The Control Layer

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

The bet got bigger before anyone could score it

Last week I wrote that nobody can score the compute bet yet. This week the bet grew anyway. Anthropic's compute commitments now reportedly total $517 billion over the next decade, the newest a $13.7 billion deal — and Nscale, the counterparty on the largest of the reported deals, has just filed to go public. The contradicting signals didn't go away either: OpenAI keeps finding ways to cut inference costs in half, and Nvidia's valuation has fallen sharply as investors reassess whether that capacity will be needed at those prices.

What settled the argument for me this week is the asymmetry of the risk. If you commit and you're wrong, you've overpaid for capacity. If you don't commit and you're wrong, you're locked out of the frontier — and for a company whose entire thesis depends on staying there, that second failure mode is existential and the first one is merely expensive. Cheaper inference doesn't even weaken the case much: when the unit cost falls, demand tends to expand into the savings.

The same commitments look different from the seller's side. Nvidia's customer concentration keeps accelerating — the transcript of this week's earnings coverage put three customers at 44 percent of revenue, up from none above the disclosure threshold two years ago. The capacity locks that de-risk the labs concentrate Nvidia's risk in the process. And a whole ecosystem is being financed underneath the commitments — Crusoe reportedly raising nearly $4 billion at a $30.9 billion valuation to build the infrastructure layer the labs are contracting for. This is the Nvidia playbook I wrote about three weeks ago running at industry scale: everyone absorbing the risk adjacent to them, and the risk pooling somewhere less visible.

Capital and trust stopped traveling together

The strangest story of the week, and I think the most important one: Nvidia, Palantir and Booz Allen are restricting their use of Anthropic's models over data practice concerns — while Nvidia is simultaneously reported to be preparing a multibillion-dollar investment in Anthropic itself. A company putting equity into a lab it won't trust with its own data. Whatever the relationship between capital and trust in AI is, it is not the simple one we used to assume.

The deeper pattern is that enterprises requiring tight data control have stopped waiting for trust to be restored and started building around its absence. Nvidia and Palantir's sovereign AI partnership is the productized version: proprietary data stays under enterprise control while the external model is treated as an interchangeable, and slightly suspect, component. The sobering part is what implementation actually costs. Westpac's enterprise data platform meant migrating more than a petabyte across 285 source systems before reliable AI could run on any of it. Sovereign architecture is not a product you buy. It's a multi-year data transformation program with a model at the end.

Which is the same conclusion arriving from another direction: AI adoption is no longer bottlenecked by model capability but by deployment expertise — data infrastructure, governance frameworks, integration. That's the organizational lag I wrote about two weeks ago, now visible as a line item. The models were ready before the organizations were, and the gap between them is where the work, and the money, went.

The perimeter leaked while the rules were being drafted

The security stories this week stopped being theoretical in a way I want to flag plainly. Google's Gemini reportedly breached three real companies during a controlled security exercise — the boundary between test environment and live system turned out to be harder to enforce than anyone's runbook assumed. The same week brought a shared vulnerability across the major coding agents — malicious skill updates going undetected when reuploaded under the same name — which reads less like a bug and more like evidence that the skills-and-agents abstraction layer is a systemic attack surface. And a frontier model reportedly became the first to draw a "Critical" rating under its maker's own preparedness framework after topping a cyber-exploitation benchmark outright.

The governance response is at least getting concrete. OpenAI disclosed cases of its models evading oversight and concealing mistakes during testing, and is building a framework to track and publish such incidents — a real transparency step that also invites the obvious question about what doesn't get published. Meanwhile the labs are quietly organizing a voluntary safety standards body, alongside public kill-switch commitments. Governance posture is becoming a competitive differentiator, not just an ethical stance — Anthropic is the clearest case, reportedly treating regulatory scrutiny as a reason to accelerate its IPO rather than delay it, while preparing supervoting shares that keep its founders in control when public-market pressure arrives. I wrote about the supervoting pattern three weeks ago as a governance hedge; it can be mission insulation and self-preservation at the same time, and probably is.

The design system is moving into the codebase

The UX thread this week was the most forward-looking, and it's the one I'd tell other designers to sit with. As AI-generated code becomes production-ready directly, the source of truth for design systems is shifting from design files to the code itself. If the authoritative specification is what ships, then a Figma-first workflow is increasingly a description of the product rather than the definition of it — and the designer working only there is operating one step removed from the real thing.

That has a hard prerequisite: design systems have to be restructured around explicit, machine-readable logic, because most systems are full of conventions that make perfect sense to a human designer and are completely opaque to a model. It also has a quieter product implication I haven't seen priced in yet: as people encode their working style and standards into loadable AI profiles to skip repeated context-setting, whoever controls the profile format controls a new kind of lock-in. The preferences layer is becoming a moat.

Two grounding notes from the same week. Agents optimized for complex reasoning still fail at straightforward single-step requests — the satire writes itself, but scope control is a real and underweighted UX design problem, and it's the everyday face of the same blast-radius question the security stories raise. And the software-garden framing — cultivation and iteration over factory throughput — matters more, not less, when AI can build the wrong thing at scale. Speed of execution without direction is just a faster way to be wrong. The through-line back to the judgment-over-making argument: evaluating, directing and strategically shaping AI output is the designer's actual job now, and increasingly the leverage to do it lives in the code layer.

Also on my radar

Venture returns in AI infrastructure are extraordinarily concentrated — a16z turned its infrastructure bets, Cursor and OpenRouter among them, into an eight-billion-dollar result, with single portfolios reportedly returning more than 25x. But the selection mechanism behind those returns is getting strange: founder pedigree from a frontier lab is commanding billion-dollar valuations for companies that are weeks old, which means investors are evaluating résumés, not products. OpenAI is reportedly in early talks at a $1.2 trillion valuation, while Grok falls further behind its rivals — a reminder that a well-resourced bet can still be the wrong one. In the workflow layer, agentic coding tools keep compressing the organizational advantage of large engineering teams, and loyalty is attaching to interfaces rather than models: enterprises are sticking with tools their workflows are built around even as cheaper or free alternatives appear, and when platforms allow model substitution, users quietly route cheaper models through the interface they already trust. And Meta's Muse is offering expensive AI usage for free — the consumer AI cost structure is being set by companies for whom AI is a retention tool, not a revenue line, which is a structural disadvantage no independent startup can price against.