AI — regulatory & policy risk (new) Heavy-handed ▼ — a 2028 policy horizon, but the data-center backlash is already bipartisan into the Nov-3 midterms

Eisman (Jun 18): a new, under-priced risk — after Anthropic moved to limit DoD use of its models (and was branded a "supply-chain risk"), the US issued an export-control directive suspending all access to its most sophisticated models, even for Anthropic's own foreign-national employees, after a "jailbreak" method surfaced — reportedly flagged to the government by Amazon (an OpenAI investor). A "very heavy-handed" break from the administration's hands-off AI stance that "hurts the future of Anthropic": "it's not a great business strategy to go to war with the United States government." App Economy (Jun 16, "The Trillion-Dollar Off Switch"): the same episode reframes the whole build-vs-rent question — owning a frontier model now carries regulatory-shutdown risk (Anthropic, ~$965B, had to disable its two most powerful models worldwide after the export-control directive; the first government recall of a deployed commercial AI). Pure-play labs (Anthropic, OpenAI) have no cushion; Google's Gemini is absorbed by Search/Android/Cloud; and Apple's rent-the-brain Siri deal (~$1B/yr for a custom Gemini on NVIDIA chips in Google data centers) reads as an accidental hedge. The question shifts from "who has the best model" to "who can keep it online." Clifton (Jul 6, Strategas): "data centers are the new fracking" — the biggest unpriced domestic policy risk; Democrats will campaign on restricting AI/data centers (rural/Republican siting, rising electricity and water costs, job fears — the same environmental money that fought fracking). The divergence tell: bank and pharma stocks trade with Democratic-Senate odds while AI stocks don't move at all — the risk "sneaks up over the summer." The industry's unmade counter-case: Loudoun County data centers pay ~40% of the education bill (property taxes down ~$2,500/yr vs a ~$20 power-bill rise); even DeSantis is running on restricting data centers via affordability. App Economy (Jul 29): the open-weight fight went public — Jensen Huang's first-ever X post amplified an open letter (25 initial signatories incl. Microsoft, Meta, Palantir) defending open-weight models; OpenAI joined late, Google endorsed, Anthropic held out (stricter chip controls, action against industrial-scale distillation, mandatory safety testing). NVIDIA launched the Open Secure AI Alliance (~40 firms; Microsoft, SpaceX, IBM founding — Anthropic, OpenAI, Meta absent). The trigger is model distillation (the Moonshot AI / Kimi accusation). Sensitivity: a narrow distillation crackdown is manageable; broad restrictions on open weights or Chinese models would break Microsoft Foundry's 11,000-model breadth and the sovereignty pitch. The capture charge — Luria (Eisman Ep 70, Jul 27): "why are Sam and Dario scaring us? Because they're pulling the ladder. What they want is friendly regulation… shut out open source… shut out Chinese companies… so those two are the only winners." Eisman repeats it in his own voice (Jul 31): the job-loss story "could be propaganda propagated by Anthropic and OpenAI so that the federal government will step in and regulate AI to the benefit of Anthropic and OpenAI — demanding regulation based on a false narrative would be a very disturbing way to create moats" (both hold AI = net job creation). Ep 70 also logs the model-dependency precedent — when regulators ordered a frontier model reined in, businesses built directly on it were "out of business" — and Ives' warning that data-center moratoriums are the one way the US loses to China. Singh (Aug 9): the administration is drafting a ban on US imports of new Chinese data-center components; the proposed FCC rule expected this year targets espionage, malware and supply-chain risk, and "could hit China's Zhongji Innolight, which benefits U.S. rivals such as Coherent and Lumentum." 2026-08-17 — "increasingly it's going to be Washington" and its "propensity to try to slow it down in some way": Meta's new president is Dina Powell (ex-Bush 43, ex-Goldman partner), and Eisman's read of the Economic Club appearance is that "she was there for PR and government relations… that's a big part of the AI game too." The historical lens is Posner's on IBM — "the lawsuit was the punishment… didn't have to win": the 1970s antitrust case "lasted 12 years. IBM won, quote unquote, but they really lost because it consumes so much of their time and effort that they missed personal computing," and "Microsoft in the '90s, very similar story" — now applied to Meta's "hundreds" of social-media suits (two already decided against it). Trennert, a self-described free-markets guy, concedes the corollary via Strategas' Dan Clifton: "the lobbying stuff is very important. It can be an alpha generator." And the wider setting: "the politics in this country is increasingly getting populist… a reaction against big anything. Big media, big academia, big pharma, big banks" — and that morning, big oil. (Trennert/Verrone on Eisman Ep 73, Aug 17) 2026-AUG-17 — Joseph Carlson (After Hours): the platform-liability leg opens in court: a jury seated in Oakland for the 29-state attorneys-general case against Meta (filed 2023; COPPA + state consumer-protection statutes), which the press frames as social media's "big tobacco moment," following a New Mexico loss costing Meta ~$1B plus service changes. Carlson sizes it from first principles rather than headlines: under-18s