Five things from this week that actually change how I'd build
Five things crossed my feed this week that, read together, are describing the same shift from five different angles: the bottleneck in AI products has moved off engineering and onto judgment — deciding what to build, and what it should cost to run.
PMs are the new scarce resource
Anthropic told its growth org to hire more product managers, not more engineers. Claude Code has pushed some teams to 2–3x their effective headcount — a five-person team shipping like fifteen to twenty — and in some periods over 80% of merged production code has come out of the tool. The constraint didn't disappear, it moved. Nobody scaled the people who decide which of those 15-person-team's-worth of features actually get built, in what order, for whom. That's now the expensive seat, and Anthropic is paying $305K–$460K for it.
The leverage question
Lenny's Newsletter ran a piece the same week on how top PMs increase their leverage with AI. Different framing, same diagnosis — once execution stops being the constraint, the PM's job stops being coordination-of-execution and becomes almost entirely judgment: what's worth building, and saying no to the rest with real conviction.
Token economics just got a standards body
The Linux Foundation announced the intent to launch a Tokenomics Foundation — a dedicated standards body for AI cost economics, backed by Google Cloud, Microsoft, SAP, Accenture and others. This is the part I actually build for a living, and it's validating in an odd way: token spend just got treated as a first-class economic category, the way cloud spend did a decade ago with the original FinOps movement. I've made the case before that AI cost isn't "cloud cost management 2.0" — it spikes for reasons a compute-hour dashboard was never built to catch. Having an industry body say the same thing, independently, is the kind of confirmation you don't get to manufacture yourself.
The compute crunch underneath all of it
The Information's AI Agenda reported Nvidia is now financially backstopping customers' AI chip purchases — because the 2026 compute crunch is worse than the 2023 chip shortage. That's not a side note to the cost story, it's the upstream cause of it. If compute itself is scarce enough that Nvidia needs to underwrite demand, then every dollar of token spend downstream carries that scarcity with it. Cost governance isn't a nice-to-have layer on top of AI products anymore; it's load-bearing.
Pricing moves whether you shipped anything or not
Anthropic shipped Claude Sonnet 5 as the new default model across Free and Pro plans, with introductory pricing through the end of August. Model upgrades used to be a footnote for most product teams. In a world where token economics are becoming their own discipline, a default-model change is a line-item change — the same feature can get meaningfully cheaper or more expensive overnight, for reasons that have nothing to do with anything your team shipped.
Put together, none of these are really five separate stories. They're the same fact, seen from HR (PMs are the scarce resource now), from the newsletter beat (judgment is the leverage), from standards bodies (token cost is now formal economics), from infrastructure (compute scarcity is the root cause), and from the model layer (pricing moves under you, not because of you). If you're building AI products right now and none of this is on your radar, it's worth putting there — it's the one shift I don't think reverses.
Built something like this? I'm always happy to compare notes.
aniket.kgp25@gmail.com →