The Adjacent Brief

TL;DR: Google's capex guidance sent AI stocks lower as Wall Street started pricing infrastructure spend as a cost rather than a growth story. A Philadelphia suburb handed a data center developer 43 conditions before approving its permit, and Ofcom found Netflix has passed the BBC as UK viewers' first-choice media source.

Worth Reading

• AI systems know your brand cold — and won't say its name — Models recognized 96% of tested brands but named almost none in outputs, which breaks the assumption that brand equity carries into AI search.
• Wall Street stopped reading AI capex as growth and started reading it as cost — The spending didn't change; the interpretation did.
• Compute could get 10x more expensive, and the reason isn't chips — Energy, land, and labor are the inputs nobody hedged.
• Google's own usage data says workers aren't automating themselves away — Behavioral data undercutting the displacement narrative, from the company selling the tools.
• AI companies are recruiting electricians and carpenters by the thousands (paywall) — The bottleneck in frontier AI is a trades shortage.

Brand & Growth

Brand recognition and brand mention are now separate assets

A study of AI search behavior found models recognized 96% of tested brands but named almost none of them in generated answers. The models have the brand knowledge but don't surface it, because naming a specific vendor reads as endorsement and endorsement is a liability. For anyone who spent the last eighteen months buying "AI visibility" services, that's a policy decision—and no amount of SEO will get you past it. The practical read for a CMO: if AI-mediated discovery grows and brand names get sanded out of it, category-level demand capture — owned channels, direct traffic, retail placement — gets more valuable. Visible human authorship has become a distribution strategy that operates independently of sentiment.

A bootstrapped games media outlet outlasted its VC-funded competitors

Simon Owens's newsletter traces how a blog with no outside funding became the most influential media company in gaming culture — consistent editorial voice, an audience that treats the outlet as an arbiter, and no growth targets forcing traffic-chasing. Compare that with the VC-funded games sites that scaled, pivoted to video, and dissolved. The asset was never audience size; it was the audience's willingness to accept a verdict.

Altman's AI-CEO reversal sends a pricing signal rather than a philosophical one

Eight months after floating the idea, Sam Altman now says nobody actually wants an AI CEO. Read it against today's market reaction to capex: when investors are rewarding demonstrated revenue and punishing speculative burn, maximalist framing stops being free marketing and starts being a liability in earnings calls. Executive rhetoric tracks the cost of capital.

Connected World

The constraint moved from chips to concrete, copper, and consent

Dwarkesh Patel's argument that compute could get 10x more expensive rests on inputs that don't respond to Moore's Law: power generation, interconnect queues, land, and skilled labor. Semianalysis documents the industry's response in the shift to modular, LEGO-style datacenter construction, and the local economies around these sites are monetizing the construction crews six-pack by six-pack (paywall). One adjacent footnote on the hardware side: The Register's requiem for Intel's Optane makes the case that a killed memory tier would have relieved today's KV-cache RAM crunch, a reminder that the cost curve is partly a product of decisions made a decade ago.

Municipal permitting is where compute economics actually get set

A Philadelphia suburb approved a data center on the condition that the developer meet 43 separate demands — noise caps, water commitments, road work, decommissioning terms. That's pricing power, and the township is using it. For anyone underwriting capacity, the relevant number is how many townships reach the same conclusion over the next four quarters and what the concession stack does to project IRR. Local government is becoming an input cost line.

The New Consumer

Attention grew, the money didn't follow

Metricool data shows YouTube long-form views rising while estimated ad revenue and average view duration both fell — more starts, shallower watch, lower yield per view. That decoupling is the operational problem for any brand that budgets against impressions: the unit you're buying is getting cheaper because it's getting worse. Prioritize completion rates and repeat visits over reach.

