The Private AI Model Explosion

Lost in the AI haze is a simple idea:

Private AI models are ultimately going to generate more revenue than the public models.

What’s the difference?

Public models — Google’s Gemini, OpenAI’s ChatGPT, Anthropic’s Claude, et al. — currently hold the world’s attention. They are all trained on the same information (the internet), the models are vast (up to trillions of tokens), they are attracting hundreds of billions of dollars of investment, and their capabilities and technology are advancing quickly.

The current narrative is that these models will subsume all of the world’s information and two or three of them will end up dominating the market — becoming as powerful in AI as Google is in search, Meta in social, or Amazon in cloud. It will be winner take all.

That’s not going to happen.

In five years, 70% of the revenue created by AI will be in private models, not public models. What is a private model?

First, let me ask a few simple questions. Do you think that your bank account is going to end up in a public model? Do you think your insurance policy will end up in a public model? Do you think the payroll system of your company will ever live inside in a public model? The answer to all three questions is of course “no.”

Where will they end up? In private models created by your bank, your insurance carrier, and your company. Why? In one word, trust. When you want to “converse” with your bank account, you will engage the Bank of America model — because you trust that company with your personal financial data and because of the decades of proprietary knowledge and experience that they have trained into their AI system.

You: “Hey, how much is in my checking account?”

BofA model: “$2,340.”

You: “Can I use $340 to pay off my credit card?”

BofA model: “Sure, should I do that right now?”

You: “Yes.”

BofA model: “Judging from your typical behavior and our experience working with you, you’re probably not going to need $1,000 for the next 30 days. Should I move those dollars into your money market account so you can get a higher return?”

You: “Yes, good idea. Please do it now.”

BofA model: “Got it. After paying your credit card and moving the $1,000 to your money market, you now have $1,000 available in your checking account.”

Companies are going to replace their websites with AI. When you get your car fixed, you’ll talk to BMW’s AI to troubleshoot your problem and schedule your repair appointment. When you buy a plane ticket to Tokyo, you’ll converse with the Japan Airlines model. When you buy a book, you’ll talk to Amazon’s model.

It’s already happening. I was at my doctor’s office last week, and he reported that his practice is using two private models. The first is OpenEvidence, a model for healthcare professionals that’s trained on data from the New England Journal of Medicine, the Journal of the American Medical Association, and other trusted, peer-reviewed scientific sources. It’s used for diagnosis, diet recommendations, pharmacological efficacy data, and other up-to-date medical knowledge. The second model is Abstractive Health, which summarizes patient medical records, enabling doctors to converse with their patients’ medical history. The two private models are making my doctor more efficient and smarter.

Despite all the attention being lavished on public models, that’s not where the the majority of revenue is going to come from. It’s going to be in private models.

That’s not to say that the public models will not have a role in the future. They will serve three critical purposes:

They will lead in AI innovation — the cool new features and capabilities will show up there first.
They will be the best source of general knowledge. In this respect, they will replace Google search and will end up driving lucrative advertising and commerce models.
They will be the foundation models for the private models.

What does number three mean? For the next several years, private AI systems will be built using context engineering techniques like retrieval-augmented generation (RAG) and post-training approaches like fine-tuning. Public models will undergird many of the private models, yielding syntax (the ability to read and write) and reasoning (the ability to think). This means that while the private models will be able to read and write as well as the public models, they will keep their customer data separate and protected. With the use of proposed services like Anthropic’s Model Context Protocol (MCP) and Google’s Agent2Agent (A2A), it will be possible (and expected) for private AI systems and public AI systems to interact dynamically — the best of both worlds.

What does all of this mean?

There is a vast misallocation of capital occurring at the moment — too much money funding public models while private models are underfunded. Smart investment should be seeking out companies that are sitting on big piles of data that will become more valuable once converted to work in a private model. Think financial data, customer data, marketing data, transaction data, supply chain information, medical data. Consider Colony’s Law (yes, that’s how modest I am): “With AI, information will double in value every year.” This will drive the value of the companies that own that data.
The business model of public LLMs is less clear than the business models of private LLMs. Yes, they will make money by serving consumers general data (likely with an advertising or commerce scheme, to be determined) and serving businesses as the foundation for their private models. But this may not be enough to generate the returns on the hundreds of billions of dollars of startup investment. As Forrester analyst Rowan Curran likes to say, “a language model is not a business model.”
There’s another possible outcome, proposed by Ted Schadler of Forrester. The public models, in their search for revenue, will turn to “hosting” private models for companies. Under this scenario, the public models will start to look like old-line enterprise software companies, taking in proprietary data, fine-tuning a version of their public model to be private, and then running that model for the customer. OpenAI and Anthropic would end up looking a lot like Oracle or SAP, building systems and charging companies a monthly fee to run those systems — back to the future.
Companies must start building their private models now. Yes, AI will help organizations be more efficient and build and run processes faster. But this is a sideshow. The real AI game will be winning, serving, and retaining customers. And that will be the sweet spot of the private-model business model.