The Adjacent Brief

TL;DR: Hollywood studios including A24, Focus Features, and Warner Bros.' Clockwork refused OpenAI licensing deals for training data, even as directors like Luca Guadagnino signed individual agreements — a split that puts the industry's relationship with AI on public display. Separately, the NSA lost access to Anthropic's Mythos 5 model mid-red-team after a dispute between the organizations, and a fake skill in AI agent marketplaces bypassed every major security scanner and reportedly reached 26,000 agents.

Worth Reading

• Your home battery is now an AI data center power source (paywall) — Tesla, Sunrun, and Renew Home are aggregating residential solar and storage to feed data center demand; the grid is becoming a two-way market faster than utilities planned for.
• Hsinchu's real estate boom is a chip-economy story (paywall) — Luxury development in Taiwan's semiconductor capital shows how AI infrastructure spending concentrates wealth geographically before it diffuses anywhere else.
• Cityside's nonprofit network model for local news — Shared infrastructure across publications cuts per-outlet costs; the most durable local news model may be cooperative, not competitive.
• What AI is actually doing to how students work — The paper assignment is disappearing from practice while institutions haven't updated their assessment methods; the gap is widening.
• Ben Thompson on memory chips, China, and Microsoft's model choices — The strategic logic behind Microsoft selectively deploying Chinese models alongside US ones, and what HBM dominance actually means for the chip stack.
• SK Hynix overtook Samsung by betting 14 years on a niche memory format — HBM was a long-cycle wager that looked marginal until AI made it the only chip that mattered.
• EVs beat combustion on emissions across every grid mix studied — The "dirty grid cancels EV benefits" argument doesn't survive contact with the data; the crossover point is lower than the skeptics claim.

Brand & Growth

The talent war is a product strategy war

Meta is offering nine-figure compensation packages to recruit external AI talent, including executives from Scale AI, according to Newcomer's reporting on this summer's foundation model talent moves. Meta's bet reveals where the company thinks the real leverage lies: in the humans who know how to build the next layer of the stack, ahead of compute and data. When a company that already has massive infrastructure is paying that much to bring in outside talent, it's betting that its current team can't get there alone.

That bet has organizational consequences. Meta's broader restructuring of product management — flattening hierarchies and collapsing the gap between researchers and product teams — reads as preparation for a faster-moving org, one where the distance between a model capability and a shipped feature needs to shrink. The PM function is being rewired around AI delivery cycles rather than traditional roadmaps.

Productivity gains have a creative cost that isn't being accounted for

Forrester's analysis of AI productivity and creativity makes the case that efficiency gains from AI tools are coming at the expense of the exploratory, divergent thinking that produces genuinely novel work. This is a business risk for any brand that competes on differentiation rather than execution. If every team is using the same tools to go faster, the outputs converge. Speed without variance is a commodity trap.

Connected World

The HBM bet paid off; the question is what comes next

SK Hynix passing Samsung to become South Korea's most valuable company is the kind of outcome that looks inevitable in retrospect and was genuinely contested for most of the past decade. HBM — high-bandwidth memory — was a niche format that SK Hynix committed to while Samsung hedged. AI training at scale made HBM the binding constraint in every major cluster build, and the payoff arrived fast once demand exploded. The Reuters account of those 14 years of bets is worth reading closely: it's a case study in the value of conviction in a long-cycle hardware market where being early looks indistinguishable from being wrong.

South Korea is now moving to consolidate that advantage. The government is in active talks with Samsung and SK Hynix over a second national chip cluster — a direct response to the CHIPS Act-era playbook that the US, EU, Japan, and India have all deployed. Industrial policy in semiconductors is no longer a US-led story; it's a simultaneous race across five or six jurisdictions, each trying to capture the next node of the stack before the others do.

Culture & Signal

Hollywood's AI split is institutional, not generational

The divide opening up in the film industry runs between individuals and institutions. Luca Guadagnino signed an agreement with OpenAI; Warner Bros., A24, Focus Features, and Netflix refused. The studios that said no are protecting IP libraries that represent decades of compounding asset value. The directors who said yes are trading on their individual creative identity, not on a catalog. Those are structurally different calculations, and neither side is obviously wrong.

What's notable is the speed at which OpenAI is moving to secure these agreements. The approach — individual deals that create optionality and public legitimacy before institutional licensing is resolved — is a familiar playbook. Lock in enough visible names to shift the cultural frame, then negotiate from a stronger position with the holdouts.

