All Models Eat the Internet; Larry Ellison Says the Valuable Half Is the Other Half

In an interview, Larry Ellison put the value of AI very directly. ChatGPT, Anthropic, Grok, and Llama have all been trained on data from the internet, but for models to reach their peak, private data needs to be connected as well.

Nicolás Echeverría pushed the point further: foundation models are turning into public commodities, and proprietary data pipelines are the real moat. Training on web text leveled the playing field; the winners may not have the best general-purpose models, but they have clean, secure access to proprietary operational data.

Linas Beliūnas summarized it even more briefly: the same internet, the same data, increasingly similar models. The advantage is not in the model, but in what you can let it see. The original tweet is here: link

This explains why the hotter large models get, the more valuable enterprise data governance becomes. Public corpora are consumed over and over, and the gap between models will shrink. Private data is not on the public internet; it lives in ERP systems, customer service records, contracts, codebases, tickets, and internal documents. To feed this data to models, cleaning, permissions, auditing, anonymization, and retention are all unavoidable. Once model capabilities are leveled, the valuable part is shifting away from the model itself and toward data integration and governance.

Some commenters are not buying it. OdinsFar compares connecting private data to letting the fox into the henhouse: either hand over financial accounts, divorce files, and offshore assets to the model, or admit that AI is just a tool used when humans think there is a payoff. This concern is not absurd. Once data is connected to a model, there must be answers to who can see it, for how long, how it is deleted, and who is responsible if something goes wrong. Gravity Research left just one line: good luck getting legal to sign off.

Another replier, NotesfromShelter, shared a more concrete experience. In an information retrieval class, they learned how database vendors sell; when old content wanted open access, they were told to discuss it offline. Four images were attached:

Linas also shared an old story. Steve Jobs once had Ellison watch 73 versions of Toy Story, just to show improvements in the shadows. In the end, Ellison said he knew the new version of RenderMan was 4% better than the previous week's version, but he did not want to watch anymore.

This anecdote fits well today. Model version iterations often revolve around improvements of a few percentage points, but where enterprises are truly stuck is whether private data can be connected securely. Someone commented that control is an illusion, and Linas replied, wasn't it always like that?

When model capabilities are leveled, whoever can safely feed private data to models will hold the next layer of pricing power. Only one question remains: your data, who do you dare let see it?