Open-source models reduce costs by 40 times, threatening the revenue of top AI labs

ME AI message: According to monitoring by Beating, after the simultaneous release of 25 open-weight models including NVIDIA Nemotron 3 Ultra and Google Gemma 4 12B, tech investor Chamath Palihapitiya noted that the performance gap between open-source and proprietary frontier models is rapidly narrowing, yet significant disparities remain in invocation costs. Estimating enterprise usage at 1 billion input and 1 billion output tokens per month, GPT-5.5 Pro costs $105,000, Claude Opus 4.8 costs $30,000, while DeepSeek R1 requires only $2,740—meaning GPT-5.5 Pro carries a premium nearly 40 times higher. Chamath warned that most company CEOs are unaware that their engineering teams are defaulting to the most expensive large model APIs without proper governance or auditing, leading to budget overruns. As model routing control planes such as Software Factory, the AI-native software development lifecycle orchestration platform co-founded and led by Chamath himself, become more widespread, enterprises will shift toward model-agnostic architectures, routing bulk inference to DeepSeek, directing high-end agent workflows to Claude Opus, and invoking GPT-5.5 Pro only on demand when clear, substantial incremental value is generated. Chamath predicts that this granular routing will significantly slow revenue growth for frontier labs like OpenAI and Anthropic, while revenue for open-source and low-cost inference ecosystems will surge. (Source: MLion)