OpenAI and Anthropic spend billions of dollars a year training models like GPT-4 and Claude, but the price wars have made business around these platforms rather precarious. Selling access to the models is quickly becoming a “zero-margin business,” Aidan Gomez, CEO of rival AI provider Cohere, said in a podcast appearance on Monday. For now, the costs of these AI models are outpacing the revenue they bring in.
“If you're just selling models, that's going to be very hard for the foreseeable future,” Gomez said in an interview with 20VC's Harry Stebbings. By “selling models,” he means selling API access to AI models. OpenAI, Anthropic, Google, and Cohere all offer this service to developers, but all face similar problems.
“There's a lot of price dumping going on, so it's going to be a zero-profit business. People are giving the models away for free. It's still a big business, and people need this technology, so there will be significant numbers. It's growing fast. But the profit margins will be very tight, at least for now.”
Companies building cutting-edge AI models are in fierce competition with each other. Currently, the most promising strategy for improving AI models is to buy more and more computing power, which means spending a lot of money on Nvidia to buy the hardware needed to make the AI models a little smarter. At the same time, there is a price war going on: OpenAI and Google have significantly reduced the price of access to their AI models to retain users, while Meta's open source models are licensed free.
“That’s why there’s so much excitement at the application layer,” Gomez said, referring to OpenAI’s $20-per-month ChatGPT subscription. Gomez said Cohere’s AI models will be an attractive business in the long term, but until then the product could be a meaningful way to generate revenue.
In other words, today's AI models are losing a lot of money. While Microsoft and Google can cover their losses or simply weather them, this is usually not the case for startups. Cohere is one of the few startups developing cutting-edge AI models, along with OpenAI, Anthropic, and Mistral. Other startups, such as Inflection, Adept, and Character.ai, have been acquired by major cloud providers, leaving behind a shell of a less profitable business model and maintaining a strong technology.
But big tech companies seem to be gobbling up these startups before they have a chance to become competitors.
“Being a subsidiary of a cloud provider is very risky,” Gomez said, noting that venture capitalists are simply looking for high profits, while the cloud provider may want more than that. “It's never good business.”
Companies that create cutting-edge AI models are in an increasingly difficult position. There is speculation that innovations in model architecture, data efficiency, and computing power will enable these AI models to generate huge profits in the future. But it's unclear when or if that day will come. And, obviously, not all of today's AI startups will see that day.