invoking national survival, it is usually because it has run out of local arguments. And this particular argument has a dangerous edge to it, because it is partially true. There is a genuine asymmetry between democratic systems, where communities can block infrastructure, and systems where land and power are allocated by administrative decree. If you believe AI leadership is a strategic asset, then the social license bottleneck is a real competitive handicap. That is what makes the framing so seductive, and so corrosive. A true premise can still be deployed as a weapon against the very people it claims to protect.
The trouble is that escalation invites backlash. If the public comes to believe that AI is being built by overriding their consent in the name of national security, the 72 percent opposition number will not fall. It will climb. You cannot win a legitimacy war by winning a jurisdiction war. When the graph spikes, the soul remains quiet โ and a graph of AI capability built on resentment is a graph with a hidden cliff in it.
Contrarian: the industry is treating a legitimacy problem as a PR problem, and that is the fatal error
Here is the counter-intuitive claim I want to leave you with. Everyone in this debate assumes the solution is better communication, better community relations, better checks written to the right local institutions. The billion-dollar commitment is the clearest expression of that assumption. Throw enough money and enough messaging at the problem, and the opposition softens.
I think that is precisely backwards, and I think the industry's own history proves it.
The reason communities are treating data centers as guilty until proven innocent is not that they are irrational. It is that they are operating under information asymmetry. Nobody will tell them how much water the facility will drink. Nobody will publish the actual power draw, the actual noise profile, the actual heat output. When institutions withhold the data that would let people reason about their own risk, people stop reasoning and start assuming the worst. The opacity is not a side issue. It is the engine of the entire conflict. The reason the revolt has spread from Inzai to Seoul to Sydney to Malaysia is not that data centers are uniquely evil. It is that they arrived without transparency, and transparency is the only currency that buys trust.
And here is the part the industry will hate: this is a problem decentralization understands better than centralization ever will. The entire premise of the systems I have spent my career building is that legitimacy comes from participation, not from authority. Quadratic funding worked โ when it worked โ because the people affected by a decision had a voice in it proportional to their conviction. That is the opposite of what a data center does. A data center is a decision made in a boardroom, imposed on a neighborhood, and announced after the fact. It is the most centralized thing you can build. And it is discovering, painfully, the oldest lesson in the book: you can centralize capital, but you cannot centralize consent.
Takeaway: the moat is consent
So where does this leave us, as we wait for a market that refuses to pick a direction?
The question was never whether AI gets built. It will get built. The question is whether it gets built with the people it touches or on top of them. The operators who understand that distinction first will hold the only durable advantage in this cycle โ not cheaper compute, not faster deployment, but the accumulated trust of the communities that host them. That trust is the new moat, and it cannot be bought, only earned, one transparent disclosure at a time.
The numbers surged. The room felt empty. And somewhere, a family is still waiting to find out what the facility next door will cost them โ in water, in noise, in sunlight. Until someone answers that question honestly, the revolt will keep finding new recruits.