The Adjacent Brief

TL;DR: Anthropic's compute commitments — a $517 billion shopping list stretching across the back half of the decade — anchored Exponential View's Monday data roundup, alongside the finding that roughly a third of companies discussing AI on earnings calls report measurable revenue from it. Also covered: law firms and investment managers deploying agents that run for weeks unattended, Australia's recording industry barring wholly AI-generated songs from its charts, and Wall Street extending credit facilities to ticket brokers.

Worth Reading

• Anthropic's $517B compute bill is the real forecast — everything else is commentary — Purchase commitments are a better read on where a lab thinks demand is going than anything it says on stage.
• Law firms are buying agents that own the case, not agents that summarize the file — Supio's long-horizon agents put the billable-hour model and the liability question on the same table.
• AI is at its 1997 gigabit moment — the demand doesn't exist yet because the use doesn't — Nate's Substack on why "good enough" local AI may matter more than frontier benchmarks.
• Post-quantum migration is a five-year program most CISOs haven't started — Q-Day planning is a procurement problem long before it's a cryptography problem.
• The governing backdrop to every AI rule that won't get written this year — Heather Cox Richardson's September 13 letter, useful context for anyone modeling US regulatory capacity.

Brand & Growth

"Awareness" became the catch-all term for the screen nobody could measure

Beet's conversation with WPP Media's Adam Shlachter has advertisers treating connected TV as a direct-response channel measured on attribution rather than reach and frequency. The honest reading is less triumphant than the framing: CTV budgets got branding treatment because the inventory was unmeasurable, and the moment identity resolution improved, the money behaved the way digital money always behaves. That's a warning as much as an opportunity. Display spent a decade over-crediting retargeting and under-funding demand creation, and the same measurement stack is being ported onto the living room screen. Brand leaders shifting CTV into performance budgets should decide now which outcomes they'll refuse to optimize against, because the attribution vendor will happily make the case for the cheapest conversions available.

The supply side is moving toward the same logic. Shlachter's other point — that creators are assuming showrunner responsibilities inside television production — matters because creators already operate on closed-loop economics. They know their cost per view, their conversion behavior, and their audience's tolerance for integration, and they bring that instrumentation into rooms that historically ran on gut and upfront commitments. For a CMO, the practical consequence is that the creator line item and the TV line item are buying increasingly similar things from increasingly similar people, and keeping them in separate budgets is now just an org chart artifact.

Compute commitments are the only forward guidance worth reading

The $517 billion figure in Exponential View's Monday numbers — Anthropic's projected compute spend — is a demand forecast disguised as a procurement line. Nobody signs multi-year capacity contracts of that size on consumer chat volume; that's a bet on agent workloads, which consume dramatically more compute per task than a chatbot query and which show up in this brief on both the legal and investing side. Worth holding next to the other number in that roundup: roughly one in three companies mentioning AI on earnings calls claiming measurable revenue impact. That ratio, not the capex, is the thing to track quarterly. It's also the tension in Dario Amodei's recent argument for a slower industry pace: the deliberate-pacing essay and the half-trillion-dollar supply lock-in are the same company in the same month, and the second one costs money.

Connected World

Where inference runs is a margin decision wearing a privacy costume

SemiAnalysis lays out the physical constraint plainly: frontier models are too large to carry in your pocket, and memory bandwidth, not clever quantization, decides the split between on-device and datacenter execution. Every consumer AI feature marketed on privacy grounds is a routing decision made on cost and latency, with privacy as the acceptable public explanation. The value is accruing to whoever owns the router — the layer that decides which query stays local and which gets shipped to a GPU cluster — because that layer sees all the traffic and captures all the cost savings. Product leads shipping "private, on-device AI" should be able to answer what percentage of requests actually stay on the handset. Most can't.

The same pressure on device real estate shows up in a far smaller form. Yanko Design's coverage of Verso, a phone case that moves note-taking to the back of the device, is a concept, not a market, and concepts rarely survive contact with accessory margins. But the itch is legitimate: the front screen is the most contested surface in consumer software, and low-stakes capture keeps getting crowded out by everything fighting for attention on it.

Fusion found a customer, and the customer will change the product

Fusion startups are sourcing deals through defense contractors rather than venture syndicates, a meaningful de-risking event for a sector with no revenue and a twenty-year credibility problem. It also redirects the roadmap. Defense buyers want compact, transportable, forward-deployable power for bases and installations, not gigawatt baseload for a regional grid. Those are different machines with different tolerances and different regulatory paths. The data center power crunch gets invoked as fusion's commercial destination in nearly every pitch deck, but the first paying customer is the one who shapes the engineering, and that customer wears a uniform. Anyone modeling fusion into a 2035 grid decarbonization scenario should note which specifications are actually being funded.

