AI Moat Digest

Open source models challenge proprietary giants

Open source models challenge proprietary giants

Commoditization continues with Kimi K3 (2.8T MoE) and Qwen3.8-Max (2.4T) open-weight releases. New distillation methods (Any-OPD enables heterogeneous distillation) further commoditize. DeepSeek V4 Flash open-sourced at low cost. Mistral released Shieldstral (3B open-weights moderation). Video-DeepResearch shows 35B model beats larger closed models. LLaDA MoE v2 scaling laws for diffusion LMs challenge AR dominance. Vertical AI moats remain nuanced: Marble, plasma proteome classifier, QuantHealth $45M, Mitti Labs, WindBorne $37M. Geopolitical risk from entity listing remains. New political framing argues open-weight models enable downloadable frontier agentic capability, shifting value to infrastructure and control plane.

Sources (21)
Updated Aug 6, 2026