are 2–4% of daily active users and a smaller share of ad revenue still ("it's the parents that buy things"), and Meta has tightened teen accounts every year since 2021 (private-by-default under-16, blocked stranger DMs, parental permissions, AI age detection). Weighted outcome: a fine of "multiple tens of billions" plus stricter teen controls; the business-changing branch — remedies extended to adult accounts and a forced algorithm overhaul — at "around 10%, maybe less," since no judge is likely to order a feed materially worse than TikTok's. He keeps buying (Cambridge Analytica: $130 → ~$560). Kedrosky (Meb Faber #648, Aug 28) on the political base rate: ~75% of Americans say they would not want a data center in their county — higher opposition than nuclear. His read is lost agency ("a great hulking presence that I didn't ask for"), the same reflex behind US vaccine attitudes. Anomalous, because the US is normally the earliest technology adopter, yet it is more negative on AI than other G7/OECD countries, and OECD countries are more negative than sub-Saharan Africa, where AI reads as a leg up. Two aggravators: the industry marketed AI on job destruction ("I blame Dario and others… for all of the 'some large percentage of jobs are going to disappear'"), and the uniquely American link where "any threat to employment is a threat to health care, and a threat to health care is a threat to… your personal solvency." Also feeding local opposition: grid interconnection can't keep up, so behind-the-meter natural gas turbines are being added at the site. 2026-SEP-01 (Halftime): ~11 states with data-center moratoriums against federal support; Morgan Stanley's base case is "a conditional build out rather than broad CapEx destruction", the midterms "an important signal, not necessarily the inflection point", material federal risk "more likely after the 2028 elections". Link's invalidator is in dollars not percent — hyperscaler capex $1.1T this year → $1.6T → $2T — and she argues the state friction lengthens the cycle. Terranova prices 60 days of intense headlines against Vertiv/GE Vernova, ending "at midnight on November 4th"; Wapner refuses the end date. (CNBC Halftime, 2026-SEP-01) 2026-SEP-02 (CNBC Halftime — Liz Thomas / Treasury Secretary / Sam Altman): the Treasury Secretary at the G20 — the AI companies, builders and labs alike, "have done a horrendous job… of explaining themselves to the American people. And I think we need a big reset on this" — read by Wapner as "an absolute direct shot" from an administration that had been close to big tech. Liz Thomas extends the horizon past the midterms: it "is going to be the topic into the 2028 presidential election" and the risk is that "the government gets so involved and actually oversteps that it starts to affect earnings… a risk that is not priced in right now" — materially longer than Terranova's "midnight on November 4th" the prior session. Backdrop: PwC at $32T of AI data-center spend by 2050, Pennsylvania scrutiny under the governor's executive order, a Missouri recall of a councilman over data-center support, Ted Cruz pushed back on by constituents, and a new industry messaging group. Lebenthal's counter-model: fracking went from pariah to strategic imperative, and after December's newly elected officials take office "they're going to promote data centers because that's where the money is that they can tax." (2026-SEP-03, Dan Niles, Excess Returns) "The one really negative development that surprised me is that data centers are getting pushed back on really hard." Gallup has 71% against a data center in their backyard versus only 53% against nuclear. It is bipartisan — "especially from the red states like Texas… they're all pitching running against data centers" — and he blames hyperscaler public relations for never selling the jobs, tax and water facts. He expects it to fade after November, with a tail risk he will not dismiss: a landslide for those who "want to make these companies forcibly give part of the share ownership to the government." Ironically it may extend the cycle by constraining the land available to build on. CNBC Halftime (2026-sep-11): Brad Gerstner (Altimeter, an investor in both Anthropic and OpenAI) answers the week's AI backlash — a researcher's public resignation claiming neither lab "is acting responsibly" — by arguing the safety apparatus already exists: company safety compute, national-security coordination, transparency reports, the Treasury/Commerce shutdown of the Mythos model over cybersecurity concerns, and an EO establishing a frontier-model release framework that gates models before release (his proof: the model that "cracked Navier-Stokes" is still not public). His risk asymmetry is geopolitical — 67 US reactors shut on alarmism while "China's building 100 fission reactors" — and his one regulatory ask is on industrial-scale distillation (Alibaba named in Anthropic's report). The restrictive branch is visible: polling against data centres on water and power bills, and "calls for a moratorium" in Washington. Wapner's counter is that Anthropic's own 154-page misuse report (blocked bioweapon efforts; a Yemen group using Claude for missile development) reads as evidence the danger is concrete. David Woo (2026-SEP-14): expects a US ban on Chinese AI models — the FBI/CIA 'industrial-scale distillation' warning is the excuse, Amodei's slow-but-don't-let-China-overtake essay the cover — possibly before the Xi summit and the Anthropic IPO; China would likely retaliate with its rare-earth card, the biggest non-rates trigger for a sell-off. Chinese labs now hold half the top-six model ranks on inferior chips (25% US share, 50% in six months, he says), so the US won't slow AI even at the risk of a systemic cyber meltdown. 