[Emergency Episode] Moonshot’s Kimi K3 has Arrived! China has a Frontier Model

Summary

Moonshot's Kimi K3 open-source model raises questions about frontier AI margins, model sizes, and who is still in the top tier. The hosts argue K3 is clearly the third best model globally, surpasses Google, and that its 2.8T parameter size highlights the need for next-gen Nvidia and AMD accelerators. They also discuss Google's embarrassing fall from the AI lead and why the closed-source labs' pricing power may be sustained despite open competition.

• Kimi K3 is considered the third best AI model, ahead of Google and Meta but behind Anthropic and OpenAI.
• Its 2.8 trillion parameter size requires Nvidia B300/GB300 or AMD MI355X to serve, pointing to upcoming hardware demand.
• Moonshot's open-weight release is delayed 10 days partly to allow inference partners to prepare performant serving.
• The model's pricing tripled to $3/$15 per million tokens, but further hikes may be limited by cost-sensitive users.
• Google is notably absent from frontier AI leadership, which could signal competitive weakness.
• Closed-source labs like Anthropic and OpenAI likely enjoy massive margins at current API prices.
• US government restrictions on Anthropic may be artificially helping Chinese labs close the gap.
• The hosts remain extremely bullish that AI adoption is still in early innings, with most people yet to use frontier models regularly.