How Development Costs Shape Open AI Model Competition

Research and development accounts for roughly 80% of computational resources spent building frontier AI models, according to recent studies from Ai2 and Epoch AI — a figure significantly larger than the compute dedicated to final model training. This economic reality creates fundamental differences between how open and closed AI ecosystems operate and compete. The Chinese AI ecosystem has structured itself around open model sharing, enabling rapid learning across competing labs while avoiding duplicated research efforts. This approach mirrors principles found in traditional open-source software, adapted for the unique challenges of training large language models. However, the broader AI industry often misrepresents these dynamics, portraying model development as purely artifact-focused and overlooking the genuine cost reductions that open sharing provides. Many companies pursue hybrid strategies: using open-source tools as foundations but maintaining proprietary internal versions. This practice undermines collective ecosystem benefits. True open alternatives face substantial challenges, such as the absence of mature recipes for large-scale reinforcement learning in mixture-of-experts models. Partially-open tools like Tinker and Prime Intellect's Lab attempt to balance accessibility with commercial viability, yet struggle to maintain momentum without complete transparency. Competitive advantage increasingly comes from vertical integration of hardware, data, and infrastructure, making shared foundation models unlikely across competing labs. Chinese laboratories gain significant advantages through transparent technical documentation and inter-company knowledge exchange. As frontier AI development remains computationally demanding, sustaining open ecosystems may require industry-wide collaboration initiatives to remain financially competitive against closed systems.