Substack built delivery infrastructure rather than a media platform

A widely-shared post argues Substack collapsing not for lack of writers but for lack of a platform. It absorbed tens of thousands of displaced journalists and never built the discovery, monetization diversity, or institutional structure that would make those careers durable. It's the same failure mode as the creator economy generally: 57% of Gen Z want this as a job, and a rounding error earns a living from it. The Prof G Pod's account of how the creator economy took over media is the optimistic version of the same facts. Both are true; they describe different ends of a power-law distribution.

Netflix passing the BBC reflects a shift in distribution; audience taste has not changed.

Ofcom research puts Netflix ahead of the BBC as UK viewers' first-choice media source. The BBC still makes the programming that defines British cultural conversation. What it lost is the default position on the home screen. For public-service media everywhere, the defining question is whether you own the interface where the session starts.

Culture & Signal

Prediction markets are pricing clinical trials, which is a problem for the trials

Doctors and researchers told the New York Times that FDA-related wagers on Kalshi and Polymarket could compromise drug development — trial participants, investigators, and staff now hold financial positions on outcomes they can influence or leak. Prediction markets have spent two years selling themselves as superior information aggregation. Here the aggregation mechanism creates the incentive to corrupt the input. Anyone with a regulatory-outcome-dependent business should read this as a preview of how contract markets will interact with their own disclosure timelines.

Wiping your phone is becoming the crime

404 Media interviewed a man charged with obstruction for allegedly triggering a duress wipe on his GrapheneOS phone, who says prosecutors want a precedent. The design question for anyone shipping privacy features: a feature that only functions when authorities are present converts your engineering choice into your user's legal exposure. Vendors selling security-by-default to consumers should assume the courtroom, not the threat model, is the venue where the feature gets evaluated.

Machines & Minds

Labor arbitrage was the BPO product, and automation is repricing it

Forrester's read on the outsourcing sector is that AI is dismantling the cost-per-seat logic BPO was built on. Clients can now benchmark what the work should cost without a wage-differential markup. That squeezes intermediaries whose value was access to cheaper hours and rewards those who own process knowledge and compliance. If you buy BPO services, this is the renegotiation window.

Quality control is the AI deliverable's actual bottleneck

GPTZero found apparent hallucinations in four PwC Middle East reports (paywall), following earlier findings that pushed EY and KPMG to retract work. Consulting is the purest test case: the product is a document, the buyer pays for judgment, and a fabricated citation is a total loss of the deliverable's value. Detection is now cheap and third-party, which means reputational risk sits with the firm rather than the client's willingness to check. On the model side, WIRED's testing found frontier systems from Google, Anthropic, OpenAI, and xAI have safeguards that simple attacks still walk through inconsistently — worse than uniformly weak, because it makes vendor risk impossible to spec.

Adoption is slower than the marketing and slower than the fear

Google's own usage data suggests workers aren't automating themselves out of jobs at anything like the rate the discourse assumes: behavioral evidence from a vendor with no incentive to undersell. Morning Brew's roundup of corporate posting-through-it moment captures the gap from the other side: the public conversation about AI's labor effects is running well ahead of observable deployment. For workforce planners, the useful posture is to plan against measured tool usage inside your own org rather than the sector narrative.

Commerce Rewired

Capex stopped reading as growth and started reading as cost

Google's spending guidance sent AI-linked stocks down, and the number itself wasn't the news. The Verge's account of AI finally getting expensive enough to make Wall Street nervous describes investors applying a different multiple to the same disclosure they applauded two quarters ago. That reframing has a direct consequence for anyone selling into or raising against AI budgets: the burn-to-growth story gets a harder audience this earnings season, already visible in Chinese model providers moving off free tiers and in tools like Cursor pricing regionally at $7/month on cheaper homegrown models rather than frontier inference. Monetization discipline arrives when capital markets ask for it.

Robots need worlds before they need bodies

a16z argues that simulation environments are the prerequisite for robotics progress, not an accessory to it. Training data for embodied systems has to be manufactured, because the physical world doesn't produce it fast enough. Read alongside the compute cost story, that's a bet on more inference demand, not less, and on a category of infrastructure spend that doesn't appear in anyone's data center narrative yet.

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