Just as Nvidia (NVDA) shares were getting hammered by the debut of a groundbreaking Chinese AI model, CEO Jensen Huang stepped onto a stage in China and did what few American tech chiefs would dare: he openly praised the source of the turmoil. His endorsement of the very model that wiped out billions in Nvidia’s market cap wasn’t a lapse in judgment—it was a calculated strategic play hiding in plain sight.
The tremor began in mid-July, when Chinese startup Moonshot AI unveiled Kimi K3, an open-weights model with 2.8 trillion parameters and a one-million-token context window. It immediately shot to the top of the Frontend Code Arena leaderboard for frontend coding and posted strong scores on the GDPval-AA v2 benchmark. Wall Street’s reaction was swift and brutal. The semiconductor sector plunged more than 20% from its June 2026 peak, and Nvidia briefly lost its crown as the world’s most valuable company.
While the market was still in panic mode, Huang appeared at the China International Supply Chain Expo with a message that flew in the face of the sell-off. He name-checked Kimi alongside models from DeepSeek, Alibaba, Tencent, MiniMax, and Baidu, calling China’s open-source AI ecosystem “world class” and globally competitive. American companies, he insisted, should “absolutely” be allowed to use these Chinese open-source models—a position that puts him directly at odds with Trump administration officials and U.S. AI labs lobbying Washington to ban them outright.
Huang is no free-trade purist. He supports restricting the export of Nvidia’s most advanced chip architectures, including Blackwell and Rubin, to China. His argument is nuanced: limit the hardware if necessary, but keep the software open. Open-weight models, he contends, are better for safety and collaboration than the closed proprietary systems that dominate the U.S. landscape.
At first glance, praising the model that just eviscerated your stock price looks like corporate self-sabotage. But a closer look at Kimi K3’s architecture reveals why Huang can afford to be magnanimous. According to semiconductor research firm SemiAnalysis, the model’s celebrated Kimi Delta Attention (KDA) innovation cuts KV cache requirements by 75% and boosts inference speed sixfold by selectively reading prior tokens. Yet Kimi K3 remains enormously compute-hungry. Its massive size and mixture-of-experts design with 896 experts employ WideEP technology, spreading experts across at least 56 GPUs just to run inference. To operate the model efficiently, SemiAnalysis notes, you need high-end reference architectures like Nvidia’s GB300 NVL72 rack systems. The reduced cache per chip, counterintuitively, drives a massive increase in cross-chip communication bandwidth, reinforcing reliance on Nvidia’s proprietary NVLink interconnects.
This is where the Jevons paradox kicks in. Efficiency gains, rather than reducing total resource consumption, often accelerate adoption and push aggregate demand higher. Kimi doesn’t lessen the need for cutting-edge Nvidia hardware; it mandates it. Huang’s embrace of the model is a bet that a flourishing open-source AI ecosystem—Chinese models included—will ultimately require an even denser foundation of high-performance compute infrastructure, a foundation Nvidia is uniquely positioned to supply.
Regulatory dynamics add another layer. Even if U.S. enterprises face pressure to favor domestic models, a fragmented AI landscape where certain jurisdictions ban Chinese open-source software could make censorship-resistant, globally interoperable infrastructure more valuable. By advocating for open software while respecting hardware export controls, Huang positions Nvidia at the center of both Western and Eastern AI build-outs.
Huang’s public nod to Kimi as Nvidia’s stock tumbled was not an act of reckless contrarianism. It was a clear-eyed recognition that a thriving open-source AI world—no matter where the models originate—ultimately runs on silicon. And Nvidia intends to remain the indispensable layer beneath it all.
