Built on Fiction

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

The valuation is a circular argument

The generous reading of the AI buildout is that every major infrastructure cycle in history ran on forward commitments made before the assets existed — railroads, telegraphs, fiber. The less generous reading arrived with the details of Nscale's IPO pitch: a thirty-five billion dollar valuation on the strength of a hundred and three billion dollars in contracts tied to data centers that haven't been built. That isn't a forward commitment, it's a circular argument — the valuation justifies the fundraise that is supposed to build the thing that justifies the valuation. And the circle has a trapdoor in it: the contracts carry walk-away clauses tied to delivery milestones, so the backlog can evaporate the moment construction falls behind. The paper-versus-revenue question I wrote about last week didn't just persist; it got sharper.

The debt side is flashing the same warning, which matters more, because debt reprices on a schedule while equity narratives can stay fictional as long as nobody tests them. CoreWeave's thirty-five billion dollar debt binge has reached what one report called a perilous reckoning, and data center bonds tied to single high-profile tenants are reportedly deteriorating faster than peer issuances — concentration risk being priced in, just slowly. Watch the debt has been my standing advice for a month now, and this is the second week running that the bond market has been the adult in the room while the promises stay load-bearing.

Power is the part you can't pitch

Even if the buildings go up, the electrons are not guaranteed. Microsoft exited a committed UK data center deal over electricity supply delays — a major cloud provider abandoning capacity it wanted, purely because the grid couldn't promise power on the timeline. That breaks every model that assumes demand is the binding variable. You can have all the contracts you want; if the power isn't there, none of it matters.

The tell for how severe this feels from the inside came in the same reports I flagged two weeks ago and which resurfaced again: hyperscalers reportedly weighing vertical integration into turbine blade manufacturing. When a software company is seriously considering making turbine blades, it is not optimizing anymore — it is in survival mode on the infrastructure question, reaching for a competency about as far from its core as physics allows.

The same dollar changes flags

The geopolitical story and the investment story are now the same story, and one transaction made that unusually literal. Accenture is weighing a one-and-a-half billion dollar sale of its entire China consulting business, driven by Beijing restricting foreign consultants from advising state-owned enterprises — and its explicit plan for the freed capital is to redirect it into AI globally. The same dollar, flowing out of China and into GPU clusters. I'd expect that pattern to repeat across professional services: the squeeze has been tightening for a while, and this is the moment firms decide the optionality isn't worth holding.

China is running its own version of the same move. MOFCOM and MOST proposed expanding the export control catalogue to cover frontier AI models, chip design, and minority-language AI — the weights themselves are now the controlled item, in the same category as dual-use hardware. That is genuinely new. A few years ago this conversation was entirely about chips and lithography; the split I started tracking in September has reached the model layer. And DeepSeek's reported fifteen billion dollar round — with Tencent and CATL said to be among the backers — sits right in the middle of it: capability built specifically to not depend on models you can't control geopolitically.

Fifteen billion is enormous even by AI standards, and the scale is deforming the instruments that are supposed to channel it. Venture firms are tapping LP pools, warehouse loans and special-purpose vehicles just to write single checks large enough to participate in frontier rounds — the round has outgrown the fund. Top firms are also reportedly securing GPU allocations as a core portfolio service, which makes compute scarcity a structural differentiator in startup competition rather than a cost line. And on the other side of the table, the story of Anthropic's cap table shows a lab deliberately capping any single VC's ownership below two percent — not a fundraising tactic so much as a statement about who gets to influence a safety-focused lab's direction, the governance-as-scarce-asset theme expressed in equity structure.

Permitted is not the same as intended

The security layer was the underrated story of the week. Fully automated AI-driven cyberattacks on financial institutions are now documented and operational — not a forecast, actual chained multi-model pipelines that removed the human bottleneck from the attack pipeline, with Claude unwittingly doing part of the work. Security researchers were reportedly turning the same trick on the labs themselves, using Claude-family models to probe OpenAI's systems, and cybersecurity stocks have reportedly rallied hard this year on exactly this dynamic. AI is the attack surface and the defensive opportunity at the same time.

