Summary
Matthew Berman breaks down what he calls the most important trend in AI right now: the accelerating shift from closed-weights to open-weights models in production workloads. The centerpiece is a Vercel platform chart tracking token consumption from June to August 2026, which shows open-weights models — led by Chinese releases like DeepSeek and Kimi — steadily taking share from proprietary models like Claude and GPT.
The data reveals a striking split between usage and revenue. DeepSeek has surpassed Anthropic in raw token share (25.2% vs. 24.5%), yet Anthropic still captures roughly 64.6% of total model spend compared to DeepSeek’s 2.8%, because Anthropic’s pricing sits at the premium end of the market. Berman argues this shows that frontier capability — even a few benchmark percentage points — commands a massive valuation premium, with Anthropic reporting an annualized run rate above $65 billion and OpenAI around $40 billion, dwarfing every open-model provider combined.
The video also documents which major enterprises are quietly building on Chinese open-weights models: Harvey (legal AI) on Kimi K3, Cursor on Kimi K2.5, Airbnb on Qwen, and Perplexity on DeepSeek. Berman frames fine-tuning on proprietary data — Harvey’s customized legal benchmarks are cited as an example — as the key reason enterprises choose open weights beyond cost alone: privacy, control, and the ability to specialize. The episode closes with a caution about measuring cost per completed task rather than per token, using a Kimi K3 vs. GPT comparison to illustrate why raw token pricing can mislead.
📺 Source: Matthew Berman · Published August 26, 2026
🏷️ Format: News Analysis