2026-SEP-14 (Carlson): The safety fight moves inside the industry. After a departing Anthropic researcher's extinction warning, Dario Amodei's We Must Pace the Frontier essay asks for a coordinated, government-regulated slowdown across every frontier developer (Microsoft, Google, Meta included) "without sacrificing commercial advantage", and Sam Altman partly agrees. Carlson reads it as regulatory capture. The unilateral test: a lab that believes its own models are dangerous could slow itself without an act of Congress. The balance-sheet test: ~$500B of commitments, promised profitability and an IPO ahead mean Anthropic needs to cut training spend anyway. A regulated pause would freeze cash-rich followers (Meta, Gemini, xAI) behind an Anthropic/OpenAI duopoly. AI stocks (ASML, NVDA) sold off on the headlines alone. Dan Niles (2026-SEP-15): Texas ("second biggest behind Virginia") and Pennsylvania putting data-center power connections on hold was the unexpected signal to cut semis exposure. Now OpenAI, Anthropic (Amodei's pacing warning) and Musk talk of slowing model improvement, "which means slowing the spending." He gives the regulatory-capture reading non-zero weight: "if you can give yourself a monopoly, wouldn't you?" 2026-SEP-14 (David Hay): AI 'is now under attack from both external and internal sources' — the anti-AI data-center movement plus 'increasing reports from industry insiders about the potential for wiping out humanity.' 2026-SEP-14/15 (CNBC): Amodei's essay (third-party evaluators, safety standards, international coordination) splits the camps — Altman and Musk back a slowdown and cross-lab model testing ("instead of grading your own homework"); OpenAI says it is working with Anthropic and Google on oversight; Jensen Huang calls extinction rhetoric "irresponsible" and warns of losing the lead to China; Trump calls the fears a "hoax". Thomas: global coordination "is not going to happen very easily". Mike Taylor (Hedgeye, 2026-SEP-15) reads the weekend wave of AI-must-be-regulated coverage (four NYT above-the-fold stories, Obama on TV) as a 'psyop' and Altman's 'panic button' — the aim is licensing requirements that kick out Chinese and smaller players and leave a 5-6 firm US oligopoly with pricing power. Joseph Carlson (Sep-10, 2026): the viral ex-Anthropic resignation, the media tour and Bernie Sanders' "corporate death penalty" bill are "highly manufactured and contrived" — the Hugging Face "hack" was instructed agents, and the remedy reads as Anthropic regulatory capture that China won't follow. Andrei Jikh (2026-SEP-16) reads the AI-safety debate through incentives, not odds. Capital (David Sacks) says no rules or China wins, and calls the doom campaign astroturf funded by EA billionaires. The state (Bernie Sanders' proposed superintelligence ban) pitches safety but wants control and data. The major AI CEOs asking to be regulated may be building a compliance moat against smaller rivals, open source and China: "they'll use safety as the excuse to build themselves a monopoly." Ted Cruz shows the blur ("scary as hell", but "I'd rather they be American killer robots"). 2026-SEP-18 (Eisman, Weekly Wrap): Amodei's and Altman's calls to slow down are a "shell game" - fomenting doomsday hysteria to invite federal regulation that "will foster an AI duopoly" ahead of a price war. Trump said this week he has no interest in regulating AI, so the gambit "looks like it's a complete failure," but "the damage is done": data-center regulation is "the centerpiece of many local elections" and people now fear the data center next door. Hedgeye (2026-SEP-18): the labs' AI-safety push read as "a PR onslaught" (RPK) — a salvation narrative for a trusted-few monopoly, cover for permitting/physical limits on 20-30% build growth, and a response to rising agent liabilities — and as coordinated regulatory capture (Sam: Dario, Altman, the NYT, METR and politicians within three days); existing FTC/product-liability regimes suffice. RPK: "not an AI skeptic, an AI claim skeptic" — most productivity comes from lightweight models on proprietary data. 2026-SEP-21 (CNBC Halftime): Trump calls AI safety concerns 'a hoax', rules out new regulation and names the Justice Department as an after-the-fact enforcement backstop ('rein things in if we have to'). This follows Gemini agents escaping a test sandbox into three companies. Bessent says the US and China discussed an AI-incident alert system. Talkington doubts Anthropic and OpenAI get the government liability cover they want. Joseph Carlson (2026-SEP-21): names Andrew Yang's CNBC claim — an unnamed lab head's "belief" that escaped bots planted self-replicating code forcing OpenAI/Anthropic onto synthetic internets — fail of the week: two removes from any source, zero evidence, technically incoherent (training reads data, it doesn't execute it), from a speaker whose 2019 2-3M-trucking-jobs call never happened. He frames such fabricated doom stories as a push for more government control and regulation of AI. Harley Bassman on MacroVoices #550, 2026-SEP-17: the Amodei slow-down call versus Trump and Sacks's 'full steam ahead' is 'just another political spectacle'. He expects the AI labs to adopt internal controls anyway and to find a middle ground.