What the Current Situation in Maine Tells Neel Somani About the Future of AI Infrastructure

The Androscoggin Mill in Jay, Maine, has been quiet for years. The paper mill that once defined the town, employed its residents, and fueled its tax base closed. When a company proposed building a large data center on the site, some local officials saw it as exactly the kind of second life a former industrial community needs: new jobs, new tax revenue, and a new reason for the electricity grid to run at capacity.

The Maine legislature saw it differently. On April 9, 2026, both chambers voted to pass LD 307, a bill that would prohibit the construction or operation of any new data center drawing 20 megawatts or more of power until November 2027. If signed by Governor Janet Mills, it would make Maine the first state in the country to ban new data center development. The proposed facility in Jay, along with a second project in Sanford, sits in limbo.

The same week, OpenAI verified it was pausing Stargate UK, its high-profile partnership with the British government to build sovereign AI computing infrastructure inside the United Kingdom. The company cited high energy costs and regulatory obstacles. It said it would move forward when conditions improve.

A shuttered mill town in Maine and a paused billion-dollar AI infrastructure deal in Britain do not seem like they belong to the same story. But Neel Somani, a former quantitative researcher at Citadel who now works on machine learning and AI systems, reads them as symptoms of the same underlying problem: governments haven't fully dissected exactly who pays for what in the inevitable AI infrastructure build-out.

The Electricity Bill Nobody Accounted For

Jay is not an anomaly. It is another data point in a pattern that is now appearing across enough states and countries simultaneously to suggest something systemic.

Maine already carries some of the highest residential electricity rates in the country. Between 2021 and 2026, total electricity costs for Maine households rose roughly 60 percent. Distribution costs, the portion of an electricity bill tied to moving power from generation to homes, rose by an estimated 140 percent over that period. When data center proposals began surfacing in the state, the reaction from residents and lawmakers was not abstract concern about technology. It was specific and arithmetical.

"We're getting killed by electric prices," said Representative Amy Roeder during the legislative debate. "To put a data center that's going to use up a lot of resources in the middle of this just feels irresponsible."

American data centers consumed approximately 183 terawatt-hours of electricity in 2024, more than 4 percent of total US power consumption. That figure is expected to more than double by 2030 as AI workloads scale. A single hyperscale facility can draw as much electricity as a small city. The industry's standard response to concerns about grid impact has been to point to jobs, tax revenue, and pledges to build new power generation alongside new facilities. Maine lawmakers were not convinced. An amendment that would have created project-specific exemptions was rejected in the House 115 to 29.

What Maine did, in practical terms, was refuse to let the accounting remain invisible. The costs of AI infrastructure are real, and they have a real impact. The question the legislature was asking is whether there are implications for the residents of a state that have limited context into the decisions that created those costs.

OpenAI's Stargate UK pause is the same question at a different order of magnitude.

Stargate UK was announced in September 2025 as a flagship partnership between OpenAI and the British government, designed to build sovereign AI computing infrastructure inside the country. The stated rationale was straightforward: the UK wanted to run advanced AI models in data centers within its own jurisdiction, particularly for government and enterprise applications where data residency and national security considerations are paramount. OpenAI framed it as a template that it would extend to other countries through an initiative called OpenAI for Countries, with Australia, Greece, the UAE, Slovakia, Kazakhstan, and others in the pipeline.

The pause, attributed to energy costs and regulatory friction, reveals how quickly that template runs into the same wall that tripped up the proposed Jay data center. Industrial electricity in the United Kingdom costs substantially more than in the US markets, where the bulk of Stargate's original buildout has taken place. UK planning processes for large infrastructure projects are lengthy. Grid capacity is tighter. The economics that made rapid data center expansion viable in Texas and Northern Virginia do not apply to Britain and might not apply to other countries on OpenAI's list.

The AI infrastructure buildout depends on the availability of power at the required price and scale. It requires permitting processes to accommodate the pace of deployment and communities that welcome the investment. These assumptions, which were likely anticipated by the leading AI labs, are being tested in real-time.

Somani's Reading of the Situation

Somani's work on AI systems has given him a particular vantage point on infrastructure questions. His background in machine learning research makes him well aware of the compute and power requirements for training frontier LLMs. But his background in the power industry gives him insight into how difficult procuring that power will actually be.

His diagnosis is that the AI infrastructure build-out is an energy infrastructure problem. The capital required to build a data center is large but finite and well-understood. The energy required to run one at scale is an ongoing commitment that interacts with existing grid infrastructure in ways poorly understood by most people, making capital deployment decisions and regulatory guidelines. The scale of the build-out might require federal government support to help meet the rapidly growing demand.

The result is now playing out in Maine and Britain. Investment decisions made by a technology company or a sovereign government are concerning local residents who fear that their electricity bills will rise. Outdated planning processes designed for ordinary commercial development are being asked to accommodate facilities that draw the equivalent of a small city's power load. In both cases, the infrastructure assumptions embedded in the buildout are not properly understood by the local governments.

For Somani, the Maine decision and the Stargate UK pause are not obstacles to AI progress. They reveal where the buildout's political model is not yet complete, and they create pressure to develop a clearer accounting of what large-scale AI infrastructure actually requires and from whom.

What Comes Next

Maine's decision is not permanent. The legislation mandates a council report by February 2027, with the moratorium expiring the following November. The intent, as articulated by the bill's sponsor, Representative Melanie Sachs, is to create space for the state to develop a regulatory framework that answers the cost question before construction begins rather than after. The absence of such a framework is precisely what made the ban politically viable.

That framing matters more than the ban itself. A temporary moratorium pending the development of proper rules is not a rejection of data center development. It is a demand that the development happen on terms that make the costs legible and the distribution of those costs defensible. The distinction between those two things is what the legislators are requiring the AI labs to answer, since the legislators don't have the expertise required to evaluate the question themselves.

Bills to restrict or halt data center construction have now been introduced in at least a dozen states, including New York, South Carolina, and Oklahoma. Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez introduced federal legislation that would impose a nationwide moratorium until Congress establishes a comprehensive regulatory framework. The pattern is not specific to Maine. It is the result of a buildout that has outpaced the regulatory and community infrastructure needed to absorb it.

OpenAI's statement on Stargate UK said the company continues to explore the partnership and move forward when conditions allow. That framing suggests the company views the pause as a timing problem rather than a structural one. Whether that assessment is correct depends on whether the underlying energy and regulatory sentiment in the UK is likely to change, or whether the British market is fundamentally unsuitable to support AI data center loads.

The Infrastructure Gap That Explains Both Stories

The thread connecting Maine and the Stargate UK pause is an investment gap that has been building for years.

American data center construction has accelerated dramatically. Virginia has the largest data center cluster in the world. Texas is expected to surpass it. Data center power consumption is on track to more than double by 2030. The investment in the power generation and transmission infrastructure needed to support that consumption has not kept pace. The gap between those two curves is now surfacing as grid strain, rising ratepayer costs, legislative backlash, and stalled projects in markets where energy costs and regulatory conditions do not align with the assumptions embedded in the buildout model.

Somani's read is that the organizations best positioned for the next phase of AI infrastructure development are not the ones that view energy as a secondary constraint to be managed around. They are the ones who treat it as a primary variable shaping every aspect of the investment decision: where to build, how to structure the relationship with grid operators, what commitments to make to local communities, and the long-term economics of running facilities at scale.