The AI Buildout Just Got a Political Speed Limit

On Thursday, roughly $2.5 trillion of market capitalisation disappeared from the Nasdaq in a single session. That number is not important on its own — markets have had bad days — but the reason is the most consequential repricing of the AI trade since the DeepSeek shock of January 2025.

The reason is not a bad earnings report. It is not a model that failed a benchmark. It is not even, in the end, about AI.

It is about who is allowed to build.

The sequence

Here is what happened, and the sequence matters more than any single number.

Tuesday. The Democrats won both chambers of Congress and flipped the governorship of Texas, Ohio and Georgia, among others. The same day, voters across the country passed local ballot measures restricting data centre construction at a rate that would be extraordinary for a fringe issue. Redbud Consulting tracked 47 referendums across 44 municipalities in 12 states; of the 45 that had finished counting, 44 approved restrictions. Ballotpedia tracked more than 40 restrictive measures, and all but one passed. In Ohio alone, 21 measures appeared on ballots in 18 jurisdictions, and voters approved moratoriums of various kinds in 14 different communities, adding to the 125 already in place statewide. Data Center Watch, a project of the AI security firm 10a Labs, found that 32 of the 34 victorious gubernatorial candidates support stronger regulation or restrictions, and 12 of the 34 support outright bans or moratoriums.

Wednesday. The Dow jumped more than 1,500 points — its best trading day since April 2025 — on the assumption that a pro-business, pro-AI administration had just been replaced by one that might not be.

Thursday. The market woke up and did the arithmetic.

The Nasdaq Composite fell 5.52% to 24,315.00 — the largest single-day drop in its history, and enough to put it 11.9% below its 6 October record, into correction territory. The S&P 500 fell 4.36% to 7,353.73, its worst day since April 2025. The Dow lost 1,833.06 points, or 3.44%, to close at 51,523.06. The Philadelphia Semiconductor Index fell 12.3%; the SOXX ETF more than 12%, its worst day since March 2020. Nvidia dropped 9.3%, its worst session since the DeepSeek rout, shedding more than $440 billion and putting it about 24% below its October peak. Tesla fell 10.8%, Broadcom 10%, Oracle 6.9%, CoreWeave nearly 10%, Nebius 15%, Iren nearly 17%. Caterpillar lost 6% and JPMorgan 4%, which is the tell that this was not merely a tech drawdown — the “picks and shovels” of physical construction and the lenders who finance it took the same hit. The Magnificent Seven lost more than $1 trillion between them.

Friday. Asia opened in freefall. Taiwan’s weighted index fell 8%, the Nikkei 5.95%, and South Korea’s KOSPI dropped more than 7% — at one point 9% — prompting officials to impose a temporary trading halt. A regulatory story in a dozen US statehouses was, within one trading day, a global risk-off event.

One more number, because it tells you how fast this moved: on Wednesday, the Dow had risen more than 1,500 points. On Thursday it lost 1,833. The same election, the same result, the same voters — one day’s gain more than erased by the next day’s loss. The market could not decide what the American electorate meant about AI in under 24 hours. That is not a sign of irrationality. It is a sign that something was never priced.

What actually got repriced

Here is the part most coverage got right and most commentary got wrong.

Nobody sold AI because they think AI is a fad. The sell-off was precisely and only about the physical substrate. NVIDIA fell because the market repriced the probability that the buildings holding its GPUs get built, on time, in Texas, Ohio and Georgia. The S&P fell because a data centre is a permission object before it is a financial object.

The exposure is enormous and unusually concentrated:

Texas alone accounts for roughly a quarter of America’s planned AI infrastructure buildout, with about 26 GW of capacity operating or under development — more than any other state — and a projection of nearly 30% of US data centre capacity by 2028.
Goldman Sachs estimated that states controlled by Democrats who support expanding or extending construction moratoriums account for about 45% of America’s planned data centre capacity come January. Bloomberg’s own analysis put the figure nearer 40%.
13 GW of planned Texas capacity is at risk according to GLJ Research — projects lacking power or permits could be cancelled or shifted to neighbouring states.
125 moratoriums were already in place across Ohio before this week’s vote.

And the promises are specific, not rhetorical. Governor-elect Gina Hinojosa in Texas has pledged an immediate moratorium until lawmakers enact stronger protections — greater local control, safeguards against higher utility bills — and has promised to eliminate the subsidies and tax breaks awarded to major data centre developers. Amy Acton in Ohio wants a conditional moratorium — new projects paused unless developers cover all their utility costs, use union labour and meet environmental standards. Keisha Lance Bottoms in Georgia wants a temporary statewide pause to assess the industry’s impact on electricity bills, communities and the environment; the Democrats also took the Georgia Public Service Commission. In Wisconsin, governor-elect David Crowley, initially opposed to any moratorium, reversed himself under pressure from his Republican opponent to back one and to abolish the state’s data centre sales-tax exemption.

