Is It AI Infrastructure Week Again?

Ryan Busse thinks Democrats have found the issue that brings working- class voters home.

Data centers.

He may be right. The electricity bills are real. The water is real. The tax subsidies are real. The transmission is real. And the people discovering that a hyperscaler would like to build a small industrial city next door are extremely real.

Busse is not vague about what he wants. He wants a two-year moratorium. Nothing built until strict regulations are in place. He says so plainly, and he criticizes former Senator Jon Tester by name for suggesting that a measured regulatory approach would be better. His reasoning is not lazy: these companies have unlimited money and move faster than any legislature, so a regulate-later posture is capture with extra steps. Debate is the delay they want.

That is a serious argument, and I want to take it seriously, because the response to it is not “calm down.”

Now look at Texas.

On August 3, Governor Greg Abbott directed the Public Utility Commission and ERCOT to conduct a comprehensive audit of every data center moving through the interconnection queue. No project advances until the audit is done. Anything that fails the requirements gets denied a grid connection. ERCOT postponed its Batch Zero transmission study. The Texas Tribune called it what it functionally is: a moratorium on approvals.

That is Greg Abbott. Not Greenpeace.

So we have a Democrat campaigning on a two-year moratorium and a Republican governor who just imposed something that works like one. The politics of this issue are not what anybody thinks they are.

But read why Abbott did it, because the reason is the most important sentence in this entire debate.

Texas asked. Twice.

The Public Utility Commission surveyed data centers on power and water usage. Of the 377 companies it notified, 28 responded, representing 92 facilities. A Republican member of the House Natural Resources Committee called the participation rate pretty pathetic. He was being generous.

The Water Development Board runs a separate survey, and that one is mandatory. It went to 341 data centers. Seventeen percent answered. Noncompliance is a Class C misdemeanor carrying a maximum fine of five hundred dollars, which tells you how seriously the obligation is meant to be taken and how seriously it is in fact taken.

That is the backdrop to the August audit. Abbott’s letter says it directly: the failure of some data centers to comply with the survey measuring water and power usage makes the directive necessary.

Sit with that. The largest data center market in the country asked its operators what they consume, once voluntarily and once under penalty of law, and more than nine out of ten companies did not answer. Then those same operators asked the state for grid capacity.

And what a request. ERCOT is looking at roughly 474 gigawatts of interconnection requests. That is more than five times Texas’s all-time peak demand, with data centers accounting for about ninety percent of the new power requests.

Nobody believes 474 gigawatts of real load is arriving in Texas. Not ERCOT, not the developers, not the utilities. The queue is full of optionality: speculative filings, duplicate applications at multiple sites,

Nine out of ten did not answer

projects hedging their bets, and some number of real facilities buried inside it. The queue is not demand. It is a wish list with a lawyer attached.

We know what happens when somebody makes the wish list expensive. Exelon began using transmission security agreements this year to separate real projects from speculative ones. Its high-probability data center load fell from roughly 18 gigawatts to about 11.

Seven gigawatts. Not from efficiency. Not from new technology. From asking.

Here is what strikes me about Busse and Abbott arriving at the same place from opposite directions.

Busse’s question is: should we let them build?

Abbott’s question is: what are they actually consuming?

Abbott’s is the better question, and he only got to it because his own state’s reporting requirement failed. But even Abbott’s audit stops short. It asks what a facility will consume. It does not ask whether the consumption is necessary.

That is the question nobody in this fight is asking.

Not whether AI requires power. It does. Not whether inference should be more efficient. The entire industry is racing on that. The question is whether the electricity being requested is doing useful computational work, or whether some portion of it is being consumed because the infrastructure underneath these systems keeps reconciling identity, policy, permissions, and authority every time an interaction crosses a boundary.

The question nobody in the fight is asking

I am not going to tell you what that fraction is. Neither can anyone else right now, which is the point. Texas just discovered it cannot get ninety percent of its operators to report a number as simple as gallons and megawatts. We are nowhere near an accounting of which watts do work.

But we know the direction, because the industry has published it.

In June, Stanford released SPIRAL, a paper on how frontier inference is actually going to work. The finding is that a single chain of reasoning is not enough for hard problems, so the answer is search: sample many reasoning traces in parallel, reason sequentially inside each, then aggregate across all of them. Up to eleven times the scaling efficiency when the compute primitives scale together. The engineering is serious, and the results are real.

Now notice what that could mean for the floor. In a research environment, those traces may run inside one system, one operator, one policy context, and the reconciliation between them never happens. But when the same scaling pattern is deployed across tools, memory systems, third-party APIs, agents, or separate authority domains, additional orchestration can also multiply the boundaries that have to be reconciled.

The industry’s roadmap is to increase orchestration. The coordination surface grows with the orchestration. On fragmented infrastructure, there is no reason to assume the coordination cost stays proportionate as traces, tools, agents, and authority domains multiply.

That is not a reason to stop building. It is a reason to ask the question before we pour the concrete, because infrastructure decisions made this year set the cost and carbon profile for the decade after.

You can build a solar farm to power avoidable overhead

One clarification before going further, because it is the difference between an argument and a slogan.

Coordination is not waste. Authentication, authorization, policy enforcement, provenance, and safety controls are necessary computational work, and any serious system spends real energy on them. Nobody should want less of that.

The issue is the avoidable part: the same interaction reconciled again at every surface it crosses, because the infrastructure underneath was assembled from pieces that do not share a boundary. Duplicative reconciliation is not security. It is the cost of not having designed the seam. That is what should be engineered out, and it is the only thing this argument asks anyone to engineer out.

With that distinction in hand, the environmental argument and the business argument stop being opposites.

