The slowdown that isn't coming

The estimates

What I would actually bet on

Market prices where markets exist; my synthesis where they don't. Every number below is an estimate, and the reasoning behind each one is in the section that follows.

Washington

Congress passes one kind of AI bill, and only one

Since 2023, exactly one significant AI bill has become federal law: the TAKE IT DOWN Act, criminalizing non‑consensual intimate imagery, which passed the House 409‑2 and the Senate by unanimous consent before being signed in May 2025. The Brennan Center counted 150+ AI bills introduced in the 118th Congress and zero enacted.

The successes share a signature: an identifiable victim, no industry opposition, no regulatory burden, and a lopsided vote. Everything that touches how models are built has stalled or been stripped. That is the whole pattern, and it has not changed in three years.

The coalition does not follow party lines

This is the single most misread feature of AI politics. The bloc most dangerous to industry is a cross‑party populist pairing, not a Democratic one. Hawley and Blumenthal have jointly produced the GUARD Act, the GRID Act capping data‑center‑driven electricity increases, an AI risk‑evaluation bill, and a bill letting creators sue AI companies over training data. Their chatbot bill passed Senate Judiciary unanimously.

And the Republican median moved away from unconditional deregulation during 2025. Blackburn led the strip of the state‑law moratorium, then in December 2025 unveiled her own conditional‑preemption framework. The emerging Republican position is not "no regulation" and not "no preemption" — it is preemption in exchange for kids, creators, and conservatives protections.

The preemption war

Congress said no twice. The executive branch found a way around it.

This is the most underrated development in the corpus. Federal preemption of state AI law failed in the two most plausible legislative vehicles. Then the administration stopped trying to legislate and started litigating — and it worked, on a six‑month clock.

1. 1 July 2025

   The Senate strips the ten-year state AI moratorium from the reconciliation bill 99‑1. Only Tillis votes no. Forty state attorneys general and legislators from all fifty states opposed it.

2. 7 December 2025

   The NDAA preemption rider is dropped after the House Armed Services chair, the Senate Armed Services ranking member, and 290+ state lawmakers object.

3. 11 December 2025

   Executive Order 14365. Creates a DOJ AI Litigation Task Force, directs Commerce to evaluate "onerous" state AI laws, conditions BEAD broadband funding, and opens an FCC preemptive rulemaking. Carves out child safety, data-center infrastructure, and state procurement.

4. 9 January 2026

   The AI Litigation Task Force is stood up, on schedule.

5. 9 April 2026

   xAI sues Colorado over its AI Act — First Amendment, Dormant Commerce Clause, due process, equal protection.

6. 24 April 2026

   DOJ intervenes — the first federal action against a state AI law.

7. 27 April 2026

   A federal magistrate stays enforcement of the Colorado AI Act, before it ever took effect.

8. 14 May 2026

   Polis signs SB 26‑189, repealing and replacing Colorado's own law with a thin disclosure regime. California, Texas, and Illinois have been named as next targets.

Three caveats keep this from being decisive. The Colorado outcome was a stay and a legislative capitulation, not a merits ruling — it does not automatically generalize. Law firms across the board judge the underlying Dormant Commerce Clause theory weak absent a federal statute or valid agency rule to conflict with, and no FCC rule has issued.

Most importantly: this is one result, against the softest possible target. Colorado's own governor had signed the law "with reservations," publicly asked the legislature to weaken it, and wanted federal preemption himself. Roughly 150 lobbyists had already forced a delay in a 2025 special session before DOJ was involved. The named next targets — California, Texas, Illinois — remain unlitigated, and California's SB 53 and New York's RAISE Act still stand. One win over a defendant that was already surrendering is not yet a repeatable playbook.

The counter‑signal is that states did not stop. Over 100 new state AI laws were enacted by 1 July 2026, against 121 by the same point in 2025 — a modest slowdown, concentrated in exactly the categories the order exempted. States are routing around the threat, not obeying it.

The states

The comprehensive model is 0-for-6

Every serious attempt at an EU‑style duty of care against algorithmic discrimination has failed. Not stalled — failed.

Zero states have a comprehensive algorithmic‑discrimination AI law in force as of August 2026.

What survives instead

A much thinner substitute, and it is remarkably consistent across red and blue states alike: disclosure, narrow intent‑based prohibitions, and a regulatory sandbox.

