> ## 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 17, 2026
- URL: https://adjacent.media/briefs/2026-09-17/
- Published: 2026-09-17T14:15:03.000Z
- Updated: 2026-09-17T14:15:03.000Z
- Description: US data centers could consume more natural gas than Germany and Japan combined by 2035 on current deployment trends, and new NYT/Siena polling found data center construction unpopular across party lines.
- Author: Jonathan Greene
- Tags: #brief

**TL;DR:** US data centers could consume more natural gas than Germany and Japan combined by 2035 on current deployment trends, and new NYT/Siena polling found data center construction unpopular across party lines. Boston cancelled its contract with Flock Safety after determining the company shared the city's license-plate reader data nationwide. And OpenAI's models out-earned Anthropic's on OpenRouter for the first time since February 2024.

## Worth Reading

- [The next AI margin fight is the cable between the chips](https://www.theregister.com/systems/2026/09/15/ai-networking-startups-race-to-replace-nvidias-nvlink/5296672?ref=adjacent.media) — A startup swarm is going after NVLink, the least-discussed and most defensible piece of Nvidia's stack.
- [Data centers are on track to become the world's third-largest gas consumer](https://techcrunch.com/2026/09/15/us-data-centers-could-consume-more-natural-gas-than-germany-and-japan-combined-by-2035/?ref=adjacent.media) — The forecast assumes a demand curve nobody can underwrite past 2028, but the pipelines get built on 30-year terms.
- [Boston proved procurement is a faster privacy enforcer than regulation](https://arstechnica.com/tech-policy/2026/09/boston-dumps-flock-says-it-shared-data-nationwide-in-violation-of-contract/?ref=adjacent.media) — No new law required. A breached contract and a city willing to walk.
- [Doom marketing has a lobbying budget behind it](https://open.substack.com/pub/zeteo/p/big-ai-wants-you-to-believe-it-could) — Zeteo's argument: existential-risk rhetoric conveniently favors the companies already holding the frontier.
- [Apple's most-copied product this year is a box with no screen](https://www.yankodesign.com/2026/09/15/the-mac-mini-form-factor-is-so-perfect-literally-everyone-is-copying-it/?utm%5Fsource=rss&utm%5Fmedium=rss&utm%5Fcampaign=the-mac-mini-form-factor-is-so-perfect-literally-everyone-is-copying-it) — Industrial design as free distribution: PC makers spent 2026 shipping Mac mini clones.
- [Python's package count is outgrowing anyone's ability to audit it](https://rushter.com/blog/pypi-packages/?utm%5Fsource=indieblog.page&utm%5Fmedium=rss&utm%5Fcampaign=indieblog.page) — Growth charts for PyPI double as a supply-chain risk chart.
- [Your floorplan already told you what hybrid work settled into](https://www.forrester.com/blogs/your-office-space-is-talking-are-you-listening/?ref=adjacent.media) — Forrester on reading occupancy data instead of running another return-to-office survey.

## Connected World

**The siting fight is the capacity constraint; the power bill is not.**

TechCrunch's reporting on [data centers landing in towns that already lost a fight with heavy industry](https://techcrunch.com/2026/09/15/the-ai-data-center-boom-is-colliding-with-cities-scarred-by-big-industry/?ref=adjacent.media) is the part of the compute buildout that doesn't show up in capex guidance. Chester, Gary, and their peers were selected because land is cheap and zoning is permissive, the same logic that put refineries and incinerators there. The difference is institutional memory: these communities have run the organized-opposition playbook before, and municipal construction restrictions are already slowing regional deployment. For anyone modeling compute availability, permitting friction is the variable to watch; power price is secondary.

That's what makes the UAE's pitch legible. When the country's AI minister told a Dreamforce audience that [data centres matter more than bridges](https://thenextweb.com/news/uae-ai-minister-al-olama-data-centres-dreamforce-salesforce?ref=adjacent.media), the substance was the implicit offer of a jurisdiction where siting goes uncontested. Sovereign compute buyers are selling speed-to-energized-rack. That's a real product, and US hyperscalers are the ones creating demand for it.

