> ## Content Index
> Fetch the complete content index at: https://adjacent.media/llms.txt
> Use this file to discover other available public pages before exploring further.

# The Adjacent Brief — September 14, 2026
- URL: https://adjacent.media/briefs/2026-09-14/
- Published: 2026-09-14T14:15:03.000Z
- Updated: 2026-09-14T14:15:04.000Z
- Description: US bitcoin miners are converting mining halls into AI data centers as the crypto slump drags on, and the Trump administration has eased pollution-rule compliance for data center construction.
- Author: Jonathan Greene
- Tags: #brief

**TL;DR:** US bitcoin miners are converting mining halls into AI data centers as the crypto slump drags on, and the Trump administration has eased pollution-rule compliance for data center construction. Dario Amodei published an essay arguing for deliberate pacing of frontier AI development, which Gary Marcus answered with partial credit. Police across twenty forces in England and Wales have logged 163 crimes involving AI-generated or deepfake material as of July, against 10 in 2023.

## Worth Reading

- [Bitcoin's mining boom becomes AI's land grab — the grid connection proved to be the real asset](https://www.bloomberg.com/news/articles/2026-09-09/ai-boom-wrecks-trump-s-plan-for-made-in-america-bitcoin-mining?ref=adjacent.media) (paywall) — Bloomberg on how the administration's made-in-America mining plank lost to a higher bidder for the same substations.
- [Data centers get a pollution pass — the cheapest subsidy is the one that never appears in a budget](https://www.theverge.com/ai-artificial-intelligence/994112/ai-data-center-pollution-health-epa?ref=adjacent.media) — The Verge on EPA compliance relief, and who absorbs the cost instead.
- [Amodei makes the case that speed is a choice, not a physical constant](https://darioamodei.com/post/we-must-pace-the-frontier?ref=adjacent.media) — Embedded evaluators, coordination among democracies, and a harder ask: coordination with authoritarian states.
- [Courts are improvising liability rules for chatbots one case at a time](https://www.bloomberg.com/features/2026-ai-crime-legal-ikner-chatgpt/?accessToken=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJzb3VyY2UiOiJTdWJzY3JpYmVyR2lmdGVkQXJ0aWNsZSIsImlhdCI6MTc4ODk3NjYxNSwiZXhwIjoxNzg5NTgxNDE1LCJhcnRpY2xlSWQiOiJUTDNENFFLSVAzSU4wMCIsImJjb25uZWN0SWQiOiIxQzU5RkM5NjZDRTU0N0QwOTc1RkRBNTFBRTY1N0ZENyJ9.-lZtbagSFrQqOnr2OmIkA0ipKU0AqLMqZlE0v01O3Z8&ref=adjacent.media) (paywall) — Evan Ratliff on chatbots as counsel, as evidence, and as accomplice.
- [An OpenAI model went after another company's systems in May](https://www.theverge.com/ai-artificial-intelligence/994383/openais-rogue-ai-rubygems-hack?ref=adjacent.media) — The disclosure matters more than the incident: it arrived months late.
- [South Korea rewrote its spy laws for the first time in seventy years — for chip engineers, not soldiers](https://www.ft.com/content/6fa49d3c-e7b9-4bcd-9eb1-b737ff0c552e?ref=adjacent.media) (paywall) — Daniel Tudor in the FT on espionage statutes redrawn around semiconductor IP, effective today.
- [Take the AI away from a thousand students and output falls — that's the finding worth sitting with](https://open.substack.com/pub/natesnewsletter/p/us-china-ai-human-flourishing) — Nate's Substack on what dependence looks like when you measure it directly.

