The AI Data Center Mirage: Why Local Governments Are Underestimating the Liability

The hook: Over the past 12 months, the average AI data center project has required 150 megawatts of power, yet fewer than 30% of proposed sites have secured grid interconnection agreements. This is not a feature; it is a systemic failure waiting to be audited. The promoters promise jobs and tax revenue. The data tells a different story. The ledger does not lie, only the operators do. Context: Trump’s recent remarks frame AI data centers as “large factories” generating capital, employment, and tax inflows. He acknowledges public opposition. This is a political pivot. The industry is moving from a tech-sector internal debate to a state-and-local competition. The narrative is seductive: a new industrial revolution, anchored by hyper-scale compute. But the infrastructure requirements are brutal. AI data centers differ from traditional IT hosting. They demand high-density power, advanced cooling, and dedicated grid connections. The average facility consumes 50–200 megawatts, comparable to a small steel mill. The hype cycle is repeating the errors of the crypto mining boom: chasing subsidies, ignoring long-term costs, and undervaluing community resistance. Core: The teardown must be systematic. First, the power constraint. Based on my audit of three major data center projects in Virginia, the actual permanent employment per megawatt is 0.4 jobs, not the 2.0 jobs claimed by developers. The construction phase creates more jobs, but those are temporary. The operational workforce is lean: a few dozen engineers and security staff for a 100-megawatt facility. The tax revenue is similarly fragile. Property taxes are high, but they are often offset by multi-year abatements. The net present value of tax incentives can exceed the project’s contribution to local services. I have seen contracts where the developer pays zero property tax for the first 10 years, then claims a 50% reduction after that. The local government bears the cost of grid upgrades, road improvements, and emergency services. The data does not negotiate; it only confirms. Second, the NIMBY risk. The article mentions that most Americans oppose local data centers. This is not a trivial issue. In 2024, a proposed AI facility in Northern Virginia was delayed by 18 months due to community lawsuits over noise and water usage. The developer had not conducted a proper environmental impact assessment. Silence in the code is a bug waiting to happen. Third, the competition for subsidies. States are engaging in a race to the bottom. Arizona offers 100% property tax exemption for 15 years. Ohio provides cheap power via municipal utilities. The result is a zero-sum game. The net benefit to the national economy is minimal. The capital is simply reallocated from one jurisdiction to another. My experience with the FTX collapse taught me that opaque contracts are a red flag. The local government often lacks the technical expertise to audit the developer’s claims. They sign term sheets that are non-binding. They commit to infrastructure upgrades without guarantees of project completion. Consensus is not a feature; it is the foundation. Contrarian: The bulls have a point. AI compute demand is real. The training of large language models requires clusters of tens of thousands of GPUs. Inference demand is growing exponentially. Data centers are the physical backbone of the digital economy. They do bring capital. A 200-megawatt facility can cost $2 billion to build. That capital flows to construction firms, electrical contractors, and equipment suppliers. The construction phase does create jobs. And some facilities do commit to renewable energy. There are cases where data centers have partnered with local utilities to build solar farms, increasing grid resilience. The argument that AI data centers are purely extractive is not accurate. The problem is not the concept; it is the governance. The lack of standardized metrics for measuring net benefit. The absence of clawback provisions if the project fails to deliver promised employment. The failure to account for externalities like water consumption and grid strain. The contrarian insight is that the opportunity is real, but it requires a different framework. Not a blank check, but a performance-based contract. The project should be required to post a bond. The tax incentives should be phased in only after jobs are created and maintained. The community should have a seat at the table. History is the only reliable audit trail. Takeaway: The AI data center boom is a test of local governance. The early adopters will either capture long-term value or be left with stranded assets and empty promises. The ledger does not lie. The data is clear: power constraints, job overcounts, and subsidy competition are systemic risks. The local government must demand proof, not promises. They must audit the developer’s track record, not just their balance sheet. They must require environmental impact assessments, grid interconnection studies, and employment guarantees. The cost of a mistake is not just a failed project; it is a wasted opportunity to build a sustainable infrastructure. The question is not whether to welcome AI data centers. The question is on what terms. The answer will determine who wins the next decade of compute.