SemiAnalysis: Kimi K3 Ranks Third Globally in Overall Strength; Overseas Leading AI May Hide High Profit Margins

Odaily reported that semiconductor research institution SemiAnalysis analysts Jordan Nanos and Max Kan released a report interpreting Moonshot AI's Kimi K3. Comprehensive benchmark tests show that the 2.8 trillion parameter Kimi K3 ranks third globally in overall strength, trailing only Fable 5 and GPT-5.6, with performance surpassing Google Gemini, clearly indicating that the frontier AI gap between China and the US continues to narrow.

The analysts pointed out that Kimi K3 cannot be deployed on a single Nvidia B200 card and requires high-end computing hardware like GB300 and AMD MI355X. This suggests that the flagship model parameter scales of OpenAI and Anthropic are comparable, with no generation-level magnitude advantage.

In terms of pricing, Kimi K3 costs $3 per million tokens for input and $15 per million tokens for output, comparable to Anthropic Sonnet, with prices three times higher than the previous generation. The analysts extrapolate that if Moonshot AI can operate healthily, the API business profit margins for OpenAI and Anthropic at higher pricing levels would be significantly higher than SaaS. This conclusion is based solely on reverse calculations using Kimi's parameters, performance, and pricing, and is not derived from publicly disclosed financial reports of the companies.