Chip, memory stocks fall amid calls for AI development slowdown

European and Asian semiconductor stocks sold off after Anthropic CEO Dario Amodei proposed slowing the pace of frontier AI development: SK Hynix and Kioxia fell 6.4%, Samsung declined 4.1%, while ASML lost 4.4% and Infineon and ASM International dropped more than 5%. The proposal revived concerns that slower AI-model development could curb AI infrastructure spending, particularly pressuring HBM suppliers whose margins depend on tight supply and favorable pricing. Bernstein remains constructive on the AI buildout, arguing that a slower training pace does not necessarily mean lower capex given compute shortages, inference demand and Anthropic's roughly $517B of capacity commitments over the 11 months through August.

Analysis

The selloff should be treated initially as a positioning and duration-multiple event rather than evidence of an AI capex reset. A moderation in frontier-training cadence would affect the marginal accelerator order more than installed-base inference demand; NVDA and AVGO retain exposure to both, while MU and SKHY depend more directly on sustained HBM tightness to protect incremental gross margin. The key distinction over the next 1-3 months is whether hyperscaler procurement schedules move, not safety-policy rhetoric.

Memory is the most asymmetric downside pocket: HBM pricing is supported by qualification bottlenecks and contracted allocations, but inventories can re-rate quickly once customers stop pulling forward supply. MU/SKHY earnings estimates would be vulnerable to even a modest delay in 2026 capacity absorption, whereas ASML's revenue is better buffered by backlog and service, though its multiple remains exposed to any reduction in leading-edge wafer-fab-equipment order visibility. European analog/industrial names—STM and IFX—have less direct AI upside and can be sold as liquid proxies for a broader semiconductor de-risking.

Contrarianly, a safety-driven pacing framework may strengthen incumbent economics. Higher compliance, compute-security, and model-governance costs raise barriers to entry, concentrating demand among well-capitalized labs and hyperscalers rather than eliminating it. Export-control enforcement is a more material downside variable for ASML and the semiconductor equipment chain than the proposed pacing language; a tightening in China-related restrictions could turn a sentiment dip into an earnings-risk event over 6-18 months.

Anthropic's reported commitments should not be capitalized one-for-one into supplier revenue until power delivery, financing, and binding take-or-pay terms are independently confirmed. Watch cloud-provider capex guidance, HBM lead times/pricing, and NVDA/AVGO backlog commentary; adverse revisions there would falsify the view that the current weakness is mainly technical.

Key Decisions for Investors

• Initiate a 1-3 month pair: long NVDA / short MU, sized beta-neutral. NVDA has broader inference and networking monetization, while MU is more exposed to HBM pricing normalization; target 10-15% relative upside, stop if MU raises HBM supply tightness or NVDA cuts next-12-month demand visibility.
• Accumulate AVGO on further sector-driven weakness rather than chase the open; use a 3-6 month horizon and a 8-10% downside risk limit. Custom silicon plus networking should benefit if customers optimize compute efficiency rather than simply maximize training clusters.
• Use STM or IFX as the preferred tactical short/underweight versus ASML for a European semiconductor risk-off basket over 1-3 months. Their cyclical end-market exposure offers less protection if AI enthusiasm fades; cover on a clear industrial-demand inflection or upward guidance revision.
• Maintain ASML as a watch-list long, not an immediate add: enter only after management confirms order-book resilience and China export assumptions. The risk/reward improves materially following a further multiple compression, but policy-driven shipment restrictions remain the thesis break.
• Set alerts for hyperscaler capex guidance cuts, HBM spot/contract price declines, or a meaningful extension in memory lead times. Any two signals together warrant reducing AI hardware gross exposure, especially MU, SKHY, ASM and BESI.