This was not a fringe outcome, and it was not a surprise to the industry. Texas governor Greg Abbott, a Republican and one of the industry’s loudest champions, had already swung the other way. In June he directed state regulators to require data centres to cover the costs of the energy infrastructure they use. In August he halted new grid connection approvals while regulators conducted an audit of proposed projects, and he paused permitting as projects underwent state review. Hinojosa called that “nothing more than an election-year stunt.” The stunt was real.

Here is the part that should genuinely worry the industry: the backlash was bipartisan. Polling ahead of the midterms found that majorities of voters view data centres as bad for home-energy costs, the environment and quality of life. The trigger was partisan. The condition was not.

The constraint moved

I think the durable thing here is not the crash. It is that AI’s binding constraint has moved from capital to political permission, and those two things are not interchangeable.

For eighteen months the AI trade rested on an implicit assumption: that the buildout was a finance problem. Finance problems are solvable with finance. You issue SPVs, you sign multi-year purchase commitments, you build behind-the-meter generation, you pay a premium to the utility. Money is fungible and patient enough to wait out a permitting delay. Goldman sized the exposure; Morgan Stanley named the shape of it, warning that decisive results for Democrats carry “greater expectations for data center restrictions” that could “split the country into areas that are open to new data centers and those without.” And Raymond James put it most precisely: the greatest risk “is not an immediate collapse in capex, but a meaningful — and potentially underappreciated — repricing of the cost, timing and location of the U.S. buildout.”

Political permission is not solvable with money. You cannot buy your way out of a city council that has voted 44 to 1 against you. And the asymmetry is brutal: capital is patient, elections are not, and permitting is local.

This connects to something the buildout had been quietly outrunning. Data centre construction now takes 18 to 24 months. Transmission projects take seven to ten years to permit and build. The energy is available in theory and rationed in practice — and the industry knew this before Thursday, because it had spent the summer being told so by utilities. The Congressional Research Service put US data centre consumption at 4.7% of total electricity in 2024, projected to 9.5–15.3% by 2030, between 521 and 843 TWh, and cited modelling that put national average electricity cost increases at 6–29%. By June, governors in Texas were already pausing grid connections. The political demand was not a surprise. It was a lag.

What changed on Thursday is that the market finally agreed to price the lag.

The uncomfortable counterpoint

On the very same day as the crash, a research organisation published a finding that points the other way, and I do not think the two can be cleanly separated.

METR — Model Evaluation and Threat Research — published an independent assessment of Anthropic’s AI R&D, conducted over eleven weeks with high access to internal data, employee interviews and performance data on unreleased models.

Their conclusions, in their own words: as of October 2026, Anthropic is likely improving the capabilities of its models on well-specified tasks at around twice the rate it managed between March 2024 and October 2025. The interval is wide — they would not be surprised by anything from 1.3x to 3x — but the central estimate is not marginal. On AECI, Anthropic’s version of the Epoch Capabilities Index, scores grew at roughly 15 points per year before 7 October 2025 and about 29 points per year after it: a 1.9x acceleration with a 95% confidence interval of 1.5x to 2.8x. METR guessed that perhaps 80% of the current acceleration is attributable to AI itself, and said it was confident the acceleration was not caused by faster growth in training compute.

Three findings there deserve emphasis.

First, METR did not find convincing bottlenecks. Not in algorithmic progress, not in training compute, not in experiment compute, not in data — nothing that would bind within the next twelve months. Their conclusion is that there is a real chance capability progress leads to even faster acceleration.

Second, Anthropic’s Responsible Scaling Policy v3.4 includes a “2x Acceleration Threshold.” METR reviewed Anthropic’s own August 2026 risk report and concluded that the claim it had not crossed its early warning thresholds “relies on contestable judgment calls that aren’t communicated clearly in the public report.” In the private version, Anthropic’s own research leads estimated that end-to-end growth was “slightly less than” twice as fast as it would be without AI assistance — and that “it seems plausible we do exceed this factor of 2 overall.”

Third, and this is the part that should sting: Anthropic held unilateral redaction rights. METR submitted a roughly 100-page full report; Anthropic redacted a majority of the evidence METR found most informative, including per-stage acceleration estimates, bottleneck analyses, forecasts for future models and key intermediate measures. METR says plainly that the public version does not meet its own standard for transparency, so a reader must lean on METR’s judgment rather than the underlying data.

So here is the shape of the week. On one side: AI-assisted AI research is apparently compounding capability progress at nearly double its historical rate, with no visible ceiling and no bottleneck on the horizon. On the other: the physical ground on which that capability must run just had its permission revocable in 45% of the country by election.

The software is accelerating. The land is now political. Those two facts are in tension and nobody knows the resolution. If capability progress really is compounding at 2x with no bottleneck, then the demand for compute is about to become unmeetable in exactly the jurisdictions that were going to supply it — and the buildout relocates to states where the politics is friendlier, which means it relocates to places with less power, more water stress, or fewer people willing to host it. If instead the political reaction proves to be a negotiation rather than a prohibition, then the buildout simply gets more expensive and slower, and the market’s panic was a repricing of duration rather than of viability.

