The Only Guardrail Is on Your Electricity Bill

A debate about brakes, interrupted by someone who wants them off

The All-In Summit is a technology conference in Los Angeles where the room is mostly investors, founders and fund managers — people whose portfolios are long compute. On Monday afternoon, a panel about the pace of AI development was under way when Jensen Huang's phone rang.

It was the president of the United States. Huang took it and put him on speaker, according to attendees' videos and reporting by CNBC and Bloomberg. What the room heard was not a policy position but a ruling: talk of AI takeover was "a hoax," data centers are "great" because they make people and states wealthy, and the push to slow things down was "playing right into the hands of a lot of people that don't want to see it happen" — possibly China. At one point he told Huang, "I'm with you all the way."

Huang agreed on the call. That is worth stating plainly, because the framing of the moment has been about a president interrupting a debate. The more accurate description is a debate in which one participant picked up the phone and invited the conclusion in.

Six days that ended with a vote

The call did not happen in a vacuum, and the sequence matters more than any single quote in it.

• Sep 8A capability researcher — not a safety researcher — resigns from Anthropic, having previously worked on pretraining at OpenAI, and says in public that neither company is acting responsibly. The post reaches more than 90 million views in under a day. §04
• Sep 12Dario Amodei publishes We Must Pace the Frontier, arguing that addressing risk now requires slowing the rate of capability improvement, and proposes a three-step framework starting with embedded third-party evaluators. §04
• Sep 13–14The White House pushes back publicly, describing AI concerns as a conspiracy and calling those worried about it "Revolutionaries for a Bad and Evil Cause." §03
• Sep 14, 09:58 ETThe president posts the sentence this whole affair reduces to: the only control or "guardrails" AI needs is a "STRONG AND SMART (High IQ!) PRESIDENT." He names Amodei directly, calls him "Dario (Anthropic!)", and says the administration already holds "tremendous CRIMINAL and REGULATORY power over these companies." §03
• Sep 14, afternoonThe call to Huang, on speaker, in front of an investor audience. §01
• Sep 15, eveningThe House brings the bipartisan Ratepayer Protection Act to the floor under suspension of the rules. It is about electricity bills, not model weights. §09

"The only control or guardrails that AI needs is a strong and smart president"

The post is worth reading as a design document rather than a provocation, because it is unusually explicit about the governance model it proposes. Three claims do the work.

First, the guardrail is a person. Not an agency, a threshold, an audit right, a liability regime, or a reporting duty — a president. Whatever protective capacity exists is located in the judgment of one individual, and it is therefore revoked the moment that individual leaves, changes their mind, or is subject to the same commercial pressures as everyone else in the room.

Second, the existing powers are asserted as sufficient. "We already have tremendous CRIMINAL and REGULATORY power over these companies." This is the strongest claim in the post and the one most testable against the record, and we do not attempt that audit here: it requires a paper trail of enforcement actions that no single source provides.

Third, dissent is reclassified as disloyalty. The post's closing warning — aimed at "Conspiracy Theorists, Treasonists, Traitors, and Leakers" — moves the safety question from the technical register into the security register. Once that has happened, the cost of raising a concern inside a lab stops being a career risk and becomes something else.

"The only control or 'guardrails' that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades!"

@realDonaldTrump · Truth Social · Sep 14, 2026, 09:58 ET · archived by trumpstruth.org

The warning that mattered wasn't from the safety team

The detail worth noticing about the resignation on September 8 is who resigned. The loudest recent warnings about AI risk have come from people whose job title already implied concern; those can be discounted as a vested position. This one came from pretraining — someone who spent three years building the capability he then said he was afraid of.

His account of why colleagues stay is the sharpest thing in the coverage. It wasn't disagreement. It was fatalism: a broad internal agreement that the race is happening, paired with a belief that the best individual response is to keep your head down and make your own corner as safe as possible.

"There's this atmosphere of almost resignation, where people have accepted that the whole race is happening and as such, the best thing they can do is put their head down."

TIME, September 9, 2026

Anthropic's head of alignment stress testing publicly agreed with him and added a number: a personal estimate above 10% that AI kills all humans within the decade, alongside a statement that the company does not yet have a plan to solve alignment for superintelligence.

That is a company telling you, through two employees, that it is proceeding without a solution. Six days later the answer from the podium was that the guardrail is the president's IQ.