Energy curriculum is a lobbying instrument

The gas industry has been funding elementary school science curriculum materials that present industry-favorable content to K-12 students, according to Heated's investigation. The mechanism — an education nonprofit acting as intermediary — creates enough distance that the sponsorship isn't visible at the classroom level. This pattern has been documented in tobacco and pharmaceuticals; its presence in climate education for children is the new element. Bill McKibben's piece on free electricity as a climate adoption lever sits in useful contrast: Australia is giving EV owners three free hours of afternoon electricity daily to drive adoption. The argument for clean energy is becoming a price argument rather than a values argument — which is harder to counter with curriculum.

The New Consumer

AI agents are spending your attention budget without asking

The premise of a piece from Ownersnotrenters is worth sitting with: the author stopped actively browsing the web, but their AI agent kept going. The web is still being consumed — just not by the human nominally in charge of the account. This is a structural change to how traffic, intent, and conversion will be attributed. If agents are doing the browsing, the relationship between a brand and its "audience" is mediated by a layer that has its own priorities and search patterns.

Those patterns are also largely invisible to web analytics. Search Engine Journal's reporting on ChatGPT's hidden web searches confirms that ChatGPT and Gemini conduct searches that don't show up in standard analytics or search console data — a blind spot that grows every time someone uses an AI to research a purchase or compare options. The traffic exists; it just can't be seen by the brand being researched.

Being findable by AI is now a distribution channel

Similarweb's data, reported by Search Engine Journal, puts a number on what this means in practice: brands recommended by AI systems saw 2.5x more site visits than those that weren't. That's not a marginal uplift — it's the difference between a functional and a broken top-of-funnel. The brands winning in this environment are the ones that AI systems have enough structured, credible information about to confidently recommend—regardless of SEO strength or ad budget size. The optimization surface has changed. Whether most brand teams have noticed is a separate question.

Machines & Minds

The NSA-Anthropic dispute is the most important AI governance story of the month

The New York Times reports that the NSA was actively red-teaming Anthropic's Mythos 5 before losing access following a dispute between the two organizations — and that the tests showed the model identifying flaws in classified systems. Each element here is significant on its own: a frontier AI model was finding vulnerabilities in classified US government infrastructure, the agency testing it lost access mid-evaluation, and the terms of the dispute that caused the cutoff remain unknown. Together they describe a situation where the capability outpaced the institutional arrangement designed to evaluate it, and where the evaluation was terminated before it concluded.

This connects to a dynamic visible across the enterprise AI market: labs are restricting access to their most powerful models based on customer tier and use case, creating situations where the people with the most legitimate reason to stress-test a system may not have the access to do it. Forrester's analysis of AI orchestration in banking describes a parallel dynamic — financial institutions deploying agentic systems as infrastructure without consistent frameworks for what those systems are authorized to do.

Agent marketplaces have a security problem that scanners can't solve

A security firm called AIR created a fake malicious skill, submitted it to AI agent marketplaces, and watched it pass every automated security scanner before reaching an estimated 26,000 agents. The mechanism of failure matters: the scanners are checking for known patterns of malicious code, but agent skills are a new enough format that the threat signatures haven't been written yet. This is the same dynamic that plagued early mobile app stores — except the blast radius of a compromised agent is larger than a compromised app because agents have broader system access by design.

WIRED's reporting on how China's AI researchers are responding to the current moment adds another dimension: the people closest to these systems, on both sides of the geopolitical divide, are anxious about where the capability curve is going. That's not a reassuring data point when the security infrastructure for the deployment layer is still catching up.

Commerce Rewired

Sanctions are creating a black market with its own price discovery

Nvidia's flagship DGX B300 server has reached $1.1 million on China's black market — more than double its pre-crackdown price — according to reporting in the Financial Times (paywall). Export controls have a well-documented history of creating secondary markets rather than eliminating demand, and AI chips are following the same pattern. The doubling of black market prices tells you two things: demand from Chinese buyers is inelastic at current levels, and the enforcement perimeter is porous enough to sustain a functioning gray market. Neither of those is the intended outcome of the policy.

The sanctions aren't failing so much as working slowly enough to give Chinese buyers time to stockpile at high cost while domestic alternatives develop. The price premium on smuggled Nvidia hardware is essentially a forced subsidy for Huawei's accelerator program. Every $1.1M black market server purchase is an argument for accelerating domestic alternatives.

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