Nvidia's July 14 Whitelist Hits Singapore and Malaysia
What Actually Shipped July 17
Moonshot AI, a Beijing-based lab, released Kimi K3 on July 17, an open-weight model, meaning anyone can download and run it rather than only calling it through an API. At 2.8 trillion parameters, it is the largest open release to date, bigger than DeepSeek's V4 Pro (1.6 trillion) and Zhipu AI's GLM-5 (744 billion). Moonshot priced API access at $3 per million input tokens and $15 per million output tokens, 40 to 70 percent below comparable US pricing, according to the Wall Street Journal. The Philadelphia Semiconductor Index, which tracks chipmakers including Nvidia and AMD, fell 10 percent the week after launch and sits more than 20 percent below its June peak; roughly $470 billion in US AI and chip-stock value disappeared in 72 hours. Treasury Secretary Scott Bessent and White House science adviser Michael Kratsios then alleged Moonshot built K3 by distilling Anthropic's Fable model, training a cheaper copy off a pricier model's outputs, using chips it obtained illegally. The timeline argues against them: Fable launched July 1, sixteen days before K3 shipped. AI2's Nathan Lambert and the Laude Institute's Braden Hancock both said distilling a model this size takes months, not two weeks.
Who The Whitelist Actually Hits
Nvidia tightened its Asian buyer whitelist on July 14, the list of customers cleared to buy its AI chips without extra review, and disqualified more than half of its prior Asian customer base. The trigger was real: a June 5 seizure at Kuala Lumpur's airport recovered $13 million in AI-chip-laden servers, and Bloomberg had reported Nvidia chips reaching Alibaba through a Thailand-based intermediary despite export controls barring direct sales to Chinese buyers. Nvidia responded by sending field inspectors and calling customers directly to verify their businesses were real. The scrutiny landed hardest on Singapore, Malaysia, and Japan, home to ASPIRE 2B, ILMU, and ABCI 3.0, the three governments' own national AI computing programs, built to run domestic research and industry, not to resell chips into China. Nvidia CEO Jensen Huang told Axios on July 22 that Chinese open models like K3 are 'excellent' and that Wall Street 'misunderstood the impact of Kimi,' arguing cheaper open models expand chip demand rather than shrink it. A procurement lead inside any of those three national programs now clears the same paperwork built to catch smugglers routing servers through Thailand to Alibaba, regardless of what they're actually building.
AI2's Nathan Lambert and the Laude Institute's Braden Hancock have already made the distillation timeline hard to defend, and that claim will likely fade once it gets read closely. The whitelist won't. Nvidia's compliance team now vets every Asian buyer against the same suspicion built for Thailand-routed smuggling cases, and that process doesn't distinguish Japan's own AI computing program from a shell company in Bangkok. The open question is whether Singapore, Malaysia, and Japan get a faster review lane, or whether Nvidia's next whitelist review runs the same way the July 14 one did.
Sources
- South China Morning Post -- Kimi K3 spec and parameter comparison (scmp.com)
- TechCrunch -- experts dispute Fable-distillation claim (techcrunch.com)
- Axios -- Huang defends Chinese open models (axios.com)
- BigGo Finance -- Kimi K3 selloff and Moonshot HK IPO (finance.biggo.com)
- Tom's Hardware -- Nvidia Asia buyer whitelist (tomshardware.com)
- Bloomberg -- Nvidia chips suspected smuggled to Alibaba via Thailand (bloomberg.com)