The Adjacent Brief

TL;DR: Texas's grid operator stopped approving new data center interconnections, saying it can't serve the load already queued. SpaceX disclosed $2.6B in AI hosting revenue — more than it earned launching rockets — and Bloomberg reported that the spread between top- and bottom-performing 2024 venture funds has more than doubled versus the 2017–2021 vintages.

Worth Reading

• AMD stops selling chips and starts selling systems — When the component vendor moves to bundled packages, it's because customers are buying capacity rather than individual silicon.
• High ratings and high attention are not the same inventory — TVision is trying to reprice TV on eyes-on-screen, which would move money away from reach-first buys.
• More AI content means more demand for editors, not less — 1440's Tim Huelskamp on curation as the scarce input when supply goes infinite.
• Cloudflare's AI bot blocker is also blocking Googlebot — The cost of crawler defense is showing up as lost search indexing.
• Telegram's App Store removal was engineered by an extortionist — Planted content plus a review policy is now a coercion tool anyone can use.

Brand & Growth

Borrowed trust is rented, and the rent is going up

A newsletter for fans of leagues the big outlets barely staff reached one million subscribers — built by founders with no prior media or sports industry background, per Simon Owens. The same asset underpins the case for routing more spend through creators as attention fragments per 404 Media, a correction issued because token consumption had become a visible proxy for AI adoption inside the company. Any organization measuring AI progress by usage volume is about to learn the same thing: engineers optimize what's counted, and tokens are cheap to burn. If your Q3 AI dashboard reports seats activated, prompts submitted, or tokens consumed, it's measuring compliance theater. The number that matters is work shipped per headcount, and almost nobody is instrumenting it.

Connected World

Power capacity, not chips, decides where compute lands

Texas's grid operator stopped approving new data center connections outright, citing demand it cannot serve, in the state that spent three years marketing itself as the frictionless alternative to permitting fights elsewhere. The engineering argument playing out in the Slashdot thread on whether this is a capacity problem or a queue-management problem matters less than the commercial consequence: interconnection timelines, not GPU allocation, are becoming the binding constraint on 2027 capacity plans. Anyone with a signed power agreement in a served region just got more valuable. Anyone whose AI roadmap assumes new US capacity comes online on schedule should re-forecast.

Export controls are building AMEC a customer base

Samsung and SK Hynix are evaluating Chinese chipmaking equipment from AMEC for their mainland fabs, Reuters reports, as a hedge against further US restrictions. The controls were designed to keep advanced tooling out of China; the second-order effect is two of the world's most demanding memory manufacturers qualifying a Chinese toolmaker's equipment — exactly the validation AMEC needs to sell everywhere else. Qualification takes years and creates switching costs in both directions. Watch this hedging behavior across the stack: every customer that diversifies under policy pressure permanently widens the field of credible suppliers.

Culture & Signal

The federal AI framework covers only the models with a corporate address

The US framework defines a covered frontier model as closed-source, state-of-the-art, and nationally security-relevant, which exempts open-weight models entirely, per Axios sourcing. Labs releasing weights benefit, as does anyone deploying those weights without disclosure obligations. Compliance burden lands on the handful of companies with lawyers and API logs, while the fastest-diffusing capability sits outside the perimeter. For enterprise buyers, vendor due diligence is doing regulatory work no agency is doing.

A stated use case is not a control

Police used Flock's license plate reader network to track a man across state lines and construct a pretextual traffic stop for a weed search, directly contradicting the company's public account of how the cameras get used. This keeps recurring in the surveillance-vendor category: the marketing describes the intended application, the deployment describes the actual one, and the gap only surfaces through court filings. Your published use-case language is a liability document when your company sells data or detection capability, regardless of how it functions as positioning.

Search results are becoming a citation lottery

Google's placement of Top Stories inside AI Overviews changes what publishers and brands can expect from news-adjacent queries: visibility without the click, and inclusion decided by systems no one can optimize against reliably. Combine that with Cloudflare's bot blocking knocking sites out of Google's index, and both defending your content and being discovered now carry technical failure modes for publishers. Owned channels, like the newsletter above, look better every quarter for reasons that have nothing to do with editorial preference.

The New Consumer

Proof of human authorship is a product now, with a price

A startup is selling cryptographic seals on manuscript drafts to prove a person wrote them: timestamped, versioned, sellable as evidence to publishers. It's a narrow service built on a broad condition: AI-generated content volume passed human-generated volume this summer, and the burden of proof has moved to the creator. Watch whether buyers, publishers, agencies, and universities start requiring provenance as a submission condition. That's the moment verification stops being a novelty purchase and becomes procurement.

The consent problem lives in the SDK, not the app

Android developers may be handing user location data to advertisers without knowing it, via third-party SDKs whose data flows they never audited. Nanit's baby monitors, meanwhile, have moved from watching an infant breathe to predicting developmental outcomes: a longitudinal profile of a person who won't be able to consent to it for eighteen years. The structural issue is the same at both ends: the company collecting data and the company monetizing it are not the same company, and the privacy policy describes only the first one. If you ship a mobile app, your actual data practices are whatever your dependency tree does.

Commerce Rewired

The AI trade split venture into two asset classes

The performance gap between top- and bottom-quartile 2024 venture funds has more than doubled relative to 2017–2021 vintages (paywall), Bloomberg reports, driven by access to a small number of AI deals. Dispersion that wide changes LP behavior before it changes founder behavior: allocators stop diversifying across managers and concentrate with the handful who get into the rounds. Smaller funds without that access are left underwriting the companies built on top of the winners' infrastructure — a real business, but priced very differently.

SpaceX is a compute landlord that also flies rockets

SpaceX booked $2.6B in AI revenue, roughly triple last year and more than its launch business. Read that against the Texas interconnection halt: the scarce asset in AI hosting is power, land, and the operational competence to run hardware at scale, all of which SpaceX built for another purpose. For anyone with industrial infrastructure on the balance sheet — utilities, REITs, telcos, manufacturers — compute hosting is the highest-margin use of a substation and a secured site right now, and the demand queue is long enough to underwrite it.

Machines & Minds

Self-replicating agents stopped being a thought experiment

Researchers demonstrated an AI worm propagating across a live agent system, WIRED reports, which makes autonomous AI malware a working technique rather than a scenario. The distribution channel for it is arriving separately: open-weight models are reaching frontier benchmark performance with meaningfully fewer safety mitigations than comparable closed systems. Anyone running agents with tool access and inter-agent messaging should treat those permissions as network attack surface and scope them accordingly. The containment question is an infrastructure decision, and the federal definition of a covered model won't answer it.

Distillation makes cheap capability the real story; benchmark scores are beside the point.

Knowledge distillation, large models training smaller ones to reproduce their behavior at a fraction of the compute, is the mechanism behind both the open-weight catch-up and the falling cost of deployment. Where that lands as demonstrated utility: more than twenty mathematicians at the 2026 International Congress described AI changing the actual practice of their work, mostly optimistically, in literature search, proof checking, and conjecture testing. That's a repeatable value loop — the work is verifiable, the errors are catchable, and the practitioner stays accountable for the output. It's a better template for enterprise deployment than any agent demo released this year.

18 articles across 6 themes · 15 sources · Powered by Folo + Claude