One post examines AI “neolabs,” loosely defined as research startups raising large sums before production to fund compute. The author sees post-training a strong open-source model as a possible route to frontier performance, but warns of millions in spending on reinforcement-learning environments and the risk of being overtaken by a newer model while tied to a base model. In the post’s example, recovering $10 million in training costs at a 50% inference margin and a blended price of $2 per million tokens would require serving about 10 trillion tokens. The author suggests alternatives to direct competition, including different kinds of models or proprietary datasets large and useful enough to surpass frontier quality in a domain. Even then, the argument goes, demand and revenue relative to compute costs must justify the investment.