The Landlords Moved In

Here's what I took from the week, and the questions I'm still sitting with.

The ballot box reached the build-out

The midterms turned data centers into electoral politics. Candidates promising to pause or restrict data center construction won fifteen governors' races, in both parties, and newly elected governors in key AI states — Texas among them — are now threatening to follow through. Below the state level, local bans have spread past five hundred jurisdictions. A year ago this was a zoning-board complaint. Now it's a mainstream platform that wins elections.

The strangest sighting of the week: Amazon and Microsoft siding with local communities against utilities in rate disputes. It's less strange than it looks — they need cheap power more than they need goodwill from utility monopolies, and the communities are angry at the same utilities for the same reasons. The old alignment, where tech companies and local governments agreed on growth, is fracturing in several directions at once.

And the political risk now has a dollar figure attached. Microsoft lowered its data center spending after a third-party cloud provider missed its build timelines, and late-delivery fees have reached five million dollars a day — enough to spawn a new insurance market around the risk. I wrote a month ago that the constraint was in the plumbing; this week the plumbing acquired a voting bloc.

The landlords moved in

Here's the structural change I keep turning over. Apollo — a private equity firm, not a cloud provider — is positioned to control seventy-five billion dollars' worth of AI chips, a position that exceeds any dedicated cloud provider's. On the other side, Nvidia is carrying roughly a hundred and seventy billion dollars in contingent liabilities for its neocloud customers — and this week it was in talks to guarantee debt on the Nscale data center Anthropic uses and moving to back Firmus after investors balked at its IPO. The two even intersect: Nvidia's answer to the data center backlash from the section above is retrofitting Apollo's old telephone buildings.

So the AI infrastructure layer now has a chip landlord on one side and a lender of last resort on the other, with the actual technology operators sandwiched in between. I called Nvidia a central bank of GPUs last week; Apollo completes the picture — the compute layer is increasingly controlled by financial intermediaries, not by the people who run the machines.

The commitments stacked on top make the sandwich heavier. One tally making the rounds puts roughly three trillion dollars in standing capex obligations across Big Tech, with three of the four major spenders reportedly already in negative free cash flow — and Anthropic's compute commitments alone now exceed six hundred and fifty billion dollars. The sequencing is what's historically unusual: these are bets that infrastructure investment must precede monetization by years, not months. Watch the debt has been my standing advice since September; the new wrinkle is who holds it.

The argument over the premium

There's a real disagreement on the table this week, and it's worth laying out both sides. The claim: in enterprise AI, capability beats cost — customers will pay up for the best model even as cheaper alternatives mature. The evidence is Anthropic's nine hundred percent net revenue retention, a number from the IPO reporting that says the customers who stayed expanded their spending ninefold.

The counter: that number measures the wrong thing at the wrong moment. It tells you the stayers expanded; it says nothing about who left before the measurement. And the leavers are starting to show up in the reporting — revenue growth is beginning to slow ahead of the IPO, corporate card data shows AI spending stalling at exactly the wrong moment, and rivals are treating Anthropic's margin as their opportunity, undercutting aggressively on price. Meanwhile distillation keeps making the cheap side cheaper — reported results this week had task-specific small models reaching sixty-seven percent cost reductions without meaningful quality loss, and SaaS startups themselves are now described as facing the risk of being distilled by AI agents.

The scale cuts the other way: revenue surged to seventeen billion dollars in the third quarter, even as the company swung back to an operating loss, and the filing's eighty-five percent gross margin is a reality check for skeptics — the operating model is efficient; it's capital investment and R&D crushing the bottom line. As I unpacked last week, most of the headline net loss is a non-cash accounting charge, which is why GAAP numbers almost seem designed to confuse here: you have to read gross margin, retention, and operating cash flow separately to see what's actually happening.

Where I land, at least this week: the capability premium is real but narrowing, and the market is bifurcating faster than the retention numbers reflect — capability holds the most complex enterprise workloads, price takes everything else, and the open question is how big each bucket turns out to be. The roadshow starts next week, so the market gets its first vote soon.

Nobody's chips are sacred

The supply chain stopped being loyal this week, in every direction at once. OpenAI — historically an almost entirely Nvidia shop — plans to buy Google's AI chips, alongside its ongoing move toward custom silicon with Broadcom. Google, for its part, is openly growing a TPU business next to Nvidia's — selling hardware to rival AI companies because revenue and ecosystem influence now matter more than preserving any compute moat. And DeepSeek is training models on Huawei's chips while helping shape the next generation of them — hardware and software co-designed from the start rather than software layered on commodity hardware after the fact. If that co-design methodology produces chips that can actually compete, it isn't just a China story; it's a playbook other players can copy.

