Note: Paid subscribers can access voice-overs for paywalled posts in podcast apps by going to Interconnects, tapping Settings, and then managing their description. Tak, fordi I lyttede! The majority of compute required to develop a top-tier frontier model is consumed by R&D efforts, not by the final end-to-end training run of the large model itself. In an open ecosystem like China’s, where all the top players collaborate openly, this creates a potentially significant cost-structure advantage that could allow labs to continue scaling far longer than outside observers expect. Two recent studies—one from AI2 on the development of OLMo 3 and another from Epoch AI analyzing public cost disclosures from frontier labs—estimate that roughly 80% of compute is spent on R&D rather than training the final model (with considerable uncertainty). In a world where research and development dominate total compute usage, the Chinese system is explicitly designed to rapidly learn from peers and avoid redundant spending on research compute or infrastructure. It’s not flawless, but it’s the nearest equivalent to the OSS ecosystem available for developing LLMs. Public discussions of AI have long stressed that models are expensive, leading casual readers to assume this refers purely to compute spent on the final artifact itself — as we saw with DeepSeek V3. This prompted me to revisit a core challenge with open-source AI: unlike open-source software, it lacks the direct feedback loops from users back to the model itself—loops that generate tremendous value through Linus’s Law (“given enough eyeballs, all bugs are shallow”). In open-source software, this self-reinforcing dynamic makes large-scale deployment the cheapest option, because the entire user community collectively shares the burden of fixing bugs and adding features. By contrast, in open-source AI the vast majority of that cost still falls on the model developer.
Interconnects AI