Insiders say the future of AI will be smaller and cheaper than you think
HSBC’s latest analysis of the financial challenge facing OpenAI shows just how big-scale the company’s thinking is. It already claims revenues of $20 billion. It has committed $1.4 trillion to build new data centers that will feed its ChatGPT interface. Even if it manages to generate revenues of more than $200 billion by 2030, it will still need $207 billion in additional financing to survive.
Those are huge amounts.
But there are a dozen or so AI insiders they spoke to luck Recently at the Web Summit in Lisbon, I described a different future for AI. This future, they say, is characterized by much smaller AI operations that often revolve around AI “agents” performing specialized, specialized tasks, and thus do not need the large, bulky language models that power OpenAI, Google’s Gemini, or Anthropic’s Claude.
“Their assessment is based on the best and the best, and that’s not necessarily the case,” Babak Hodjat, chief AI officer at Cognizant, told Fortune.
“We are using large language models. We don’t need the largest models. There is a threshold at which a large language model is able to follow instructions in a limited domain, and be able to use tools and actually communicate with other agents,” he said. “If this threshold is exceeded, that is enough.”
For example, when DeepSeek introduced a new model last January, it sparked a sell-off in tech stocks because it reportedly cost only a few million dollars to develop. Hodgate said he was also working on a model that used fewer parameters per request, which was much smaller than OpenAI’s ChatGPT, but was relatively capable. Once they get below a certain size, some models don’t need data centers, they can run on a MacBook, he said. “This is the difference, this is the trend,” he said.
A number of companies orient their services around AI agents or applications, assuming that users will want certain applications to do specific things. Superhuman — formerly Grammarly — runs an app store filled with “AI agents that can sit in the browser or in any of the thousands of apps that Grammarly has already got permission to run,” according to CEO Shishir Mehrotra.
At Mozilla, CEO Laura Chambers has a similar strategy for Firefox. “We have some AI features, like Shake to Summarize, smart tab grouping on mobile devices, link previews, and translations that all use AI. What we do with them is we run them all locally, so the data never leaves your device. It’s not shared with models, it’s not shared with MBA students. We also have a little slideshow where you can choose your model that you want to work with and use AI in that way,” she said.
At chipmaker ARM, chief strategy/chief marketing officer Ami Badani told Fortune that the company was out of compliance with models. “What we’re doing is creating custom extensions on top of LLM for very specific use cases. Because obviously those use cases vary greatly from company to company,” she said.
This approach — AI agents that are largely focused on acting like separate companies — stands in contrast to massive, general-purpose AI platforms. In the future, a source asked luckwill you use ChatGPT to book a hotel room that suits your specific needs — perhaps you want a room with a bathtub instead of a shower, or a west-facing view? — Or will you use a specialized proxy that has a deep database underneath only Contains hotel data?
This approach attracts serious investment money. IBM Ventures, a $500 million AI-focused venture fund, has invested in some unattractive AI efforts that fill obscure institutional niches. One such investment is in a company called Not Diamond. This startup notes that 85% of companies using AI use more than one AI model. Some models are better than others at different tasks, so choosing the right model for the right task can become an important strategic choice for the company. Not Diamond creates a “template router,” which automatically sends your task to the best template.
“You need someone to help you figure that out. At IBM we believe in a fit-for-purpose modularization strategy, which means you need the right model for the right workload. When you have a modular router that is able to help you do that, it makes a big difference,” Emily Fontaine, project lead at IBM, told Fortune.
2025-12-01 08:28:00