Texas's TRAIGA is the template. The original 43‑page draft was modeled on the EU AI Act; the rewrite eliminated nearly all of it. The load‑bearing change is one sentence — disparate impact alone is not sufficient to show intent — which makes the law close to unenforceable against algorithmic bias. Private developers have almost no affirmative duties. Government agencies must disclose. There is a 36‑month sandbox.

Connecticut proves the pattern from the other direction. Maroney's bill failed twice; SB 5 finally passed in 2026 — 32‑4 in the Senate, 131‑17 in the House — only after adding a regulatory sandbox Lamont specifically requested and narrowing employment scope. Utah got there first, in 2024, and its SB 149 Office of AI Policy remains the model everyone else converged on.

Where the risk actually lives

Data centers are more unpopular than fracking

And it is genuinely bipartisan. Net favorability runs −75 among Democrats, −65 independents, −43 Republicans, −63 among rural voters. Support for a construction pause splits Democrats 67%, independents 66%, Republicans 58% — a strikingly narrow gap. Data Center Watch found opposing local politicians split 55% Republican, 45% Democratic.

Pew's most useful finding kills the industry's favorite defense. Among Americans who have heard a lot about data centers, 67% say they are bad for home energy costs — against 42% among those who have heard a little. Familiarity makes opinion worse, not better.

Why this matters more than AI opinion does

Because AI itself has no salience. In Pew's July 2026 open‑ended poll of what candidates should talk about, the economy and cost of living took 29%, government ethics 9%, immigration 7%. AI did not appear at all — it fell below the 4% reporting cutoff. Eighty‑six percent of voters want stricter AI regulation and it still doesn't crack the list.

AI regulation risk does not arrive labeled as AI regulation. It arrives as a utility bill.

The scale check

Everything above is real. Almost none of it is big.

Data Center Watch, the only national tracker, reports $130B blocked or delayed in Q1 2026 across 75 projects, with opposition groups active in 49 states and 150+ local bans passed in July 2026 alone. Those are period snapshots, not a cumulative total.

Set them against the pipeline. Morgan Stanley puts global data center spend at roughly $3 trillion for 2025–2028; US megacap capex alone runs about $1.1 trillion for 2026–2029. Even the largest single reporting window is around five percent of that.

And blocked does not mean cancelled. In Data Center Watch's own first breakdown, $18B of $64B was blocked outright and $46B — 72% — was merely delayed. Tract's $14B Arizona project was withdrawn under zoning resistance in May 2024 and rebuilt in the same county at a friendlier site by August.

Virginia is the ceiling test, and it failed

Virginia gave the opposition everything it could ask for: a 15‑point governor's win, a 64‑36 House supermajority, and JLARC's finding that data centers will push a typical residential bill up $14–37 a month by 2040 while the sales‑tax exemption cost $928M in FY23.

What that mandate produced: a new 1.1¢/kWh consumption tax capped at $600M a year and sunsetting in June 2028 — and the tax exemption survived untouched. The sharpest ratepayer win of the year came four months after the session closed, from the SCC ordering Dominion to shift transmission costs onto data centers, which cut the residential surcharge from a proposed $2.90 a month to $0.94.

Meanwhile Northern Virginia's under‑construction capacity hit 2,078 MW in H1 2025, up 80% year over year, and the growth is relocating within the state — Google broke ground in Botetourt, Stack is moving in Pittsylvania, Wythe County landed a campus.

The 1.6% number

One statistic should discipline every "hundreds of gigawatts" claim on either side. ERCOT's large‑load interconnection queue holds roughly 233–239 GW of requests, up about 300% year over year, 70‑plus percent of it data centers. Only ~9 GW has approval to energize, and observed peak large‑load consumption is ~3.9 GW. A request‑to‑operating conversion of about 1.6%.

Nationally the same picture: 3,969 US data centers are planned and only 802 — about 20% — are under construction. Of a 565 GW announced pipeline, roughly 180 GW is judged realistically buildable. An ERCOT official put it plainly: the review process was built for forty or fifty large loads at a time, not hundreds. Queues are wish lists.

The actual constraint

Gas turbines, not statutes

If every moratorium were repealed tomorrow and Congress passed nothing forever, the buildout would still be capped — by a supply chain that cannot physically deliver.

The bottleneck is castings, welders, and factory slots — not fuel and not permitting. Plant construction lead times have gone from 3.5 years in 2023 to about 5 now, and up to 7 for some models.

