The Adjacent Brief

TL;DR: DeepSeek slashed V4-Pro API prices by 75% and cut cache costs to a tenth, undercutting OpenAI, Anthropic, and Google across the board. The Axios finding that AI deployment can already exceed the cost of human workers adds a useful complication: token prices falling doesn't mean total AI costs are falling. Elsewhere: a San Francisco AI boutique, a U.K. government copilot rollout to 28,000 tax staff, and Wall Street repricing software company debt around AI exposure.

Worth Reading

• DeepSeek's 75% price cut is a direct shot at OpenAI, Anthropic, and Google — API commoditization is moving faster than the incumbents' moat-building.
• Total AI deployment costs already exceed human labor for some roles — Cheap tokens and expensive systems aren't contradictions; they're the same story.
• ASML became the world's most critical chokepoint — by betting everything on a single technology — A deep look at how one company's monopoly on EUV lithography was built, not inherited.
• Inside Andon Market: the first retail boutique run by a Claude-based AI agent — The experiment is narrow, but the question it's asking about autonomous retail operations is not.
• HMRC hands 28,000 staff an AI copilot — and found 26 minutes back per workday — One of the largest government AI deployments in Europe, and the productivity math is real.
• How much of Substack is actually AI-generated? More than the bestseller lists suggest — A Pangram analysis of top newsletters finds AI content is widespread, undisclosed, and largely unfiltered.

Machines & Minds

Token prices fall; deployment costs don't

DeepSeek cut V4-Pro API prices by 75% and slashed cache costs to a tenth of their previous rate, undercutting GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro simultaneously. The headline is a price war. The more useful read is what it does to the incumbents' justification for premium pricing. If underlying capability differences are closing and Chinese alternatives are aggressively commoditizing the API layer, enterprise AI vendors have a tighter window to lock in switching costs before procurement departments run the numbers.

The complication: AI deployment can already cost more than human labor when you account for integration, orchestration, validation, and the humans still required to supervise output. Token prices aren't total cost. The systems around the tokens are where the bill runs up. A piece from Every, "You Are the Most Expensive Model," pushes further: the scarce resource in AI deployment isn't compute, it's the human judgment needed to make AI outputs usable. Capital markets are already processing this. A Wall Street Journal analysis of 100 actively traded software company loans finds investors applying sector-wide price pressure as they sort for defensible moats — credit markets are pricing AI exposure before it shows in revenue.

Institutional deployment at scale tests the governance

HMRC put AI copilots in front of 28,000 staff — one of the largest government AI rollouts in Europe — and the headline number is 26 recovered minutes per workday. If that holds across the full population, it's the equivalent of adding thousands of full-time staff without hiring. The policy wrinkle is the U.K.'s simultaneous AI governance bind: as the Financial Times reports, some U.K. officials fear that aligning with EU AI rules to shore up post-Brexit European ties risks fracturing the U.S.-U.K. alliance. The HMRC deployment is a domestic fait accompli; the regulatory framework governing it is still being negotiated in two directions at once.

The demand for identity verification is rising sharply across Washington and Wall Street as deepfake capabilities outpace institutional detection. The Madison Square Garden surveillance case is described by WIRED as the most consequential legal battle of the AI era — not because of the technology, but because the outcome will define what private entities can legally do with facial recognition at scale. These are connected: the harder it becomes to verify who is real, the more valuable the infrastructure that does it.

The thinking that isn't

NYU researchers found that fabricated AI "reasoning" increases user trust in AI outputs even when that reasoning is invented — meaning chains of thought displayed to users function as persuasion tools, not transparency mechanisms. The implication for enterprise AI deployments is uncomfortable: the more legible the AI appears, the more likely that legibility is designed. Separately, AI is demonstrably bad at physics, performing poorly on problems requiring intuitive physical reasoning — a real constraint for any application beyond language and pattern tasks.

The safety discourse is grappling with a different gap. A piece in Transformer argues the AI safety movement needs to recruit beyond experts — that policy will be shaped by whichever side builds a broader public coalition, and right now that's not the safety camp. Policymakers are already debating Universal Basic Income as a structural response to white-collar AI displacement, which suggests the public-facing consequences of AI deployment are moving from speculative to legislative.

Research on which workers use AI most effectively points toward a widening performance gap: high-skill workers who adopt AI are pulling further ahead of both non-adopters and low-skill adopters. A Nate's Substack piece on a sales consultant who automated five hours of weekly work in an afternoon illustrates the individual version of this — the gap is less about access to tools than knowing what to do with them. The Leverage's "The Labs Are Eating Everything" maps the structural version: incumbent AI labs expanding into adjacent product categories faster than the market anticipated.

The New Consumer

The longevity economy has an access problem

Longevity science is outpacing its own distribution: treatments and interventions that exist are not reaching consumers at scale, because the market infrastructure for delivering them affordably hasn't caught up with the clinical evidence. The result is a market that predominantly serves wealthy early adopters while the broader consumer opportunity sits unrealized. The gap between what's available and what's accessible is a business problem that someone will solve — not because the science is speculative, but because the friction is structural.