$96.2 billion in ninety-one days

Any argument about slowing down runs into a number reported on August 26. Nvidia's second quarter of fiscal 2027, ended July 26, produced $96.2 billion of revenue — up 106% year over year — of which $89.0 billion was data center. GAAP gross margin was 75.0%. GAAP net income was $59.7 billion. For the current quarter the company guided to $108 billion, and explicitly assumed no data center compute revenue from China at all.

Read the guidance again with the call in mind. A company forecasting $108 billion in a quarter has already priced the next quarter of the buildout. A moratorium, a pacing agreement, a genuine slowdown — these are not abstractions in that document. They are a line item with a sign in front of it, and the sign is negative.

Everybody is all-in, and nobody can find the power

The demand behind that guidance is the five largest US hyperscalers, whose 2026 capital expenditure plans have been reported in the range of $660–690 billion — against roughly $443 billion in 2025. That figure is a secondary aggregation of company guidance and should be treated as directional, not audited; we grade it C for that reason.

The more interesting part of the same reporting is physical. One major provider has been described as holding a large backlog of orders it cannot fulfil because it lacks the electricity to install the hardware it already owns, with power transformer lead times quoted in the range of two and a half years. This is the hinge of the whole story: the binding constraint on the buildout is not chips, capital, or regulation. It is megawatts, delivered to a specific place, on a specific date, past a specific set of neighbours.

The lab asking for caution is winning

The tidy version of this story — safety is a posture that losers adopt — does not survive contact with the market data. According to Menlo Ventures' 2025 enterprise survey, Anthropic's share of enterprise LLM API spend climbed to about 40%, from 24% the year before and 12% in 2023, while the former category leader fell to 27% from 50%. Total enterprise AI spend reached roughly $37 billion, up from $11.5 billion the year before.

Menlo is an investor in Anthropic, and that conflict is disclosed by the reporting itself; we carry it forward and grade the finding B. But if the direction is even roughly right, then the lab publishing the pacing essay is the lab gaining share. Caution, for them, is cheap. For the incumbent losing share, it is not cheap at all.

That is a less flattering conclusion than "nobody believes what they say," and a more useful one. The participants are not lying. They are each responding to a price. And no one in the argument — not the seller of compute, not the gaining lab, not the losing lab, not the president — is currently being charged for being wrong about the risk.

77% worry about their own electricity bill

The politics of AI safety are loud and the politics of AI infrastructure are quiet, and the second one has better numbers.

In polling reported in late August, 64% of respondents said building data centers at a rapid rate is not a good thing. 77% said they worry data centers will increase their electricity rates. Only 14% said they would be willing to live close to one. Those positions cut across party lines, and candidates in both parties have begun running against data centers — in at least one competitive gubernatorial race, both directions.

Notice what this coalition is not. It is not asking whether models can be aligned. It is asking who pays for the transmission line. That is a question with a legislative answer, and one that does not require anybody to adjudicate the nature of superintelligence.

One day after the call, the House moved a bill about wires

On the evening of September 15, the House brought the bipartisan Ratepayer Protection Act to the floor under suspension of the rules — the fast-track procedure reserved for measures expected to command two-thirds support. It was sponsored by one member from each party, and it has one operative idea: facilities above a certain size should pay for the generation and transmission they require, rather than passing those costs to everyone else on the system. The reported threshold is 100 megawatts.

The bill is weaker than that description sounds. It does not compel state regulators to adopt the standard. It requires them to hold a hearing on it within two years. Critics on both sides have said so out loud — that it is a nudge dressed as a rule. And its provenance is awkward for the week's narrative: the measure would codify parts of a ratepayer pledge the same White House released in March.

Set the two instruments side by side and the asymmetry is stark.

The row that moved is the row that is about money. The rows that are about models did not.

Three measurable tests, not three opinions

One: does the pacing framework acquire teeth? The framework's own first step — embedded third-party evaluators with access to training pipelines — is the only part that is verifiable. Watch for a named evaluator, a publication cadence, and an incident-reporting channel. Watch for silence.

Two: does a single state adopt the ratepayer standard? The federal bill asks states to hold a hearing. A hearing is not a tariff. The test is whether one state commission, under electoral pressure, actually sets a large-load tariff that a hyperscaler has to sign.

Three: does anybody resign over being wrong? The one thing no participant in this debate currently faces is a cost for a bad forecast. In a market where the seller of compute makes $59.7 billion in a quarter, predictions are free. Watch for the first place where someone posts a number and is held to it.

A reusable checklist

Where every number comes from