A Chinese open-source model just triggered another AI selloff. The panic says more about US spending than about Kimi.

This is less a clean fundamental negative for semis than a valuation shock to the AI trade's scarcity premium. The first-order move is a de-rating of anything priced on model exclusivity or linear monetization, because cheaper frontier performance compresses the perceived moat of incumbent US AI stacks. The second-order effect is more nuanced: if model quality becomes easier to source, enterprise adoption can accelerate, which is ultimately positive for inference-heavy infrastructure and the cloud platforms that monetize usage rather than scarcity.

The biggest near-term losers are the crowded, high-beta baskets: SMH/SOXX, AI software names with stretched multiples, and any vendor whose pitch depends on proprietary model leadership. The longer-duration winner set is the compute supply chain that benefits from a larger total addressable market for inference and agentic workloads: NVDA, TSM, ASML, and hyperscalers such as MSFT/AMZN/GOOGL if capex guidance stays firm. The market is likely over-interpreting "better model" as "less compute"; historically, efficiency gains have often been reinvested into more usage, not lower spend.

The key risk is that this evolves from sentiment shock into a real capex revision if hyperscalers decide model progress is now cheap enough to slow training budgets for 1-2 quarters. That would hit semis first and the broader AI complex second, with the feedback loop showing up in order books and guidance rather than press headlines. Conversely, if the next major earnings round shows unchanged or rising capex plans, this becomes a fast-reversible positioning flush rather than a structural repricing.

Contrarian view: the market may be underpricing how much Chinese model progress pressures domestic China AI winners to adopt faster, cheaper stacks while leaving US infra vendors relatively insulated. The most durable competitive damage may land on application-layer vendors with weak workflow lock-in, not on the chipmakers that sell the picks and shovels.