You can build a solar farm to power avoidable overhead. You can build a nuclear reactor to power it. You can put the most efficient GPU ever manufactured inside a data center and still spend watts reconciling a fragmented system around it.

Clean electricity changes where the watt came from. Efficiency changes whether you needed the watt.

Those are not the same intervention, and almost everything in the current debate addresses the first one. More renewables, better cooling, water recycling, carbon offsets: all useful, all downstream of a demand figure that nobody has audited.

Before deciding how to produce another megawatt, it seems reasonable to establish whether the megawatt has to be produced.

That is not anti-growth. It is what any competent operator does. Nobody running a factory says the process wastes twenty percent of its input, so let us build a second factory. They fix the process, then they expand.

So here is the policy, and it is not a moratorium wearing a lab coat.

Before a project receives a major tax incentive, preferential utility treatment, public transmission investment, or scarce grid capacity, require four things.

Report what you already owe. Texas asked for power and water usage and got answers from fewer than one operator in ten. Compliance with the reporting a state already requires is not a new regulatory regime. It is a prerequisite.

Post something you lose if the project is not real. Non-refundable interconnection deposits, credit obligations, milestone requirements, shortfall payments. This is the part that already works. Exelon began using transmission security agreements to separate real projects from speculative ones, and its high-probability data center load fell from roughly 18 gigawatts to about 11. That is not free. It imposes a cost a serious project can justify and a placeholder cannot, which is the entire point. The price signal is what produces the revelation.

Bring your own power where you can. What portion of the load will the operator generate behind the meter rather than shift onto existing ratepayers.

And attest to the compute. Not an audit. An attestation: the operator states the methodology it uses, reports the categories that methodology distinguishes, discloses its resulting estimate of how much of the requested load is payload versus coordination, and is held to what it said the way any representation made to obtain a public benefit is.

Efficiency before expansion

That fourth one is where the objection lands, so let me meet it directly. No public utility commission can inspect a frontier inference pipeline. They regulate voltage and water rights, they do not have the staff to reverse-engineer an orchestration layer, and no hyperscaler is opening its architecture to a county board. All of that is true, and none of it is an argument against the condition, because the condition does not ask the state to compute the number. It asks the operator to state it.

That is how much of regulated disclosure works. The party holding the information makes the representation, the representation enters the record, and false representations have consequences. There is no generally accepted standardized metric today for separating useful computational yield from avoidable coordination overhead at the interaction level. Requiring operators to state a methodology is how such a metric comes to exist: competing methodologies become visible, comparable, and eventually standardizable, and the burden of building one sits with the only people who can.

Measure before subsidizing. Attribute before offsetting. Optimize before expanding. Then build.

Notice this answers Busse’s strongest objection. He is right that these companies outrun legislatures and that a three-year rulemaking is a gift to them. But efficiency before expansion does not necessarily require waiting for a comprehensive new regulatory regime. It is a condition attached to benefits the public already controls. Texas just demonstrated the point. Abbott did not wait for the 2027 Legislature before using the authority already available to require verification before projects advance through ERCOT’s process.

And it puts a very different question to an AI company. We are not telling you that you cannot have electricity. We are asking why our citizens should finance electricity you did not need to consume.

That is a hard sentence to argue with in front of a county commission.

There is one more objection worth answering, because it is the sophisticated one. Jevons paradox: make compute cheaper and aggregate demand expands rather than falls, so efficiency gains get eaten. Historically that is often true, and I will grant it entirely.

It is an argument about total consumption. It is not an argument about who pays. The ratepayer question is not whether national demand will fall. It is whether the public should underwrite load that the applicant cannot account for. Jevons does not answer that allocation question. If anything, it cuts the other way: if demand will expand fastest wherever the cost is externalized, the case for measuring before subsidizing gets stronger, not weaker.

Political problems where the solution gives almost everyone something are rare enough to notice.

The environmentalist gets fewer watts burned, not cleaner watts burned. The utility gets a load forecast that reflects something real. The community gets leverage. The taxpayer gets a test before the subsidy. The ratepayer gets protection from cost shifting. The technology company gets lower operating expense, which it wanted anyway. The investor gets more output per dollar of capital. The country gets more AI capacity inside the same energy envelope.

And the trades still get the construction project.

That last one matters more than it looks. In Montana, the Carpenters endorsed a MAGA Republican for Congress, the state AFL-CIO told the Democratic Party to stop criticizing data centers, and Busse read it as an inside job. I would read it more simply. When the only options on the table are build and don’t build, and one side says build, the people who build things will go where the work is. That is not betrayal. That is arithmetic.

Something for nearly everyone

Efficiency before expansion is the position that keeps the work and protects the bill. It is very hard to run a moratorium and a jobs argument at the same time. It is not hard to run this one.

Two roads here, and we have seen both before.

One party can spend a decade promising to repeal something without ever answering what replaces it. We watched that movie with health care, and the second act was always the same question: replace it with what?

Or everyone can hold another Infrastructure Week. Total agreement that the problem is real, total agreement that something must be done, and a running joke instead of a project, because identifying a problem has always been easier than building the answer.

The advantage on this issue will not belong permanently to either party. It will belong to whoever moves first from grievance to architecture.

Busse is right that Democrats should grab this. Abbott is right that nobody should get a grid connection they cannot account for. Neither one has yet asked whether the load itself is necessary, and that is the question that turns a protest into an industrial policy.

Yes to AI. Yes to the infrastructure. Yes to the jobs. Yes to protecting ratepayers.

With one condition.

Do not ask the public to build around overhead you could have engineered out.

I have written elsewhere about where that overhead lives and why the architecture produces it. The politics are simpler than the engineering.

The fork

Somebody is going to be the first to say yes, but, and that person is going to win an argument the other side cannot answer.