The bottom line: New studies show: 80% of AI compute costs go to research, not training. China’s open ecosystem leverages this structure for significant cost advantages – a benefit open-source AI lacks due to the absence of direct user feedback unlike traditional software.
In open AI ecosystems like China’s system, substantial cost advantages emerge because research and development spending dominates – not the final training phase. This structural advantage could enable labs to scale longer than outside observers expect.
Approximately 80 percent of computing capacity for developing leading AI models is expended on research and development, not on the final training itself – as shown by recent studies from AI2 on the development of OLMo 3 and from Epoch AI on publicly disclosed costs. In an ecosystem like China, where leading players collaborate openly, this creates significant competitive advantages: the system is explicitly designed to learn quickly from competitors and avoid redundant expenditures on research resources.
Unlike open-source software, open-source AI cannot rely on direct user feedback loops – those mechanisms that function in traditional software through “Linus’ Law”: “Given enough eyeballs, all bugs are shallow.” While in open-source software the entire user community collectively contributes to bug fixes and further development, these cost savings do not accrue to AI models. Here, the model developer still bears the majority of the burden.