Nvidia's response to all of it is the move I first wrote about in August: convert rivals and strained customers into partners through investment, the Firmus backstop being this week's example. Supply chain resilience has become a strategic priority rather than an engineering preference — even Anthropic is now touting Claude's chip design skills to IPO investors, which tells you model capability and silicon strategy are collapsing into one conversation.

The design loop goes physical

Two stories this week point at the same frontier: the engineering design loop itself. OpenAI announced GPT-6 Astra, aimed at engineering for the physical world — the detail that stuck with me is the model reportedly designing printed circuit boards and mechanical enclosures without direct human edits, with test programs running at Chip Ganassi Racing and Mazda. And Applied Intuition is fielding offers at a thirty billion dollar valuation — not for building autonomous vehicles, but for the simulation software and tooling underneath them. Full-stack vehicle manufacturing is expensive and capital-intensive; software that makes every manufacturer better scales without limit, and investors just made that conviction concrete.

Physical product companies haven't had to reckon with full automation of the design loop the way software teams have. Capability arrived first in software and the organizations lagged; the same sequence is now starting on hardware's doorstep.

Attacks without operators

The security threshold I've been tracking crossed into the mundane this week. Hackers are using Claude and Grok chatbots to automate attacks on banks — chaining commercial and open-source models to remove humans from the attack pipeline entirely. The human bottleneck was the main rate limiter on complex attacks; that constraint is being engineered away. And the physical edge of it got vivid: European supermarkets hit by suspected cyberattacks on their refrigeration systems, spoiling food at scale — a reminder that connected industrial equipment outlives its software support and then stays exposed indefinitely.

The market is pricing the response as large and urgent. Datadog offered nearly four billion dollars for Upwind, a cloud security startup focused on exactly this surface, and CrowdStrike shipped a kill switch for AI agents. The deeper point is the one I flagged two weeks ago, and it keeps getting reinforced: an agent acting within its permissions can still be acting against you. Identity-based access control matches credentials to permissions; what's needed matches behavior to intent. That's a new security paradigm, not an extension of the old one.

Two bets on the interface

The week's clearest framing for designers is a corporate divergence. Salesforce began charging fees for outside AI agents like Claude working inside its apps — pricing for a world where the browser is optional and the agent is the user. Figma's entire bet runs the other way: the interface remains the product. Those are incompatible architectural positions. One of them will look prescient in five years, and the other will look like Kodak doubling down on film.

Underneath both sits the interface problem itself. Chat imposes a sequential, human-paced interaction model on systems built for parallel autonomous action — a mismatch, not a natural fit — which is why tools like Mean, a visual handoff layer between humans and AI agents, are interesting: they assume chat is not the right substrate and build the handoff mechanics directly.

For the profession, the evidence for bifurcation keeps piling up — one designer's account of watching her company design without designers made the rounds this week, and the generalist middle keeps thinning toward either high-level judgment and strategy or AI-augmented execution. The skills that make you valuable in each camp are nearly opposite. Two of them look newly load-bearing to me: writing formal design intent specifications an agent can interpret — NN/g's piece on UX in agentic software factories is the best current map of that territory — and curating a context library, actively auditing what the AI knows, as a discipline separate from prompting. Neither was taught in school, and the designer-as-operator shift I've been mapping since September now runs straight through both.

Also on my radar

The entrepreneurship threads this week mostly compressed into warnings. Vibe-coded MVPs are generating hidden technical debt that can outweigh the speed advantage — and AI has simultaneously raised the interest rate on all technical debt, since vulnerabilities in neglected code now get discovered faster. The hiring shift compounds it: when agents are reportedly running six times the work of your human employees, you stop hiring for junior execution and hire only for judgment and autonomy — judgment over making, a year early. The entry-level pipeline absorbs the hit first, which is how you get essays this week telling would-be developers to learn a trade instead. I don't fully buy the trade-school pivot, but the question under it — where senior capability comes from once junior roles compress — has no good answer yet, and the organizations that find one first gain a durable talent advantage.

Three reinforced threads worth logging. There were reports of China blocking SoftBank's acquisition of ABB's robotics business — geopolitical leverage over strategic hardware is becoming a standing constraint on M&A, the same category as the data center politics above. Multi-model routing keeps strengthening: Microsoft's Copilot Harness explicitly orchestrates agentic work across competing models behind one interface, which means the model is becoming invisible and the orchestration layer is becoming the product. And power stayed the week's most cross-domain constraint — SpaceX is laying groundwork for a turbine blade factory (the turbine squeeze I've flagged twice now has a company attached), one fund is raising two billion dollars to turn idle EV charging stations into AI power, and lenders are growing pickier about financing the build-out even as Nscale carries its unbuilt backlog toward an IPO. One management note I keep coming back to: running AI agents is practice for managing humans, because bad direction produces immediate, legible consequences with no social friction to soften the lesson — a low-stakes gym for a high-stakes skill.