Senior White House official claims China’s K3 model stolen from Anthropic

Michael Kratsios claims Moonshot AI’s Kimi K3 (a 2.8T-parameter open-weights model released July 16) was developed via large-scale covert distillation of Anthropic’s Fable, and allegedly involved access to Nvidia GB300 accelerators (including via servers in Thailand). The accusations triggered a near-term drop in US AI stock sentiment as investors worried the model could undermine US AI businesses. Treasury Secretary Scott Bessent said open-source AI is supported, but that “covert, industrial-scale” distillation that amounts to IP theft could lead to sanctions and Entity List designations.

Analysis

The market is reacting less to any one model and more to the policy signal: Washington is telegraphing that perceived IP theft can become an export-control issue, which raises the discount rate on China-linked AI revenue and on any supply chain that depends on routed access through third countries. For NVDA, the direct P&L hit is probably incremental rather than transformative because China exposure is already constrained, but the headline risk can still compress the multiple if investors start underwriting a broader sanctions spiral.

Second-order, the bigger loser may be the ecosystem that monetizes model scarcity rather than raw compute. If open-weight Chinese models can close the quality gap through distillation, US frontier labs face faster commoditization of model access and lower pricing power, while the aggregate demand for inference may actually rise as cheaper models get deployed more widely. That is constructive for the long-duration compute stack, but it shifts power away from closed-model moats and toward whoever can sell the cheapest tokens at scale.

The contrarian risk is that the knee-jerk selloff in AI hardware is too broad. Over 6-18 months, more efficient and more abundant open models can be net positive for total token volume, which supports accelerator demand even if some premium services get de-rated. What would break the bearish case on NVDA is no meaningful new Entity List action within 2-4 weeks and evidence on the next guide that China is not a step-down driver; what would validate it is any formal expansion to third-country hosting or GPU-rental intermediaries.

Key Decisions for Investors

• Do not chase a full-size outright short in NVDA; instead buy a 4-8 week put spread 5-10% below spot to express the policy-headline risk with defined downside. Best if entered on any intraday bounce, since the first move is likely overdone but the catalyst window is open.
• Relative value: long AMD / short NVDA for 1-3 months. Thesis is that NVDA carries the heavier China/export-control overhang while AMD has less headline sensitivity to routed GPU access. Risk is that a broad AI de-rating hits both; stop if NVDA reclaims the pre-news high on no incremental policy action.
• If you are already long AI beta, hedge with a short-dated NVDA call spread overwrite rather than selling the whole basket. The goal is to monetize elevated implied volatility while preserving exposure to the broader inference-demand upcycle.
• Watchlist alert: if BIS names any Thailand/third-country GPU host or expands Entity List language within 2-4 weeks, add to the NVDA hedge. If there is no follow-through by the next policy print, fade the event-driven weakness and rotate back into semis.