**Gas contracts outlive the demand forecasts they're justified by**

The projection that [US data centers could burn more natural gas than Germany and Japan combined by 2035](https://techcrunch.com/2026/09/15/us-data-centers-could-consume-more-natural-gas-than-germany-and-japan-combined-by-2035/?ref=adjacent.media) puts the sector third globally among gas consumers. The forecast extrapolates current deployment plans, which is its weak point, but the infrastructure being signed against it is not speculative. Pipelines and turbines amortize over decades against AI demand nobody can credibly underwrite past 2028\. If inference costs keep falling and open models keep closing the gap, the stranded-asset risk lands on utilities and ratepayers, not on the labs that triggered the buildout.

## Culture & Signal

**Streamers turned the greenlight into an option, and producers eat the downside**

The Entertainment Strategy Guy tracks the [return of the un-order — and now the un-un-order](https://open.substack.com/pub/entertainment/p/un-orders-are-back-along-with-un), where platforms cancel shows they've already ordered, then occasionally reverse themselves. High Rotten Tomatoes scores are losing to viewership and subscriber-retention math, with platforms cancelling critically acclaimed shows on pure performance calculus. This repricing of creative risk mirrors the shift of showrunner responsibility toward digital creators: platforms are pushing development cost and uncertainty onto the supply side while retaining the kill switch. If you're financing content, the commitment you signed is closer to a call option than a contract.

**The privacy enforcer turned out to be procurement**

Boston cancelled Flock Safety after finding the vendor [shared license-plate data far outside the city's authorized scope](https://arstechnica.com/tech-policy/2026/09/boston-dumps-flock-says-it-shared-data-nationwide-in-violation-of-contract/?ref=adjacent.media). No statute triggered it, no regulator acted: a city read its own contract terms and walked. Municipal purchasing is currently doing more surveillance governance than any federal body, which matters for every vendor selling data-collecting infrastructure into local government. Your contract language is your compliance regime, and a single city's audit can cost you a reference customer and a category.

Why the regulatory layer stays quiet is partly explained by Popular Information's look at the [consultants and lobbying shops that sell influence to both parties at once](https://open.substack.com/pub/popularinformation/p/corporate-mercenary-is-a-bipartisan). When the same firms staff both sides, corporate policy outcomes are hedged against election results. That's the structural reason enforcement keeps arriving from buyers rather than legislators.

## The New Consumer

**People are opposing the specific AI experiences that have reached them**

NYT/Siena polling adds to a run of surveys finding [data centers and AI expansion unpopular across essentially every demographic cut](https://www.theverge.com/ai-artificial-intelligence/995917/data-center-nyt-midterm-poll-september?ref=adjacent.media). Attitudinal data is the weaker kind, so the more useful companion is what people are actually experiencing: 404 Media documented an [AI agent platform that reinvented mass email spam and flooded inboxes worldwide](https://www.404media.co/ai-agent-platform-reinvents-spam-floods-inboxes-worldwide/?ref=adjacent.media). For most consumers, the first agentic AI that ever contacted them was a spam campaign. That's the product surface generating the polling numbers, and it's why "AI-powered" is becoming a liability in consumer marketing copy while remaining a selling point in enterprise procurement. Brand teams should stop treating the two audiences as the same market.

**Low-production text now reads as proof a human made it**

The Grudge Report's riff on [American history taught through Notes app screenshots](https://open.substack.com/pub/besskalb/p/more-american-history-via-notes-app) is a joke about a real format shift. The most-shared artifact on Instagram is often a photo of unstyled text, no edit, no voiceover, no B-roll. In a feed where polish is cheap and synthetic, the absence of production value is doing authorship verification. Any brand currently investing in higher-fidelity social creative should check whether fidelity is now working against it.

## Machines & Minds

**Four months of lead time costs five times the price**

Kottke flags the number worth putting in every AI procurement deck: open-weight models are [roughly four months behind the Anthropic and OpenAI frontier at about one-fifth the cost](https://kottke.org/26/09/0049605-the-frontier-anthropic-op?ref=adjacent.media). For classification, summarization, extraction, and routing — the actual bulk of enterprise workload — four months of capability lag is invisible. The frontier premium is only rational for the narrow band of tasks where the delta shows up in output quality you can measure.

Forrester's argument that [frontier labs are shipping ahead of what enterprises can absorb](https://www.forrester.com/blogs/frontier-ai-lab-are-out-of-pace-with-the-realities-of-enterprise-ai/?ref=adjacent.media) sits directly on top of that math. Model releases arrive on an eight-week cadence; enterprise integration, security review, and change management run three to four quarters. Companies are paying a premium for capability that sits in a pilot queue. The gap is not a technology problem, and buying the newest model won't close it.