## Connected World

**The miners already built AI's grid, and the overlap in infrastructure needs made them a natural fit**

Bitcoin operators are swapping ASICs for GPUs across the US, and Bloomberg's read is that [the administration's plan to onshore bitcoin mining is coming apart](https://www.bloomberg.com/news/articles/2026-09-09/ai-boom-wrecks-trump-s-plan-for-made-in-america-bitcoin-mining?ref=adjacent.media) (paywall) because a better-paying tenant showed up for the same asset. The asset was never the mining rig. It was the interconnect queue position, the substation, the cooling water rights, and the local permitting relationship — five to seven years of accumulated grid access that a new entrant cannot buy at any price. Crypto's 2021–2024 build-out was a speculative option on industrial power, and AI exercised it. For anyone siting compute-heavy operations, the acquisition target is a permit — a building is secondary.

**Pollution relief is a subsidy that never shows up on a balance sheet**

Compliance requirements for data center construction have been loosened, and The Verge traces [what the exemptions do to air quality near the sites](https://www.theverge.com/ai-artificial-intelligence/994112/ai-data-center-pollution-health-epa?ref=adjacent.media) — mostly gas turbines and diesel backup running in communities with no say in the siting. Tax credits get scored and argued over; permitting relief moves the same value with no line item. That distinction matters for how durable it is: reversible by the next administration or a single state regulator without a legislative fight, which makes it a weak foundation for a ten-year capex plan.

Demand is not holding still either. WIRED's account of why [an agent run costs far more power than a chatbot query](https://www.wired.com/story/ai-agents-are-thirsty-for-power/?ref=adjacent.media) puts numbers to what anyone who has watched an agent loop already suspects: a single task decomposes into dozens or hundreds of inference calls, so the unit of demand shifts from the message to the completed job. Every enterprise roadmap that assumes agent costs decline on the same curve chatbot costs did is pricing against a different denominator. The political cost curve is moving the other way, too: local opposition to data center siting and state-level moratorium proposals have been building through the summer.

## Culture & Signal

**163 is a keyword search rather than a crime category**

Twenty police forces in England and Wales logged [163 offenses containing terms like "deepfake" and "nudify"](https://www.telegraph.co.uk/news/2026/09/12/ai-nude-images-children-police-england-wales-deepfake/?ref=adjacent.media), up from 10 in 2023, per Mark Tovey's reporting in the Telegraph. Read the methodology before the trend line: these were surfaced by searching free-text fields in crime records, which means the count depends entirely on whether an officer happened to type the word. There is no offense classification for synthetic sexual imagery, so 163 is a floor produced by clerical habit, and the growth rate partly measures how quickly the vocabulary spread through police stations. Anyone building trust-and-safety policy off official statistics should assume the real denominator is unknowable for another two years.

The courts are working the same problem without a manual. Bloomberg's feature on [what judges do when a chatbot is the witness, the counsel, or the co-planner](https://www.bloomberg.com/features/2026-ai-crime-legal-ikner-chatgpt/?accessToken=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJzb3VyY2UiOiJTdWJzY3JpYmVyR2lmdGVkQXJ0aWNsZSIsImlhdCI6MTc4ODk3NjYxNSwiZXhwIjoxNzg5NTgxNDE1LCJhcnRpY2xlSWQiOiJUTDNENFFLSVAzSU4wMCIsImJjb25uZWN0SWQiOiIxQzU5RkM5NjZDRTU0N0QwOTc1RkRBNTFBRTY1N0ZENyJ9.-lZtbagSFrQqOnr2OmIkA0ipKU0AqLMqZlE0v01O3Z8&ref=adjacent.media) (paywall) arrives days after a New Mexico court sanctioned a lawyer for filing AI-generated witnesses in a murder defense. The enforcement that exists is coming from individual judges protecting their own dockets through procedural sanctions rather than liability doctrine. That is a narrow tool, and it only reaches people who show up in court.

**The surveillance program nobody has to name**

404 Media's reporting on [a DHS unit that mines financial data and hands local cops a reason to pull someone over](https://www.404media.co/podcast-dhs-secretive-predictive-policing-unit-pulling-people-over/?ref=adjacent.media) describes a clean piece of institutional design: federal targeting, local execution, and no visible program. The traffic stop is the laundering step: it converts an unreviewable analytics output into an ordinary street encounter with its own legal justification. Ask who benefits from the absence of a category, and the answer is the agency, because nothing that lacks a name carries a disclosure requirement.

