Maine's Data Center Moratorium and the Case for Smaller AI

Maine became the first U.S. state to pass a moratorium on large data centers last week, covering new facilities that consume more than 20 megawatts of power for the next eighteen months.¹ For context, 20 megawatts is roughly the scale of large hyperscale campus infrastructure, not typical regional deployments. The concerns are straightforward: electricity costs, water usage, and land use. Communities that hadn't expected to host the infrastructure powering the internet's next chapter pushed back, and their legislature listened.

That seems reasonable. It's also not obvious that this is a problem for AI.

Some will read it that way, as a statement that slowing data center expansion means slowing AI. That worry tends to assume that AI and data centers are basically the same thing. They're tangled together right now, but not necessarily because they have to be.

Hyperscalers are on track to spend something like $700 billion on data center infrastructure in 2026 alone.² That level of spending is hard to read as purely demand-driven. It looks more like a race to secure capacity ahead of consolidation, a hedge against being caught short if demand materializes as projected. Much of this hardware will be depreciated within a few years, even if it remains operational beyond that.³ Companies understand this and are building anyway, because the upside of owning that capacity is enormous if demand meets expectations. The result is a buildout shaped as much by financial positioning as by technical necessity.

Training frontier models genuinely requires enormous compute. Runs for models like GPT-4, Gemini, and Claude are widely understood to consume millions of GPU-hours.⁴ Inference, by contrast, is often far less demanding. While some large-scale products benefit from centralized infrastructure, many inference workloads run comfortably on a single machine or small cluster. Training is an episodic cost; inference is the day-to-day use, and for many tasks its requirements are far more modest than the current conversation suggests. That doesn't mean centralized infrastructure isn't essential — large-scale systems and consumer platforms depend on it. But it does suggest that not every application needs to be built at that scale.

If anything, restrictions like Maine's could be good for the field, not because they limit AI, but because they force more careful use of infrastructure.

In practice, constraints force better questions. When compute isn't effectively unlimited, you have to ask what the task actually requires instead of defaulting to the largest available model. That shift tends to produce more efficient systems, not worse ones.

There's also a sustainability case. Data centers are already among the fastest-growing consumers of electricity and water in the country, and that demand is accelerating.⁴ While they can bring real economic benefits to some regions, those benefits are unevenly distributed. Slowing the buildout can limit the extent to which those costs are externalized onto communities that didn't choose to bear them.

And then there's trust. The pushback in Maine wasn't about AI. It was about power bills, water tables, and land repurposed without anyone asking. AI infrastructure that can operate at smaller, more distributed scales doesn't impose those tradeoffs in the same way. That makes it easier to argue, genuinely, rather than rhetorically, that AI is working for people rather than extracting from them. A 2023 Pew Research study found that 52% of Americans feel more concerned than excited about AI, up significantly from previous years.⁵ Building systems that don't need to commandeer local infrastructure is one concrete way to address that anxiety.

That question of what AI actually needs ended up shaping how I approached a project of my own.

What I Built on a Pi

I've been working on a project called Civitas, a platform that scores U.S. senators and House representatives on how closely they represent their constituents. It pulls from voting records, campaign finance data, floor speeches, and stated platforms, and produces scores across four dimensions: funding independence, promise persistence, independent voting, and funding diversity. The scoring system is necessarily imperfect, more a structured lens than an objective measure.

It runs on a Raspberry Pi 5 with 16GB of RAM, sitting on my desk at home. It pulls data from government APIs like Congress.gov, the FEC, and GovInfo, but all AI inference happens locally. No cloud compute. No GPU. No calls to OpenAI or Anthropic or anyone else.

The main model is all-MiniLM-L6-v2, a compact sentence-transformer with a model size of about 90 megabytes.⁶ About 95 percent of the classification work runs through it: bill policy areas, donor industries, party alignment, stance direction. A batch of 64 documents takes around 100 milliseconds. It earns its keep.

The remaining five percent goes to Qwen 2.5 at 1.5 billion parameters, running via llama.cpp compiled for ARM. It handles the things that genuinely need generation: senator narrative summaries, promise evaluations, issue summaries for the action center. About 500 calls a day, three seconds each.

At cloud API rates, 180,000 LLM calls per year would run somewhere between $1,800 and $9,000 depending on the model and pricing tier.⁷ On the Pi, it's roughly $20 in electricity.

This isn't a model for replacing cloud infrastructure. It's a demonstration of how much can be done without it.

Doing More with Less

The 1.5B model is a tradeoff, but one the project's requirements made easy to accept. Civitas uses the LLM for structured extraction, completing templates rather than open-ended generation. Before each call, the pipeline retrieves the relevant context (key votes, campaign finance data, platform positions) and injects it directly into the prompt. The model doesn't need to store knowledge; it just needs to reason over what it's given. That's the core advantage of retrieval: a smaller model performing above its weight because the hard work of finding the right information has already been done. It raises the question: does a two-sentence summary of a voting record truly benefit from a trillion parameters, or is it simply a matter of feeding the right context to a smaller, more efficient engine?

The same thinking shaped the classification approach. It's tempting to pipe everything through a capable language model, but for something like categorizing a bill as "Labor" or "Healthcare," a sentence-transformer with cosine similarity is faster, cheaper, and more explainable. You can trace exactly why a classification happened. Send the same question to an LLM and you'll often arrive at a similar answer, but with significantly more compute and less transparency.

Maine's moratorium targets facilities over 20 megawatts. The Pi draws about 12 watts under full load.

The Models We Already Have

We often act as if the utility of AI depends on the next training run. But for many human-scale problems, the bottleneck isn't a lack of parameters. It's a lack of thoughtful application. The models powering Civitas are not new, and they didn't require retraining. They're already doing useful work.

Many of the most interesting applications likely haven't been built yet. A lot won't require larger models either, just engineers willing to reach for the right tool at the right scale and build accordingly. Those models already exist. What may not be necessary, at least for many applications, is another hundred gigawatts of new capacity to support them.

None of this is to say scale doesn't matter. There are real applications that need it, and the policy questions around energy and infrastructure deserve serious attention. It's possible this view underestimates how much future applications will rely on centralized infrastructure. But it's not obvious that today's trajectory is the only path forward.

The next time you reach for a large model API to power a feature, it's worth pausing to ask: is the limitation really model capability, or is it that the work of structuring the problem hasn't been done yet in a way a smaller model can actually use? Not because large models aren't valuable, but because they aren't always necessary.