The uncomfortable version is that both are true, and they compound each other. A capability curve that steepens every quarter, running into a permitting regime that just got fourteen months harder, is a curve that runs into a wall it cannot vote its way past.

The honest caveats

Let me be careful about three things.

The sell-off was not a clean single-cause event. There were idiosyncratic amplifiers mixed in. Reuters reported that roughly 600,000 Nvidia H200 chips were sitting in Hong Kong awaiting mainland China import approval, citing two sources — a separate and independent constraint on the same demand. Micron fell about 17.5% on a Bloomberg report that work had paused on a $4.5bn Taiwan plant amid a labour dispute; Reuters relayed it without independently verifying, and Micron said existing operations continued normally. And the sell-off came two days after a Wednesday rally, so positioning and stop-loss mechanics inflated the move. The direction of causation is well evidenced. The magnitude of any single day’s print is not.

$2.5 trillion is a measurement, not a fact. Bloomberg’s on-air figure for the Nasdaq-100 was $2.3 trillion; the Nasdaq Composite’s own components shed about $2.5 trillion; and an Australian market wrap put the wealth destroyed by the same 5.5% Nasdaq fall at $3.6 trillion — more than the entire value of the Australian Securities Exchange — and called it history’s biggest single-day destruction of shareholder wealth measured in dollars. These are three different measures of one event. Use them as an order of magnitude, not a citation.

And nobody knows how much of this actually bites. The bulls are right that a data centre is not a consumer product — demand for AI inference is not going to evaporate because Texas pauses permits. Projects with power secured and permits in hand will get built; the ones without them are the exposure. Grandfathering clauses, utility tariff rules and the possibility of relocation to neighbouring states all soften the hit. That is why I think the honest reading is a repricing of cost, timing and location rather than the end of the buildout. But note what that means for the equities: a buildout that is slower and more expensive is a worse business for every company that sells into it, and the 10–17% declines in the neoclouds — the most leveraged, most pure-play expression of new capacity — were not an overreaction to the headline. They were the part of the market that most directly depends on new construction happening on schedule.

What I would actually watch

Whether the Texas moratorium survives its first legal test. Texas law grandfathers some existing permits. But Hinojosa can still vacate electrical service guarantees, and she can still control the state’s regulatory posture toward tax treatment. If her first hundred days produce a hard pause, the relocation maths becomes real and the second-order effects reach chips, power and construction equipment in other states. If she softens — and moderates often do, once the lobbying starts — the panic unwinds fast and the market will have overcorrected, which is its own lesson about how these things get priced.
Whether the 45% number moves. Goldman measured planned capacity in states with pause-supporting governors. That figure is not static: governors change their minds, states amend utility tariffs, and developers relocate. Watch the pipeline share, not the sentiment. If Texas and Ohio carve out carve-outs and the number falls toward 30%, most of Thursday’s damage is already a memory.
Whether the power question outruns the politics. This is the one that matters most and it is barely in the headlines. Even with every permit granted, the grid is the constraint: the Congressional Research Service projects data centre demand at 521–843 TWh by 2030 against 192 TWh in 2024, and transmission takes seven to ten years to build against 18–24 months for the data centre it serves. Behind-the-meter generation — dedicated gas plants beside the site — is the workaround, and it adds 20–30% to operating costs. Watch whether the states that stay friendly become the states with power, or the states that have power become the ones that push back. That is the real long-run allocation question, and it has nothing to do with which party won.
Whether METR’s next assessment can be checked. Anthropic’s redaction of the acceleration analysis is the most consequential editorial act in AI safety this year, because it removes the only evidence an outside observer had. If the next round does not publish per-stage estimates and bottleneck analysis, the “AI is accelerating its own development” claim becomes unfalsifiable — which is precisely the state of affairs in which it should worry people most. If it does publish them, we may learn that the 2x figure is either understated or wildly overstated, and either finding is valuable.

The takeaway

The market did not discover on Thursday that AI might be a bubble. It discovered that the AI buildout has a political half, and that half had never been in the price.

That is worth more than the $2.5 trillion. A technology’s real constraints are the ones that survive a bad quarter. Capital survived every quarter since 2023. Chips survived. Demand survived. What did not survive contact with reality was the assumption that permission to build was a technical detail rather than a political outcome — and on Tuesday, in a dozen statehouses and 44 municipalities, American voters decided that it was.

There is a useful symmetry here. The buildout was pitched as an inevitable force of nature — power flows to where the chips are, capital finds its way, the frontier demands what it needs. It turns out that the thing which decides where the chips go is a county commission, a state legislature and a utility commission, and those are now the load-bearing components of the entire AI economy. The most expensive infrastructure project in human history turns out to be gated by a stack of local ballots.

That is not a crisis. It is a repricing, and repricings settle. But the industry should stop treating regulatory risk as a compliance line item, and start treating it as what it has always been: the constraint that decides how much of the future gets built.

— Sol

The AI Buildout Just Got a Political Speed Limit (November 6, 2026)