The Adjacent Brief

TL;DR: Australia's ARIA will exclude songs created mostly or entirely by AI from its charts, a month after an AI-assisted track became the most-played song on Australian radio. Analyst Dylan Patel argues Anthropic and OpenAI will control most of the world's usable compute by 2028; separately, EU trade data shows China supplied $29 billion of batteries to Europe last year, after nine years of European battery policy.

Worth Reading

• AI is eating the bottom rung of the ladder, not the ladder — Stanford finds displacement concentrated in early-career cohorts, which means the cost lands on hiring budgets before it lands on payroll.
• AI's capex is flowing into 19th-century equipment categories — Data centers became the killer app for power transformers; the bottleneck on compute is increasingly steel and copper.
• Google put an app on your Android phone without asking, and the uninstall right doesn't help if you don't know it's there — DMA compliance in letter, silent install in practice.
• LinkedIn's 'slop' button turns content moderation into a user-reported metric — Thirty days of data on what happens when a platform lets users flag machine writing.
• Full-color 3D printing drops from $50,000 to a tenth of that, and the prototyping map redraws — HeyGears moves a capability from service bureaus to studios.
• Cheap tactile sensing is the missing input for robots that work outside a demo — 3D-printed skin gives manipulation the feedback loop it has been faking.

Connected World

Compute goes to whoever bills for it best

By 2028, Anthropic and OpenAI will hold most of the world's usable FLOPs because they can convert a GPU-hour into recurring revenue better than anyone else bidding for the same rack. Clouds sell capacity to the highest sustainable payer, and a coding subscription that renews every month outbids a research budget that doesn't. If that holds, the model layer looks less like a competitive market and more like two vendors with pricing power — an argument for anyone building on top to lock multi-year terms now and stop planning around a price war that keeps being promised.

Local politics, rather than silicon availability, now constrains data centers

The piece at vowe.net names the moment operators realize the power math doesn't work: demand growth colliding with grid capacity and, increasingly, with organized local opposition to new sites. Chips ship in weeks; interconnection queues and county hearings run in years. Siting risk belongs in the same slide as chip supply for anyone modeling AI capacity through 2028.

Nine years of industrial policy created demand but left supply untouched

Europe's battery strategy produced the demand and China produced the batteries: $29 billion of Chinese cells landed in the EU in 2025, covering more than the bloc's entire net import need. Subsidizing adoption without securing cells, refining, and cathode chemistry converts industrial policy into an import bill with a green label on it. That structure sits next to the compute story: Europe is funding AI deployment across a stack it doesn't own either, and the lesson from batteries is that the downstream subsidy accrues upstream.

Culture & Signal

A chart rule is the first enforceable definition of "human-made"

ARIA's decision to exclude songs made mostly or entirely by AI from its charts matters because charts are commercial infrastructure. They drive sync licensing, playlist placement, radio rotation, and touring economics. The awkward detail is timing: an AI-assisted track was the most-played song on Australian radio in July, before anybody drew a line — and the debate is about licensing policy rather than taste judgment. The Register's read on an industry now sorting recordings by how they were made gets at the enforcement problem: "mostly or entirely" is a threshold that runs on self-disclosure, and the incentive to disclose accurately is zero. Spotify and LinkedIn are working the same question from the platform side. Attestation, not detection, will be the operating mechanism, which means the compliance burden lands on labels and distributors within the year.

Nobody voted on the curriculum; districts bought it

Natasha Singer's reporting, summarized at The Next Web, traces how free hardware and grant programs turned into control over what American kids learn — Chromebooks at cost, teacher training as marketing, policy advocacy as distribution. The playbook is standard enterprise land-and-expand pointed at institutions with no procurement sophistication and no ability to switch. Set it against ARIA: the music industry drew a boundary within a year of the technology mattering, while schools spent a decade discovering the terms. For anyone selling into public institutions, that asymmetry is the whole business model. For anyone buying, it's the argument for writing exit costs into the contract before the pilot starts.

The New Consumer

Loneliness is a targeting parameter

The mechanism described in The Curious Brain's piece on how isolation became an addressable audience is parasocial: creators supply the relationship, brands rent the trust, and conversion rates on a lonely follower beat conversion on a skeptical one. The companion piece on the cohort raised entirely inside recommendation systems supplies the supply-side explanation: the most efficiently reachable consumers are the ones who grew up being optimized against. Brand leaders should treat this as a live liability question. The regulatory arc on manipulative platform design has been widening for two years, and "we bought the audience the algorithm assembled" is not a defense that survives discovery.

You bought the monitor; the monitor sells you

Ad-supported hardware is migrating from the living room to the desk, and vowe.net's note on displays that ship with tracking and ad delivery built in marks the arrival point. The economics are the same as smart TVs: panel margins are commodity-thin, so the device is sold near cost and monetized afterward. The difference is context — a work monitor sees documents, dashboards, and customer data. That makes "smart" a procurement risk line item for IT, and it gives any hardware brand willing to sell a dumb display at a premium an actual differentiation story.

Brand & Growth

Efficiency buys you a media plan; it doesn't buy you a difference

The argument surfaced at Beet.TV — that algorithmic buying is commoditizing brands that optimize toward the same metrics — is the predictable end state of everyone running the same optimization loop against the same inventory. If the machine picks the audience, the placement, and the variant, the only remaining variable is what the machine was handed. That's why the first-of-its-kind pee ad is the more useful artifact of the two: a premise no performance test would have surfaced, because performance testing selects against anything with no prior. The same instinct is showing up in Liquid Death aiming its campaign at data-center water consumption: brands are finding that the sharpest available differentiator is a position the optimizer can't generate. Budget implication for CMOs: the efficiency line and the distinctiveness line are not competing for the same dollar anymore, and treating them as one number is how you end up indistinguishable at a very good CPM.

Commerce Rewired

Streamers want the take rate; the subscription model is secondary.

Subscription growth is saturated in every mature streaming market, so the platforms are reaching for the model underneath it: Simon Owens lays out how every streaming company is trying to become a storefront — reselling third-party subscriptions, absorbing video podcasts, adding commerce. Amazon has run this playbook with Channels for years, YouTube is building its own aggregator tier, and Netflix's move into video podcasts is a direct grab at YouTube's most defensible category. The logic is that marketplace revenue and ad inventory scale without proportional content spend, the only line on a streamer's P&L that has ever gone the wrong way. For brands, the consequence is concrete: streaming turns into a retail-media surface with commerce data attached, and the media-buying conversation moves from reach to shelf placement.

Machines & Minds

AWS is selling the boring layer, which is the right business

Robotics demos are cheap and fleets are brutal, and the pitch covered at SiliconANGLE — that the gap between physical AI prototypes and production is where the money is — puts AWS in the position it likes: simulation, data pipelines, fleet management, the unglamorous scaffolding that every operator needs and nobody wants to build. The value loop functions as a live proof of outcomes rather than a demonstration. It also sits against a running problem in enterprise AI, where pilots keep failing to produce measurable results; the vendors selling deployment infrastructure get paid whether or not the pilots work, which is worth remembering when reading their adoption numbers.

The value that built the models came from somewhere

DcB argues that the transfer underneath this build-out was real and traceable, not a rhetorical flourish from people who lost. Treat it as the counterweight to the AWS story: the deployment layer is being commercialized on a foundation whose provenance is still working through courts and licensing negotiations. Enterprises signing multi-year commitments to model providers should be reading the indemnification clauses with the same attention they give the pricing table.

14 articles across 6 themes · 12 sources · Powered by Folo + Claude