The Scarce Things Changed

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

Nvidia buys the hub

The reported $12.9 billion Nvidia–Hugging Face acquisition is the clearest single expression of that shift. Nvidia already owns the chip layer; Hugging Face is where developers find, share, and deploy models — the community layer, the distribution layer, and a massive data asset in one move. It's also a departure from the licensing-plus-hiring playbook I wrote about last week: this would be a full acquisition of a platform whose value is partly open-source goodwill. The moment developers perceive the hub as an Nvidia sales funnel, that goodwill starts to erode. And notably, reports conflict on whether the deal is actually signed — which itself is a signal that Nvidia wants credit for the intent.

The timing explains the logic. Open-source models just captured 62% of token share on a major frontend cloud platform within two months — not a gradual trend, a flip. If open-source wins production workloads, the platform that distributes open-source models becomes the moat. But it sharpens a tension I flagged before: Nvidia is now a chip vendor, a model developer, and potentially a platform owner. At some point its customers have to ask whether Nvidia is a partner or a competitor.

Priced for perfection

Last week I asked whether supervoting shares at an AI lab were mission protection or founder entrenchment. This week suggested the question is becoming moot: dual-class supervoting structures are turning into a standard precondition of AI founder-led IPOs, not an exception. Anthropic's reported structure pairs it with something more unusual — letting some existing shareholders sell in the offering while locking others up longer than usual. That's a pressure valve for employees and early investors sitting on enormous paper gains with no exit.

The numbers around it are staggering: a $965 billion private valuation, with bankers floating $1.5 trillion for the public offering — for a company still pre-profit in any meaningful sense. And alongside it, a reported $45 billion, six-year compute commitment in West Virginia. That's not an operating-expense decision; it's building a utility. West Virginia makes sense not for the talent pool but because you can actually get the megawatts. If the safety framing is the justification for keeping control, public markets will eventually hold the mission to that promise.

Compute was never the moat

The most important correction of the week: stockpiling GPUs doesn't make you the research leader. Musk publicly admitted Grok lags behind Anthropic despite xAI's aggressive GPU accumulation — remarkable candor that confirms what the data was suggesting. Meanwhile Google is losing the people who built modern AI: Jeff Dean, Demis Hassabis stepping aside, Noam Shazeer to OpenAI, John Jumper to Anthropic. You can throw money at recruiting, but you can't buy back institutional knowledge on a hiring timeline.

So what is actually scarce? Three things kept surfacing. Power: data-center build-out is hitting zoning bans and grid constraints that pure capital cannot solve. Talent: see above. And data: with public internet data increasingly exhausted as a training source, proprietary workplace behavioral data — Slack threads, code commits, meeting transcripts — becomes the next frontier input. Even Anthropic's acquisition of a consultancy reads two ways: implementation support on the surface, and underneath, proximity to the enterprise data those workflows are sitting on. Services-led growth as a data-access strategy — whether or not that's the intent, it's the structural consequence.

The designer's job is judgment

The UX signals this week were scattered but coherent: the designer's job is shifting from making things to judging things. Component systems are now being built machine-readable-first, specifically for AI agents to consume — the design artifact isn't a Figma file, it's an explicit system an agent can interpret and apply. At the other end, "software for one" is becoming real: when AI can build a personal app for a single user in a day, you're no longer designing for the average user — you're enabling personal adaptation. And when production is democratized like that, functional parity is trivial; experience quality becomes the only differentiator left.

The uncomfortable organizational half of this: design functions lose influence not because the work is bad but because leadership can't interpret it as business value. If design's value is increasingly judgment — evaluating AI output, directing generation, setting standards — that's even harder to make legible to a CFO than shipping screens was. One framework from this week that I'm sitting with, for problem spaces too new to research conventionally: what could this technology do, what should we do with it, what might we do as it develops, and what shouldn't we do regardless. Those four questions bound a design space before user research can even start.

Also on my radar

The venture-return concentration story: a16z's infrastructure practice reportedly cleared 25x on a batch that included Cursor and OpenRouter — the same OpenRouter I held "with appropriate uncertainty" last week. The returns look real; the repeatability is genuinely uncertain, and Grok is the counterweight to any narrative that resource concentration reliably wins. Two smaller patterns worth keeping: hybrid moderation architectures — cheap classifier first, expensive model only when confidence is low — as a template for any fast-cheap/slow-accurate pipeline; and developer tools holding enterprise users through price increases because switching costs moved from contracts to cognition. Once your workflow lives inside a tool, leaving is technically easy and practically painful.