OpenAI and Anthropic spend billions of dollars a year training models like GPT-4 and Claude, but competitive price dumping is making the business around these platforms rather precarious. Aidan Gomez, CEO of competing AI provider Cohere, says that selling access to models is quickly becoming a āzero margin businessā in a podcast appearance on Monday. For now, these AI models cost more than they make.
āIf youāre only selling models, for the next little while, itās gonna be a really tricky game,ā said Gomez in an interview with 20VCās Harry Stebbings. By āselling models,ā he means selling API access to those AI models; OpenAI, Anthropic, Google and Cohere offer this service to developers, and theyāre all facing a similar problem.
āItās gonna be like a zero margin business because thereās so much price dumping. People are giving away the model for free. Itāll still be a big business, itāll still be a pretty high number because people need this tech ā itās growing very quickly ā but the margins, at least now, are gonna be very tight.ā
Companies building AI models on the cutting edge are in fierce competition with each other. The most reliable strategy for improving AI models today is adding more compute, which means cutting big checks to Nvidia for the hardware needed to make AI models a hair smarter. At the same time, thereās a race to the bottom. OpenAI and Google have slashed prices for accessing their AI models in order to retain users ā while Metaās open source models are simply free to license.
āThatās why there is a lot of excitement at the application layer,ā said Gomez, referencing OpenAIās $20 a month ChatGPT subscription. Gomez says Cohereās AI models will be an attractive business in the long term, but products could be a meaningful way to generate revenue until then.
In other words, todayās AI models lose money ā lots of it. While Microsoft and Google can subsidize or simply weather that loss, thatās not usually the case for startups. Cohere is one of the last remaining startups developing frontier AI models, alongside OpenAI, Anthropic and Mistral. Other startups like them ā Inflection, Adept, Character.ai ā have been acqui-hired by large cloud providers, leaving an unprofitable business model husk behind while preserving their powerful technology.
However, Big Tech is kind of eating these new companies alive before they have a chance to become competitors.
āItās really dangerous when you make yourself a subsidiary of your cloud provider,ā said Gomez, noting that venture capitalists just want a nice return, while cloud providers may want something more. āItās just not good business.ā
Companies creating cutting-edge AI models are in an increasingly difficult position. Thereās speculation that innovations in model architecture, data efficiencies or computing power will generate huge returns for these AI models some day. However, thereās no telling when, or if, that day will come. And evidently, not every AI startup today will be around see it.