The New Consumer

The thing you bought keeps being governed by the company that sold it

Ars Technica's reporter actually used John Deere's self-repair diagnostic tool and found the mechanism works, then found that farmers won't touch it, because using it raises questions about warranty coverage and future repair eligibility. Shipping the capability while leaving the penalty structure intact produces a compliance artifact that nobody actually adopts. Right-to-repair concessions prove this: the tool exists to satisfy regulators and headlines, never to become a product people use. The terms surrounding it tell customers what the company actually wants them to do, and customers read terms better than press releases.

The automotive version is less negotiated. The Verge's column on cars monetizing location, behavior, and cabin data describes collection that most owners never meaningfully consented to and can't inspect, let alone switch off. Deere and the connected-car industry are running the same play from opposite directions: one restricting what owners can do with hardware they own, the other extracting value from data they generate without asking. For brands in any durable-goods category, the post-purchase relationship is where trust is actually won or lost, and it's the part of the customer journey most companies have delegated to legal.

Blaming screens is cheaper than fixing schools

Marginal Revolution's look at the PISA decline questions whether screens are the culprit at all — the timing and demographic distribution of the score drops don't line up cleanly with smartphone adoption curves, which points back toward curriculum and instructional quality. The screens story persists because it identifies a villain outside the education system and suggests an intervention that costs a district roughly nothing. Phone bans are politically legible and cheap; pedagogy reform is neither. Anyone selling into K-12 should expect the device-restriction budget to stay well-funded and the evidence base to stay thin.

Machines & Minds

The pitch moved from answering questions to owning the file

Supio is selling law firms agents that run multi-week matters autonomously, delivering sustained work across a case timeline. The parallel arrived the same day from finance, where investment firms are described moving from research assistants toward agents that participate in decisions. Both categories share a structural problem their vendors handle gently: these are professions where liability attaches to a named human, and no court has settled what happens when the agent's work product is wrong. A New Mexico attorney was held in contempt this month over unverified AI-generated citations. That's the small version of a question that gets expensive fast when the agent has been running unsupervised for three weeks. Firms buying this should be negotiating indemnification rather than dwelling on feature lists.

Output quality and expertise formation are different variables

The research Tyler Cowen flags on whether AI assistance builds expertise or quietly erodes it is the missing input in every agent deployment plan. A firm can improve output quality this quarter and degrade its bench in the process, and the two effects appear on different time horizons with different measurement difficulty. The long-horizon agent story runs on the same logic: the person reviewing a three-week autonomous workstream needs judgment that was historically built by doing three-week workstreams manually. Leaders deploying agents into junior-heavy functions should be explicit about where apprenticeship now happens, because the default answer is that it doesn't.

The criminal application has better unit economics than the defense

Persuasion's piece on coordinated AI agents running crime as a business describes the asymmetry clearly: fraud scales with cheap inference, while detection and enforcement scale with headcount and jurisdiction. Attack capability is priced per token; defense is priced per investigator. Fraud teams at banks, marketplaces, and payment processors are the first to feel this, and the operating assumption should be that any verification step a human can pass, a swarm can pass more cheaply, which pushes the whole industry toward provenance and identity infrastructure rather than behavioral heuristics.

Culture & Signal

Gatekeeping is how "human-made" becomes worth money

Australia's recording industry barred wholly AI-generated songs from charts and radio airplay, a defense of the chart itself. A chart is only valuable as a filter; flood it with zero-marginal-cost output and the signal it carries goes to zero, taking the licensing and promotional economy built on it along too. The enforcement problem is obvious and unsolved: "wholly" is doing enormous work in a production chain where AI touches stems, mastering, and vocal correction as a matter of routine. Expect the fight to move quickly from prohibition to provenance documentation, which favors labels with paperwork infrastructure over independent artists without it. The same machinery is visible in Instagram relabeling synthetic accounts and in the detection of thousands of self-identifying AI agent posts: verification is becoming an industry, and the standards are being set by whoever moves first.

Capability disclosure is the biosecurity policy; everything else is rhetoric

The New York Times examines whether biology-trained models measurably raise bioweapons risk (paywall), and the useful part of the piece is the gap between what labs can evaluate internally and what any external body can verify. Frontier labs have spent recent weeks calling for slower development. Amodei, Altman, Musk, and Hassabis have all said some version of it publicly. Voluntary pacing statements cost nothing and bind no one. Third-party access to capability evaluations for bio-relevant models costs something real, which is precisely why it's the thing to watch for. Until an outside party can run the tests, "we evaluated it and it's fine" is a press release.

Commerce Rewired

When arbitrage gets a credit facility, the markup becomes permanent

Wall Street institutions are extending capital to ticket resale operations as a defined investment category, which changes the economics of scalping in a way enforcement never has. A broker with a credit line can bid on inventory at a scale and risk tolerance a hobbyist can't match, and that capital demands returns, which means the spread between face value and resale price stops being an inefficiency and becomes the yield on a financed asset. Artists, teams, and venues underpricing tickets at primary sale are effectively transferring margin onto a broker's balance sheet, and the more sophisticated the financing, the harder it gets to recapture through dynamic pricing or fan verification. The parallel to watch is agentic commerce: the same logic applies anywhere fast, well-capitalized automated buyers meet manually priced scarce inventory.

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