Descriptions:
Matthew Berman breaks down the release of Kimi K3, a 2.8 trillion parameter open-source model from Moonshot AI that he argues represents a genuine frontier-level achievement from China’s AI ecosystem. The model supports a 1 million token context window, native multimodal inputs, and benchmarks comparably to GPT 5.6 and Anthropic’s Fable — at roughly half the API cost, coming in at $3 per million input tokens versus $5 for GPT 5.6 (though Berman notes K3 tends to use approximately twice as many tokens per task, making cost-per-task the more meaningful metric).
Berman uses the release as a foundation for explaining the strategic logic behind Chinese AI labs open-sourcing their most capable models. He covers ecosystem control (become the standard, then shape its future), the ‘scorched earth’ margin-destruction strategy — historically paralleled by Linux, Android, Chromium, and Meta’s Llama — and China’s heavy state subsidization of frontier AI research as a geopolitical lever. The goal, he argues, is for Chinese AI and eventually Chinese chips to become global infrastructure that other countries depend on.
The video also addresses why open-source is not actually free at frontier scale: running a 2.8 trillion parameter model requires expensive hardware and significant electricity, with real costs simply shifted from API fees to compute infrastructure. With the US government reportedly considering restrictions on Kimi K3, Berman positions this as essential context for anyone tracking the intersection of foundation model development and great-power competition.
📺 Source: Matthew Berman · Published July 21, 2026
🏷️ Format: News Analysis