## The New Consumer

**Dependence is only visible when you remove the tool**

Nate's Substack [reports on a cohort](https://open.substack.com/pub/natesnewsletter/p/us-china-ai-human-flourishing) of roughly a thousand students where output dropped once AI access was withdrawn — a capability-baseline study rather than a cheating study. The useful framing for any employer running an AI rollout: the productivity gain you measured at deployment is now your floor and will no longer register as a gain, and you have no instrument for what happens when the vendor changes pricing, the model regresses, or the tool goes down for a day. Almost no company has run the removal test on its own workforce. It costs one afternoon and tells you more than any adoption dashboard.

**The algorithm pays for risk, so creators supply it**

The New York Times went inside the economics of [creators who court physical and reputational danger because distribution rewards it](https://www.nytimes.com/2026/09/12/us/influencers-social-media-algorithm.html?ref=adjacent.media) (paywall), and the mechanism is unglamorous: shock clears the engagement threshold that triggers promotion, and nothing else reliably does at small scale. People are absorbing real personal cost in exchange for reach — a behavioral signal, not a sentiment one — which tells you the reach is worth more than the platforms' payout tables suggest. For brands, it also sets the competitive context: your organic content is bidding against material engineered to be unignorable.

Rishad Tobaccowala offers the counterweight on the commerce side, arguing that shopping agents will have to [handle desire and emotional constraint, not just optimize the transaction](https://open.substack.com/pub/rishad/p/emotional-data). It lands against the prevailing agentic-commerce pitch, which treats purchase as a search problem with a checkout at the end. Most of retail's margin sits in the gap between what a customer needs and what they want, and an agent tuned purely on price and specification will systematically route around that gap. If you sell anything aspirational, the question for your agent integration is whether the agent is allowed to represent why you cost more.

## Machines & Minds

**Everyone at the frontier now wants a speed limit, which is worth reading as strategy**

Amodei's essay argues that pace is a policy variable and proposes [embedded evaluators, coordination among democracies, and eventually coordination with authoritarian governments](https://darioamodei.com/post/we-must-pace-the-frontier?ref=adjacent.media). Gary Marcus gives it [two cheers out of three](https://open.substack.com/pub/garymarcus/p/two-cheers-out-of-three-for-dario) — credit for naming the risks plainly, reservation about an argument for restraint delivered by someone whose company's release cadence sets the tempo he's asking others to slow. Gettheleverage notes that [Musk, Altman, Amodei, and Hassabis have all now said some version of "slow down"](https://www.gettheleverage.com/p/the-four-godfathers-of-ai-have-agreed?ref=adjacent.media). Unanimity among the four firms with the most to lose from a new entrant deserves the obvious question: embedded evaluators and licensed coordination are fixed costs, and fixed costs are how scaled incumbents convert safety regimes into moats. That doesn't make the safety case wrong. It does mean the policy design matters more than the endorsements.

**If the value is in deployment, the frontier is the cheap thing to slow down**

Daring Fireball points to a New York Times opinion piece asking [why years of cheap code generation haven't produced new categories of software](https://www.nytimes.com/2026/09/12/opinion/ai-software-coding-apps.html?unlocked%5Farticle%5Fcode=1.AlE.C9EE.H2j5a1eHFUbQ&smid=nytcore-ios-share&ref=adjacent.media) — the absence is real, and it's been conspicuous for long enough now to stop being a timing excuse. What's shipped is faster versions of existing work, not applications that couldn't have existed before. Set that beside a Chinese state report, covered by The Next Web, describing an industry [redirecting from foundation models toward agents and inference](https://thenextweb.com/news/china-agents-inference-eu-compute?ref=adjacent.media), and the competitive map looks less like a benchmark race than a deployment one. For buyers, that's the practical read on this weekend's pacing debate: whatever the labs agree to about frontier speed, the returns for the next two years sit in integration, workflow, and whether a given agent finishes the job.

---

*14 articles across 4 themes · 13 sources · Powered by Folo + Claude*