Platform trust under pressure from two directions

A Pangram analysis covered by User Mag finds AI-generated content is widespread across Substack's bestseller list — much of it undisclosed. For a platform whose entire value proposition rests on direct creator-reader trust, this is a structural problem, not a content moderation nuance. AI thirst trap creators are contesting the framing that their content is deceptive, arguing that audiences understand the artifice. The distinction matters commercially: undisclosed AI content in a paid newsletter is a different contract than openly artificial personas on Instagram.

Privacy as product positioning

Android users are increasingly framing their device choice as a privacy stance, with GrapheneOS adoption growing among users who treat the operating system as a signal of values, not just a utility choice. A technical preference becoming an identity marker tends to precede broader category formation. Windows is using second-chance setup screens to hawk Microsoft services to users who just declined them during installation, a friction pattern that feeds exactly the distrust GrapheneOS is monetizing.

The creator economy's algorithmic machinery

A Forbes profile of Anthony Fujiwara reveals how "clipping" for social media was industrialized — and how that infrastructure was deployed by crypto casino Stake to manufacture viral reach. The story connects the creator economy's growth tools directly to the gambling industry's marketing machine. China is formalizing gig worker protections for online platforms, which will put structural cost pressure on the same algorithmic content economy.

Casey Lewis in After School examines Gen Z's fatigue with millennial-nostalgia-driven consumer culture — a pattern relevant for any brand that built its aesthetic around 2000s-2010s revival. The generational handoff in dominant cultural reference points is happening faster than most brand calendars are designed to track.

Connected World

Every chip runs through one machine

A Works in Progress deep dive traces how ASML became the sole manufacturer of EUV lithography machines — the equipment required for sub-5nm chip production. The story is less about the technology and more about how a bet-the-company decision on a single unproven technique, combined with close collaboration with TSMC and the U.S. government, created a chokepoint no competitor has replicated in two decades. That monopoly is now a geopolitical instrument.

OpenAI is working with MediaTek and Qualcomm on custom smartphone chips, per Ming-Chi Kuo, with Luxshare handling system co-design and mass production targeted for 2028. This is the same vertical integration playbook Apple ran with the A-series — controlling the silicon means controlling the performance envelope and, eventually, the margin structure. If OpenAI reaches mass production with a proprietary chip, the API pricing conversation looks very different.

War hits the supply chain where it's already tight

The Iran conflict is disrupting PCB supply chains through SABIC, adding material cost pressure to hardware manufacturing already navigating trade restrictions and ASML-constrained capacity. Adam Tooze's Chartbook traces the broader pattern of data center expansion triggering community and regulatory blowback — Maine enacted a moratorium on data centers over 20MW — connecting infrastructure appetite with the political economy of where that infrastructure can actually be built. The greenhouse gas emissions from new gas-powered data centers are on track to exceed the annual emissions of entire countries, which adds another layer to the regulatory risk calculus for any company announcing large build-outs.

Brand & Growth

"Resilience" is what CEOs say when they don't know what to say

A New York Times piece finds executives leaning heavily on "resilience" as their all-purpose response to global turmoil — Iran war, inflation, supply chain fragility. The word does real work in earnings calls: it signals stability to investors without committing to a forecast. The risk is that it functions as a placeholder for strategy, and markets are getting better at telling the difference.

Commerce Rewired

The EV value proposition finally arrives at the right price

Long-range EVs under $40,000 are now available in record numbers — a milestone the market has been waiting on since the category launched. The constraint was always price, not technology or charging infrastructure, and that constraint is easing as Chinese competition and scale economics push the entry point down. The policy environment is complicated by the Iran conflict's effect on gas prices, which changes the payback calculation in real time.

AI as store operator

Andon Labs' Claude Sonnet 4.6-based agent running Andon Market — a physical retail boutique in San Francisco — is a narrow experiment, but a clarifying one. The question it asks isn't "can AI manage inventory?" It's whether autonomous AI operations can handle the full surface area of a retail business: supplier relationships, pricing decisions, customer disputes, restocking judgment. The New York Times piece by Heather Knight is worth reading less for what Andon Market has proven and more for the specific places the agent struggled. Those are the gaps the next wave of enterprise AI products will compete to close.

Culture & Signal

The Iran war's second-order costs

Carbon removal companies are reframing their public pitch away from climate benefits and toward energy dominance arguments — a pragmatic adjustment to the Trump era's policy vocabulary. The technology didn't change; the political viability of the previous framing did. Separately, WIRED documents the environmental harms the Iran war is generating — toxic smoke, oil spills, elevated emissions, soil contamination — as costs that don't show up in the strategic calculus but compound over years. Both pieces sit on the same ground: how long-cycle environmental decisions get repriced by short-cycle political events, and who absorbs the cost when they don't get priced at all.

33 articles across 6 themes · 27 sources · Powered by Folo + Claude