Capability keeps deflating in the open; defensibility retreats to the stack
The board
▼ CRWD 5.7% ▼ UMC 4.6% ▼ AMD 4.2% ▼ LRCX 3.1% ▼ META 2.8% ▲ ASML 2.9% ▲ PLTR 2.8%
- Models — Gemini 3.6 Flash: same intelligence as 3.5, about half the time-per-task, cheaper output. Poolside ships an open-weights coding model; Korea's Motif-3 (341B) stays closed.
- Value — Kimi K3 tops the model board on value: $6/M blended, 9.52× intelligence-per-dollar.
- Status — OpenAI degraded for part of the day.
The read
The through-line, again, is deflation in plain sight. Google's Gemini 3.6 Flash holds its predecessor's intelligence while halving time-per-task and cutting output pricing — efficiency, not capability, is the release. The sharper argument of the day was Ben Thompson's claim that frontier tokens are non-fungible luxury goods; OpenRouter's routing volumes and the relentless optimisation coming out of Chinese labs both push back. Mark caught the shape of it in passing: every thread bends the same way — raw capability is deflating in public while defensibility migrates to the layers where copies don't help.
Those layers were on display. On the software side, the day's best writing converged on which layers survive commoditisation — convenience dies, the expression of decisions stays — alongside two mental models for agent-augmented work, Block's bet that signed identity makes agents accountable, and Poolside's class-leading open-weights coding model. On the physical side the cost floor is rising even as token prices fall: TSMC will lift wafer prices up to 10%, Nvidia's Vera CPU pushes into the data-centre, and the buildout is now visible in share prices, land and cooling patents — and, most starkly, in the IRGC's claimed cruise-missile strike on an AWS facility. Even Airbus leaving AWS reads as the same story from the demand side: control of the stack is worth paying for.
The policy front hardened in step. The same day a judge approved Anthropic's US$1.5bn author settlement, Treasury began framing Chinese "distillation" as theft — the double standard John flagged: Chinese labs breach terms of service but pay, while US labs breached copyright wholesale and, Anthropic aside, paid nothing. Lab lobbying hit a record. Set against the sell-side euphoria a China-AI correction warning punctured, the market is still pricing permanent scarcity into a technology visibly busy dismantling it. A quieter marker of the same shift: mathematicians report being "out-counterexampled" by their own tools.