The Adjacent Brief

TL;DR: Data center developers are hiring acoustic consultants to survey noise emissions before neighbors do it for them. SpaceX is laying groundwork for a turbine blade foundry in Bastrop, Texas, to relieve its own power constraints. In music, producers are running forensic audio investigations on tracks they suspect were made with Suno while major labels litigate against and sign deals with the same generation platforms.

Worth Reading

• SpaceX would rather build its own turbines than wait in the grid queue — Job listings point to a Bastrop foundry making blades for large gas turbines. Vertical integration as a response to a sold-out supply chain.
• The discount compute tier skips the security basics you assumed were included — SemiAnalysis on why cheap GPU capacity is cheap.
• Nvidia's moat moved from the chip to the room it sits in — Competing on rack-and-datacenter architecture is much harder to price-shop than competing on a processor.
• Meta's data centers look like a cost center until you price them as a compute business — The argument that idle internal capacity is an unmonetized external product.
• Asked 100 companies for his data, some deleted it instead of handing it over — Deletion is the cheapest compliance path when your data architecture can't answer the question.
• 89 of Douyin's top 100 animated dramas were AI productions — the microdrama boom is already synthetic (paywall) — DataEye's May count, via the FT.
• Samsung's privacy-first smartglasses still walk into a legal minefield — On-device processing doesn't help you with two-party consent statutes.

Connected World

Noise is the permitting fight nobody underwrote

Data center developers and the communities next to them are both retaining acousticians to measure what the chillers and transformers actually emit at the property line (paywall), Bloomberg reports. Noise is a useful proxy because it is measurable, locally regulated, and doesn't require anyone to litigate climate policy — a decibel reading at 3am is admissible in a county zoning hearing in a way that a water-usage projection isn't. That matters more given the EPA's recent proposal to let data centers shield pollution disclosures: as federal environmental data gets harder to obtain, enforceable constraints migrate to whatever local ordinance still has teeth. Anyone modeling site acquisition should price acoustic mitigation and community relations into the pro forma rather than deferring them to the contingency budget.

SpaceX would rather own the turbine supply chain than queue for it

Job listings and permitting activity indicate SpaceX is building a foundry in Bastrop, Texas, to cast blades for large gas turbines. The AI power crunch framing is the obvious read. The heavy turbine order books at GE Vernova, Siemens Energy, and Mitsubishi are booked years out, which makes on-site generation a manufacturing problem rather than a procurement problem. A rocket company already casts high-temperature superalloy parts at volume; the adjacency is real. Watch whether other compute buyers follow the same logic — capital going into blade casting says more about grid pessimism than any press release about a power purchase agreement.

Cheap capacity is cheap because the security line item was removed

SemiAnalysis's audit of the neocloud tier finds operators missing controls that enterprise buyers assume come standard — exposed management interfaces, weak tenant isolation, inconsistent key handling. The relevant question for a buyer is whether the delta between neocloud pricing and hyperscaler list covers the compliance work you now have to do yourself. If you're running training jobs on proprietary data, the neocloud discount transfers security spend from the vendor's balance sheet to yours, and that trade is not obviously favorable.

Culture & Signal

Verification is becoming an unpaid job in music, done by the people with the most to lose

Producers in the EDM scene are doing forensic work on tracks they suspect were generated with Suno — comparing stems, catalog velocity, and artist histories to expose acts they believe are AI fronts, as The Verge documents in the H4RRIS and Nihil Young case. Note who's absent from that process: the distributors and DSPs that ingested the uploads. Detection cost is being absorbed by working musicians whose royalties are diluted by the flood, which is not a stable arrangement. The same adverse-selection dynamic showing up in natural diamonds — where a superior stone now triggers more suspicion, not less — applies here: a prolific, polished release schedule reads as evidence against you.

Labels suing and licensing the same companies is price discovery

The music industry's simultaneous litigation and partnership posture toward AI generation platforms, laid out in the NYT DealBook piece on whether this is thievery or innovation (paywall), follows the Napster-to-YouTube playbook almost exactly. Lawsuits establish that a license is required; the license terms get negotiated afterward, from a stronger position. The endpoint is licensed catalogs with per-use attribution, which is what Content ID became. If you're a brand or agency licensing music, demand provenance documentation in the contract now, because the indemnity language on this is about to get much more expensive.

Microdramas are a distribution economics story that celebrities are only the surface of

Recognizable Hollywood names are taking roles in vertical microdrama apps, TechCrunch reports — a format with 90-second episodes, pay-per-unlock monetization, and production cycles measured in days. Talent is following unit economics. A microdrama recoups on a few thousand paying viewers; a streaming pilot needs a platform to say yes. The pressure comes from the other side of the format: per DataEye, 89 of the top 100 animated dramas on Douyin in May were AI-generated. Celebrity attachment is what a human-made microdrama can offer that a synthetic one can't yet, which makes these deals a hedge with a shelf life.

The New Consumer

The AI enthusiasm gap runs backwards from the digital-native assumption

Glassdoor's analysis of what employees write voluntarily about their employers finds 47% of Gen X workers describing company AI use positively, against 40% of millennials and 33% of Gen Z (paywall). This is closer to behavioral data than a survey — nobody prompted these reviews — and it inverts the standard adoption narrative. The likely explanation is positional: mid-career workers have status meetings, summaries, and decks to hand off, while early-career workers are watching the exact tasks that constituted the first three rungs of their ladder get automated. If your internal AI enablement program targets "digital natives" as early adopters, it's aimed at the cohort with the least to gain. The retention exposure sits in the 0-5 year band.

Personal recommendation is filling in where search stopped working

The pull of a newsletter whose entire premise is links a friend would have sent you is a function of what happened to the alternatives. Travel planning through general search and social is now reliably degraded by SEO farming and sponsored placement, and the pattern repeats across product research, restaurants, and health queries. What's replacing it is a named individual with taste and a track record. For brands, this reframes the earned-media question: placement inside someone's personal forward carries the credibility that ranked results above it have spent a decade shedding, and it can't be bought at rate card.

Brand & Growth

"Forward-deployed" is what implementation consulting is called when you need engineers to take the job

Startups are hiring for forward-deployed roles that involve sitting inside customer organizations and making the AI product actually work (paywall), the NYT reports — a Palantir coinage now spreading across the AI vendor landscape. The title arbitrage is real: engineers who'd decline a customer-success role will take a forward-deployed one at engineer compensation. The rebrand also inverts the tidy version of the AI thesis. If the value of an AI product depends on months of on-site configuration against a customer's specific data and workflows, then the AI era is generating services headcount, and services headcount is what separates a 75% gross margin from a 45% one.

For buyers, this is a useful diligence question: ask how many forward-deployed people come with the contract. A high number means you're purchasing a consulting engagement with a software wrapper, which may be exactly what you need, but you should price it that way, and you should ask what happens to the deployment when those people rotate to the next account. Deloitte's finding that roughly one in five enterprises has mature AI governance suggests most buyers aren't yet asking. The vendors know it.

10 articles across 4 themes · 7 sources · Powered by Folo + Claude