China cracked the chokepoint, and the whole stack repriced
The board
▼ AMKR 24.7% ▼ UMC 9.0% ▼ MU 8.9% ▼ AMD 8.2% ▼ ASML 4.4% ▲ WDAY 8.2% ▲ FIG 6.3%
- Markets — the entire semiconductor complex repriced in one session: Amkor −24.7%, Micron −8.9%, AMD −8.2%, Applied Materials −7.8%, ASML −4.4%, the sector ETF −3.5%, and the power names dragged with it (Vertiv −6.3%, Vistra −5.4%). Enterprise software took the other side: Workday +8.2%, Figma +6.3%, IBM +5.2%.
- Open weights — Kimi K3 is now the most-trending model on Hugging Face: 99,000 downloads and 7,800 likes within days of release.
- Models — Opus 5 still holds both ends of the model board: top intelligence at 60.7, and the best value on the board at $10/M blended.
The read
One announcement did most of the damage. China began mass-producing its own immersion DUV lithography, and the market repriced the chokepoint the entire Western equipment stack has been valued on — Amkor down a quarter, ASML, Applied Materials and Lam all sharply lower, the sector ETF with them. The same session, Apple passed Nvidia as the world's most valuable company, which reads less as an Apple story than as capital rotating out of the AI hardware trade altogether. And the week's other export-control skirmish inverted: distillation became a Google product and a Chinese counter-accusation in the same 24 hours — the technique Washington called theft a week ago now sold as a feature, which is roughly how every technical-sovereignty argument ends.
The open-weights position hardened into something more consequential than weights. Moonshot shipped the entire recipe for Kimi K3, not just the model — training data mix, curricula, the lot — which is the difference between publishing an artefact and publishing a capability. Both frontier labs answered in the same week and drew opposite conclusions from the same premise: Anthropic set out its open-weights position (John's read: all three ships have sailed), while Meta's superintelligence manifesto argued the reverse. Underneath the politics, the engineering keeps compressing: an 8-billion-parameter model now ships as a 3.88 GB file with no float weights in the decode path, and a sharp paper argues an uncensored model is a different decision-maker, not the base model minus refusals — a warning for anyone treating alignment as a removable layer.
The measurement story is where the thesis moved. The harness now shifts the coding benchmark more than the model does — the clearest statement yet that value has left the weights for the scaffolding around them. Two new benchmarks put a number on the autonomy ceiling, and the number is sobering. Tom Tunguz's aftermarket-harness piece also supplied the counter-argument we needed: most firms will not build their own tooling, which concedes the obvious objection to the tiny-teams thesis and then reverses it. Poolside's training cadence as the competitive variable points the same way — what you learn per iteration beats what you know — and OpenAI's own scientific-computing field report finds the bottleneck is stewardship, not capability.
Two proofs and a security result show what verification is now worth. Claude Mythos halved a post-quantum signature scheme's security in sixty hours — cryptanalysis at machine cadence, against a scheme chosen precisely for its durability — while the first formally verified 3D mesh intersection landed in Lean 4 and Beyond Zero proposed shrinking the trust boundary to the individual. Verification is becoming both the attack and the defence. Meanwhile the consequences arrived on schedule: displacement stopped being a forecast at Visa and across customer service, agentic trading turned up with a tenfold volume claim attached, the EPA moved to let data-centre power sidestep pollution law, and the ACM began debating whether to hand its Digital Library to the machines — the institutions deciding, one at a time, what they are prepared to feed.