**The binding constraint is the auditor, not the model**

Workiva's [bet that trust and verifiability are the product as AI agents enter financial reporting](https://siliconangle.com/2026/09/15/workiva-bets-trust-ai-agents-enter-financial-reporting-amplify/?ref=adjacent.media) identifies the unresolved problem cleanly: auditors have no established method for verifying an autonomous agent's decision logic. SOX attestation requires a control you can describe and test. An agent that reaches the right number by an unreproducible path fails that bar regardless of accuracy. Whoever solves agent auditability sells to every public company, which is why governance tooling is a better place to look for enterprise AI revenue right now than another copilot.

**The safety pact is a market-structure move**

OpenAI, Anthropic, and Google [quietly coordinating on AI safety work](https://siliconangle.com/2026/09/15/openai-anthropic-and-google-secretly-joined-forces-to-collaborate-on-ai-safety/?ref=adjacent.media) is being read as encouraging, and the technical work may well be. Ask the standard question anyway: who benefits from the resulting standards. Three firms holding the frontier, jointly defining what responsible development looks like, produce criteria that the open-weight ecosystem four months behind them cannot easily satisfy. Anthropic's leadership has spent the year arguing publicly for slowing down while shipping capability faster than its customers can deploy it. Both things can be sincere and still function as a moat.

Apple's version of the trust argument is asking users to [opt in to training Siri and its foundation models](https://mjtsai.com/blog/2026/09/15/opting-into-training-siri-ai-and-apple-foundation-models/?ref=adjacent.media). Michael Tsai's read is worth sitting with: consent-based collection is genuinely differentiated, and it's also Apple requesting the data its privacy posture prevented it from gathering. The privacy frame is doing double duty, brand asset and data acquisition pitch in one screen.

## Brand & Growth

**The manager's day is turning into a review queue**

A first-person account from The Landing Pad — ["I used to review my team's work. Now I review Claude's."](https://thelandingpad.substack.com/p/i-used-to-review-my-teams-work-now) — describes a change in managerial labor that hasn't reached any job description yet. The work shifted from developing people to evaluating output, and the reported cost is that the feedback loop which used to build junior capability now terminates in a model that doesn't retain the lesson. One essay alone does not make a trend, yet it aligns precisely with what runs through today's enterprise coverage: the scarce input is judgment about whether the output is right, and organizations have been delegating that function too.

**Meta's creator economics only run one direction**

Simon Owens details Meta's move to [charge creators for content they currently supply for free](https://open.substack.com/pub/simonowens/p/meta-wants-to-charge-creators-for), increasing the platform's take from monetization programs. Read it alongside the collapse in referral traffic to independent publishers and the McClatchy layoffs: the supply side of the feed is being squeezed from both ends, less distribution out, higher rent in. Any brand whose content strategy depends on organic reach through Meta properties should price that channel as rented, not owned, and budget for the rent to rise.

## Commerce Rewired

**Model loyalty at the API layer is thin**

OpenRouter reported that [spending on OpenAI's models exceeded Anthropic's in the week of September 7](https://x.com/openrouter?ref=adjacent.media), the first reversal since February 2024\. Don't over-read one week from one aggregator: OpenRouter users are self-selected for price sensitivity and model-agnostic routing, and that's the actual finding. At the API layer, switching costs are approximately zero and allegiance lasts as long as the price-performance ratio holds. Anthropic's durable revenue lives in enterprise contracts with integration depth, not in developers who can change a string in a config file. The "who's winning" leaderboard built from aggregator data also measures something different from what enterprise buyers are doing.

**Retail media is buying the top of the funnel now**

Retail media networks are pushing past onsite search and product listings into CTV, in-store screens, and off-site inventory, which Beet argues [forces marketers to restructure the funnel rather than optimize within it](https://www.beet.tv/2026/09/retail-medias-expansion-beyond-the-aisle-is-forcing-marketers-to-rethink-the-entire-funnel.html?ref=adjacent.media). The consequence is a measurement takeover: awareness spend routed through a retailer's network gets judged on closed-loop attribution to basket, because the retailer can measure that. Brand budgets moving into these networks get repriced as performance budgets whether or not the CMO intended it, and the metrics that justify long-horizon brand building don't survive the transition.

---

*18 articles across 6 themes · 16 sources · Powered by Folo + Claude*