Why Model Prices Keep Falling, and What Cloud Providers and Model Makers Want from Each Other

Originally published in Chinese on May 21, 2024. In May 2024, Chinese cloud providers and model makers were cutting API prices one after another. This piece asks what the cuts are for, and how the companies that sell models and the companies that sell cloud services will relate to each other afterward. What a price cut actually lowers A price cut lowers the cost and the threshold for developers. For products still in an early stage of adding AI, API spend is a real budget item, and the more advanced models cost a lot. Cheaper access lets developers give their products AI capability faster. That is one motive behind the cloud providers' cuts: get products that have no AI to start thinking about what AI could add to them. In effect it converts existing users. Say I already run a note-taking product and a utility product, neither with AI. When the barrier to adding AI is this low, adding it is the obvious first move. The aim is to let existing products add AI. People with a startup idea, or a scenario where they need AI to solve a problem, come second. They are not the main factor behind this round of cuts. For products that start from an API or a capability, the scenario is still defined too broadly and will take longer to settle, and what they already spend on APIs is a psychological cost weighed against a much larger expected return. With that in mind, higher API usage means the services around the API will be used too, and the cost of those services is where cloud providers expect most of their revenue. Later, once API costs reach a basic level of service, the state of play may change again. Basic text APIs could become free while multimodal models are charged for. By "basic service" I mean, for example, that plain speech transmission and speech synthesis are still billed as separate items today. If a packaged model becomes a basic product in another form (see Doubao), that is one more line of thinking and one more opportunity. Where model makers and cloud providers go from here