Which is why the demand-side numbers keep getting revised up while the buildout does not accelerate to match. BloombergNEF's July 2026 forecast — data centers reaching about a fifth of total US electricity by 2035, around 200 GW — was 83% higher than BNEF's own December 2025 forecast. Every hyperscaler raised 2026 capex; none cut. Big-five capex is running about $507B against roughly $317B in FY25.

But these constraints compound — they don't queue politely

The tempting version of this argument is that turbines bind first, so politics is harmless slack behind a hard ceiling. That reasoning is wrong, and it is the weakest link in the case above.

The grid has no slack to absorb political friction. PJM has missed its reliability requirement two years running. The Department of Energy has issued 40+ emergency orders since May 2025 — more than any comparable twenty-year period — purely to keep roughly 4.4 GW of coal plants from retiring, and each one lasts only ninety days before it must be re-justified. Grid-connection waits already exceed four years in primary markets.

In a system that tight, hundreds of local bans, two statewide moratoria, and a wave of new large-load tariffs do not sit harmlessly behind the turbine ceiling. They consume the same scarce interconnection, permitting, and regulatory-attention capacity the physical shortage is already straining. The honest formulation is that turbines set the ceiling and politics determines how much of the space beneath it actually gets used — not that one makes the other irrelevant.

Where this could break

Four ways I could be wrong

One of these is not really a tail risk. It is a live channel that my own framing above understates, and I want to be explicit about why.

The liability channel, and the circularity problem

Everything in this report defines "does policy slow AI?" as "does anything cap compute or gate deployment?" That framing has a built-in flaw: it excludes liability, and by excluding liability it guarantees a low answer. Nobody is proposing a FLOP cap, so a test that only counts FLOP caps will always come back near zero. The underlying research calls Section 230 erosion the single biggest legal-exposure shift in the entire landscape — and my headline framework scores it as noise. Here is the case for taking it seriously.

In Garcia v. Character.AI, a court held in May 2025 that chatbot output is first‑party product speech, not Section 230‑immune third‑party content. The case settled confidentially in January 2026 along with four related suits, so the holding stands without appellate review. Once model output is a "product," the entire products‑liability apparatus becomes available against developers — no legislation required. Roughly eleven wrongful‑death suits including Raine v. OpenAI are now coordinated in one San Francisco proceeding.

Anthropic's $1.5B Bartz settlement, finally approved July 2026, established the parallel doctrine on the copyright side: training on lawfully acquired data is fair use; liability attaches to how the data was sourced. That converts an existential question into a licensing line item — but the Third Circuit heard argument in Thomson Reuters v. Ross on 11 June 2026, and a decision going the other way would reset the doctrine.

The other three

Watchlist

Ten dated things that would move these numbers

Two things worth knowing separately

Money, and the rest of the world

The money is enormous and has not obviously worked yet

Leading the Future — a16z at $50M, Greg and Anna Brockman at $25M, Joe Lonsdale, Perplexity — has raised $125–140M and spent $24M+ across federal primaries, backing 28 candidates of whom 25 won. It endorses in both parties. Public First Action, funded by Anthropic at $40M, has raised $80M on the other side.

Their marquee fight was New York's 12th district, where Alex Bores — author of the RAISE Act — was targeted with $8M in attack ads. He lost the June primary 38.9% to 34.9%. But precinct analysis credits the winner's institutional endorsements, and the anti-industry side actually outspent the industry side roughly 2.5 to 1 in that race. The "AI money bought a seat" reading is not supported by the spending data.

Against which: the industry lost the only federal vote that mattered, 99‑1, with unified funding and direct presidential backing. And the clean framing is dead — both leading labs now fund advocacy on the pro-regulation side of specific fights, and OpenAI funds a "Parents & Kids Safe AI Coalition."

Europe delayed its own rules and has issued zero fines

The EU AI Act's high‑risk obligations were pushed from August 2026 to December 2027 for most Annex III systems and August 2028 for Annex I — a sixteen‑month slip agreed in trilogue on 7 May 2026, driven partly by the fact that the CEN‑CENELEC harmonized standards were not going to be ready. The compliance clock had started before the yardstick existed.

Enforcement began 2 August 2026. There have been no fines. The AI Office has "more than 125 staff" and was still recruiting forty more enforcement roles with applications due in September. And the documented cases of labs withholding products from Europe — Meta's multimodal Llama, Apple's phased feature rollout — were driven by GDPR and the DMA, not the AI Act.

The distinction that matters throughout: European regulation slows product rollout in Europe. It has not been shown to slow US frontier R&D, because training happens on US compute regardless of where the model is later sold.