The Adjacent Brief

TL;DR: Microsoft pulled out of its deal to anchor Nscale's planned Essex data center, where 23,000 Nvidia chips may not get grid power until the 2030s. AI infrastructure debt sold off further: CoreWeave's bond yield passed 15%, insuring SoftBank debt got more expensive, and some Oracle debt traded at 80 to 90 cents on the dollar. Separately, the Financial Times reports that Anthropic is consulting law firms as it builds agents to negotiate deals for users.

Worth Reading

• Anthropic hires lawyers to teach its agents to bargain, so the agent is the product and the chatbot is the interface (paywall)
• Bernstein finds only about 35% of the planned US data center pipeline is credible (paywall)
• CoreWeave's real counterparty risk is Nvidia, not its lenders
• A cheap motor controller could free up gigawatts of grid capacity
• Boomers got 38% richer in three years while under-35s lost ground
• Polymarket recorded $600M in election trades after the results were known
• Meta's new commercial terms may let it skip millions in AI licensing fees

Commerce Rewired

Agents are spending money before anyone agrees on the rules

Stripe's Link data shows AI assistants from competing vendors already completing real purchases through one shared payment rail. Stripe sits between every agent and every merchant, and whoever holds the stored credential and the fraud model decides which agent transactions clear. For a retailer, the useful question is whether your checkout can be read and trusted by an agent you have never heard of. Building your own agent matters less.

The AI bill is moving from the income statement to the bond market

CoreWeave's bond yield passed 15%, the cost of insuring SoftBank debt rose, and some Oracle paper changed hands at 80 to 90 cents on the dollar. These are credit markets, not equity investors reacting to a demo, and they are pricing a real chance that the leveraged layer of AI infrastructure doesn't get paid back on schedule.

Microsoft is pushing from the other side. Its effort to make AI cheaper to run raises the question of whether investors ever see the savings or whether they get competed away in pricing. The two stories are linked. If hyperscalers succeed in cutting cost per token, the GPU landlords that borrowed against today's scarcity pricing are the ones exposed. Cheaper inference helps buyers of AI and hurts the people who financed the last generation of capacity.

Connected World

Power, not chips, sets the build-out schedule

Microsoft walked away from anchoring Nscale's Essex data center, a commitment to rent 23,000 Nvidia chips at a site that may not reach the grid until the 2030s. A delay like that sinks a plan no matter how much capital is behind it. The Financial Times widens the picture: America's $10tn infrastructure boom is starting to fray (paywall) as projects run into permits, local opposition and interconnection queues.

Headline capex announcements are a weak guide to future capacity. Energized megawatts are the number to watch. Treat any compute roadmap, including your vendor's, as conditional on a utility's timeline.

Quantum's progress is in the plumbing

Researchers ran the first successful hardware test of a magic-state cleanup method that cuts errors 40-fold. Magic states are the resource that lets a quantum machine do more than a classical computer could simulate, and purifying them cheaply is one of the expensive steps on the way to fault tolerance. It is not a product, and nobody should rewrite a roadmap because of it. Engineering results like this shorten timelines more than qubit-count press releases do.

Culture & Signal

Surveillance arrives through the ad stack

Cory Doctorow details how an Israeli ad-tech company served as a delivery system for mercenary spyware. Programmatic advertising was built to put arbitrary code in front of precisely targeted people, which also describes an exploit pipeline. Brands buying through opaque supply chains should ask what else is riding in the same auction.

WIRED's column on how encryption is being squeezed from all sides covers the policy side: lawful-access demands, client-side scanning proposals and platform pressure on end-to-end messaging. Read together, the two pieces describe a state seeking a way in through policy and private vendors selling a way in through ads, while users are told the lock still works.

Liability is looking for a name

Giorgos Tsetis argues in Fortune that AI executives should face personal legal liability, the way Sarbanes-Oxley made CEOs sign their financials. That is unlikely to pass soon. The argument still matters because it moves the target from models to signatures. A boardroom where a named person signs off on AI deployment decisions would ship differently.

The New Consumer

Screens keep getting cheaper, so people keep buying more of them

Marginal Revolution's point is plain economics: quality-adjusted screen time is about 39% cheaper than in 2000, and demand followed price. Much of the moral panic about screen addiction attributes to design what price explains. For anyone competing for attention, the rival gets cheaper every year. Matching it on price is a losing strategy, so compete on something else.

Platforms consolidate your identity without asking

Google has started linking phone numbers to accounts users deliberately kept separate, with no explicit opt-in. A work persona, a side project and a pseudonymous account all collapse into one identity graph that is more useful to Google than to the user. Platforms keep turning user data into infrastructure for themselves.

Housing demand splits by balance sheet

Some buyers are totally unfazed by 7% mortgage rates, because rates are mostly a problem for people who need to borrow. A market that clears at 7% is selling to buyers with cash and equity. "The consumer" doesn't describe one group facing one set of conditions. Segment by assets; age and income are poor dividing lines.

Machines & Minds

Negotiating agents are where autonomy becomes a legal product

Anthropic is consulting law firms on negotiation tactics (paywall) as it develops agents that bargain on users' behalf, including against other AI agents. The decision to bring in lawyers is telling. Negotiation involves anchoring, concession strategy and authority limits, and an agent that misreads its mandate can bind a principal to a bad deal. The value is a repeatable task that enterprises already pay humans to do, a stronger case than another chat interface.

Agents trust each other too easily

Nate's Substack breaks down the Hugging Face incident, where one agent treated a peer agent's go-ahead as authorization. Set beside Anthropic's negotiators, it shows the design problem: multi-agent systems need a way to verify who is allowed to approve what, and most don't have one yet. Any team deploying agent swarms should audit where one model's output turns into another model's permission.

Training data comes from the customer base

Meta has granted itself the right to scrape business customers' content for model training. The advertisers paying Meta are now also supplying its data. CMOs should read the new commercial terms before the next creative upload, since reading them afterward leaves no room to adjust.

Scientific AI is earning its keep

MIT researchers are giving AI a sharper view of crystal structures in materials. This narrow, domain-grounded model work feeds directly into R&D pipelines for batteries, semiconductors and catalysts. It has less spectacle than a frontier model launch and a clearer line to revenue.

Brand & Growth

Premium moves to what airlines can actually change

Airlines can't retrofit fleets fast enough, so they're upgrading the menu instead of the plane, using celebrity-chef partnerships and better catering. Food is the cheapest part of the premium experience to change and the easiest for a passenger to notice. Other brands stuck with long capital cycles can borrow the approach: when the hardware is frozen, differentiate on the parts customers touch and photograph.

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