The contested question is whether agent security is actually a new problem or the authorization problem we've always had. I went back and forth on this, and I've landed on: it's new. With a human, you can reason about intent from context, and a social layer catches most permitted-but-wrong actions before they land. With an agent, an action can be technically within its permissions and completely contrary to what the user meant, and nothing catches it before it executes. The skeptic's reply — that "behavioral governance and intent-matching" is just the problem restated as an architecture — is fair as far as it goes. What makes it more than a restatement is that the layer is actually shipping: the defense patterns coming out of real agent incidents are converging on runtime enforcement above the model, and Google Cloud introduced an agent compute engine with expanded agent governance in the same week. The control layer I described in September is becoming a distinct product category, not a feature.

The squeeze at the top of the stack

The pricing-pressure story got its most concrete version yet. Shopify is turning open models into fine-tuned task experts for Sidekick — roughly ninety percent cost reduction, seven times faster, matching frontier accuracy on the narrow task. That's not a procurement anecdote, it's a structural claim: if the winning strategy for production workloads is fine-tuned open models rather than frontier API calls, the revenue model at the top of the stack has a problem. And the squeeze comes from the side as well as from below — major vendors are reportedly routing workloads across competing foundation models behind a single interface, decoupling the user-facing brand from any single provider, and there were reports this week that OpenAI's models are coming to Google Cloud's marketplace. Distribution is decoupling from exclusivity; "OpenAI on Google Cloud" is a sentence that would have been strange eighteen months ago.

So what actually wins up there? The moat is migrating from model quality to distribution, infrastructure access and financing capacity — which is why the labs' strategic moves increasingly look like infrastructure finance rather than research. Google Cloud and NVIDIA's Vera Rubin general availability fits the same read: the A5X instances are purpose-built for large-scale reinforcement learning and agentic inference — infrastructure betting on the agent paradigm as the primary competitive surface, not on raw model performance.

The spec is the instruction set

The design thread worth pulling out of the infrastructure noise: every product's underlying workflow architecture is now effectively a public API. Workflow friction that blocks human users will equally block your users' agents — bad workflow design used to be a human UX problem you could partially paper over with training and habit; now it's a compounding liability, because you're also degrading every automated pipeline that touches you.

The source of truth is moving too. As AI tools generate production-ready code, Figma-first workflows are increasingly misaligned with how products actually get built — the codebase is becoming the canonical artifact. One designer walked through building an entire product in Claude Code without opening Figma, and EditUI makes the new loop explicit: point at a UI element, tell a coding agent what to change, and it edits the code directly — no design file touched. Which is why spec writing is becoming a critical design skill: not documentation for other humans, but literally the instruction set for the AI generating the code. The designer-as-operator shift, now with syntax. One honest caveat on the research side: NN/g found synthetic users reach the same conclusions as human respondents with less variation — they surface majority patterns and lose the rare behavior. And rare user behavior is rare precisely because you didn't anticipate it; a synthetic population approximates a distribution you defined, it doesn't discover the one you didn't know existed.

Under all of this sits the human capital question I keep circling. Anthropic's own look at what happens when agents work six times more hours than employees points where you'd expect: hiring shifts toward judgment and autonomous oversight, and entry-level technical pipelines compress before senior ones do. Pair that with the skill-atrophy argument — taste, judgment and critical thinking are maintained by deliberate practice — and the uncomfortable arithmetic appears: if you outsource execution to agents, you may be outsourcing the practice reps that build senior judgment. Judgment over making has been my thesis since August; this week asked what happens to the ladder that used to produce the judgment.

Also on my radar

The tooling buildout around agents is hardening into a real market. Daytona raised sixty million dollars at a billion-dollar valuation for AI agent cloud infrastructure, and Blacksmith hit a billion-dollar valuation with annual recurring revenue up seventeen-fold on a thesis as simple as it is telling: AI coding agents generate so much code that CI infrastructure has to scale to match. The argument behind the category — AI-written software needs a flight-simulator-style stress test — is directionally right even if I'd quibble the analogy: flight simulators test known failure modes of understood systems, and AI-generated code fails in ways we don't fully understand yet. The gap is real; the name for it matters less.

And one number almost too good to believe: OpenAI reported legacy-code modernization accuracy — COBOL migration included — jumping from roughly five percent to the mid-eighties across model generations. If that holds, it dissolves a category of enterprise problem people had basically given up on. The "if" is doing a lot of work, and the week supplied its own reason for skepticism: the drama inside biology's AI-versus-human contest is a case study in how headline benchmarks collapse under scrutiny. Hold the excitement and the doubt at the same time.