AI Broke Software's Best Trick (ardonio.com)

A recent analysis highlights a fundamental shift in the economics of artificial intelligence (AI), revealing that the existing software model of “zero marginal cost” is no longer viable for generative AI (genAI) applications. Unlike traditional software, where the cost of adding users is negligible, genAI incurs substantial costs with each interaction due to the high computational demands of model training and inference. This shift creates a challenging landscape for AI companies, as the current subscription models fail to cover the rising costs associated with hosting powerful datacenter hardware needed to support user engagement. Power users are often subsidized by a larger base of less active users, making profitability difficult to achieve. To address these economic strains, industry experts suggest two potential solutions: increasing prices for high-ROI tasks, or extending the lifecycle of older models by routing simpler tasks to less intensive models. This evolution hints at a return to software distribution models where users can leverage their own hardware for running AI applications, thereby reducing usage costs and making profitability feasible. As frontier models evolve and become more efficient, the transition to running sophisticated AI locally could not only democratize access but also shift the economic burden away from centralized labs, leading to a newly structured AI landscape.