Moonshot AI is intensifying its hunt for Nvidia’s Blackwell chips to train Kimi K4, an ambitious upgrade from its already massive 2.8 trillion-parameter Kimi K3. Released just days ago, Kimi K3 stunned with a context window of 1 million tokens, placing it among the world's largest open-weight AI models. The push to build Kimi K4, expected to be significantly larger, means the startup’s chip demand will spike well beyond its previous needs.
This pursuit unfolds amid growing tensions. The U.S. government has tightened semiconductor export controls targeting China, yet Moonshot managed to build Kimi K3 with Blackwell chips, raising eyebrows. A public accusation came swiftly after the launch: Michael Kratsios, White House science policy director, alleged Moonshot used servers in Thailand to access restricted compute resources. This move reportedly aimed to replicate U.S.-made AI models, including Anthropic’s Fable, bypassing export rules through geographical loopholes. Moonshot has stayed silent on the details.
Nvidia’s CEO Jensen Huang recently stressed that more powerful AI models like Kimi K3 don’t reduce chip demand; they actually increase it by driving wider AI adoption and heavier inference workloads. Moonshot’s plan to release Kimi K3’s full weights to the public highlights its commitment to open collaboration, with partners like Nvidia, AMD, and vLLM involved. Meanwhile, U.S. scrutiny over Southeast Asia’s role in serving restricted tech could impact cloud and data center companies indirectly supporting startups like Moonshot.
As the AI arms race heats up, Moonshot’s quest for Blackwell chips reveals how hardware bottlenecks and geopolitical friction are shaping the future of massive AI models. The outcome of these tensions might redefine supply chains and enforcement across borders, affecting not only Moonshot but the broader AI ecosystem.
This material is for informational purposes and does not constitute financial advice.
