Special edition: A pause is not a pause. It is a decision about who continues.

A bill introduced in Congress on September 3 would permanently ban the development of superintelligent AI, pause advanced AI development until a new federal regulator sets guardrails, create a cabinet-level AI agency, and attach penalties including corporate charter revocation and up to twenty years imprisonment.

I want to give that bill more credit than it will get from most of this industry. The people behind it are not confused about the technology and they are not grandstanding. The concern underneath it, that we are building systems whose behaviour we cannot fully predict at a speed that outpaces our ability to check them, is the correct concern. It is the one serious people inside the labs have been raising for a decade.

The bill is still the wrong instrument, and this piece is about why, and what I would rather we did instead.

The word pause implies a shared stop. Everyone lowers their tools, the world takes a breath, and work resumes when the rules are ready.

That is not what a unilateral pause is. A unilateral pause is a decision about who continues without us.

Our competitors have published their timelines. They are not waiting for a US regulator to write guardrails, and no American statute reaches their laboratories. So the practical effect of a domestic prohibition is not that the frontier stops moving. It is that the frontier keeps moving in jurisdictions with no congressional oversight, no Freedom of Information Act, no plaintiff's bar, no discovery, no state attorneys general, and no obligation to publish anything about what a system can do.

That is the trade, stated plainly. The safest version of a powerful technology is the one built somewhere it can be watched, sued, subpoenaed, regulated, and voted about. A ban forfeits precisely that and keeps none of the safety it was meant to buy.

I would put it more sharply. If you believe advanced AI is dangerous, you should want it developed where the courts work.

Set the geopolitics aside and the bill still has a drafting problem that would defeat it in practice.

It proposes to ban the development of a class of system that has no agreed technical definition, identified by a threshold nobody can currently measure, enforced by an agency that does not yet exist.

Laws written that way do not stop the largest developers. Those firms employ people whose entire job is to establish that whatever they are building sits on the permitted side of an ambiguous line, and they will be right often enough. What a law like that stops is the university group that cannot afford the opinion letter, and the startup whose investors will not accept the risk.

The result is the opposite of the intent: a frontier concentrated in the handful of firms large enough to litigate, with the independent researchers who audit and criticise them pushed out first. If your worry is unchecked concentration of capability, a vague prohibition delivers it.

SRJ publishes theworldofai.org, which rebuilds itself every day from primary sources. The numbers below come out of that pipeline, not from a think tank's projection, and they are what convince me this is an infrastructure argument rather than a philosophical one.

Capital is already committed at national-program scale. From the most recent SEC filings we track, Amazon reported 54.2 billion dollars of capital expenditure in the quarter ending June 30, 2026, and 44.2 billion the quarter before. Alphabet reported 35.7 billion, Microsoft 30.9 billion, Oracle 28.5 billion, Meta 19.0 billion. CoreWeave, which did not meaningfully exist as a public company three years ago, reported 7.7 billion.

Those are quarterly figures from a handful of firms. This is not a debate about whether to build an industry. The industry is being built, with money already spent, and the question left to policy is whether it gets built here.

The physical footprint is real and it is everywhere. Our registry holds 1,843 data centre facilities across 51 US states and territories. Six small modular reactor projects are tracked specifically against data centre load. This is a heavy industry now, with land, substations, water, and neighbours.

And it is being stopped locally while we argue nationally. We have logged 43 separate moratorium stories in the last ninety days and 229 data centre news stories in the last thirty. One of them this week is Philadelphia, where a moratorium is being urged amid what the coverage describes as spreading AI fear.

That is the actual constraint on American AI. Not a shortage of models. A shortage of somewhere to put them, and a permitting process that a determined county can stop cold.

Model development, semiconductor manufacturing, and data centre capacity with the power to run it are one supply chain. A model is a function of chips. Chips are a function of fabs. Fabs and data centres are both functions of electricity, water, and permission to build.

We currently regulate the first, subsidise the second, and let the third get decided parcel by parcel in county zoning hearings. Those three policies are not merely uncoordinated. They actively work against each other. We fund fabs while the interconnection queue for the power to run them stretches past the useful life of the equipment being installed. We debate model safety in Senate hearings while the physical capacity to run American models is blocked in Philadelphia and a dozen places like it.

No adversary has to outpace us if we can simply fail to build.

Not an industrial policy in the sense of picking winners. A plan in the sense of deciding, at national level, the things that cannot be decided locally.

Power first, because everything else is downstream of it. Interconnection queue reform with a statutory clock. Federal backstop authority for transmission serving critical industrial load, the way we treat other infrastructure of national consequence. A licensing path for advanced nuclear that matches the timeline of the demand it is meant to serve rather than the timeline of the 1970s.

Siting that is predictable. A large facility should face a defined process with a defined answer, including the answer no. What we have now is worse for everyone: developers cannot plan, communities cannot extract real commitments because there is no forum that binds, and the outcome turns on who shows up to a Tuesday meeting. Give communities genuine standing, real terms on water and ratepayer impact, and a decision that arrives inside a year.

Chips and packaging together. The fab investment is real and the advanced packaging and substrate capacity to finish those wafers largely is not. A leading-edge wafer that has to cross an ocean to be packaged has not been onshored in any sense that matters in a crisis.

Guardrails aimed at behaviour, not at existence. This is where I would spend the regulatory effort that the pause bill spends on prohibition. Pre-deployment evaluation with published results. Mandatory incident reporting with real consequences for staying quiet. Independent assessment by people who did not build the system. Liability that attaches to deploying something you did not test, rather than to the act of researching it.

That last cluster is not hypothetical, and here is the thing worth noticing. California enacted the scaffolding for it this month. AB 1405 requires that an AI auditor may not audit its own work, must keep evidence for ten years, and must disclose in every report what it did not examine. That is a serious rule, it constrains conduct rather than existence, and it was written by the same state legislature that is usually accused of overreach.

The choice is not between racing and stopping. It is between shaping a technology that is going to exist anyway and forfeiting the ability to shape it.

Every hour spent debating whether to permit AI development is an hour not spent on the interconnection queue, the packaging gap, and the siting process, which are the things that will actually determine whether the American AI industry exists in ten years.

Our adversaries have made their decision. They are building. The only real protection available to us is to keep building too, in the open, under law, with independent people checking the work and putting their names on it.

That is not a compromise between safety and speed. On this technology, at this moment, it is the same thing.

Stephen Jordan is the founder of SRJ Consulting & Services LLC, an accounting and AI advisory practice in Frisco, Texas. SRJ publishes theworldofai.org, which tracks AI legislation, litigation, data centre buildout, and grid interconnection from primary sources, rebuilt daily.

Figures in this edition are drawn from that pipeline as of September 14, 2026: capital expenditure from SEC filings, the facility registry, and our news archive. The underlying pages are at theworldofai.org/data-centers/ and theworldofai.org/ai-laws/.