The $100 billion poured into scaling language models may be the most expensive wrong turn in tech history.

Here’s an unsaid fact: ChatGPT doesn't “understand” anything.

It predicts the next word. Give it text, it returns text, and it's extremely good at that one specific task. But the real world is made of continuous sensor streams, camera feeds, pressure data, temperature readings, and motion signals. These streams change every millisecond in ways that are physically impossible to predict frame by frame, the way you predict the next word in a sentence.

You can't compress a moving robot arm into tokens. The physics of reality doesn't work that way.

And that's the exact problem Yann LeCun just raised $1.03 billion to solve.

LeCun is one of three scientists who won the Turing Award for inventing deep learning techniques that power virtually every AI system in use today. He then spent a decade at Meta running FAIR, their Fundamental AI Research lab, the team responsible for Meta's core AI breakthroughs.

In November, he walked into Zuckerberg's office, said he was leaving, and four months later had taken half of Meta FAIR with him, including Meta's former research director, a Google DeepMind lead, and Meta's senior AI research director.

AMI Labs is four months old. 

The CEO said publicly, on the record, that they won't ship in three months, won't have revenue in six months, and won't hit $10M ARR in twelve months. He called it "a long-term scientific endeavor."

Investors gave him a billion dollars anyway.

Why?

Because the investors aren't betting on a product. They're betting on a ceiling. 

NVIDIA, Samsung, Toyota Ventures, Dassault – all these are companies that need AI to understand physics and geometry. A language model that writes poetry is useless to a robotics engineer trying to predict what happens when a mechanical arm applies 12 newtons of force at a 30-degree angle to a flexible surface.

Language models fail here because they are architecturally incapable of doing so. They generate statistically plausible outputs.

AMI is building what are called world models using an architecture LeCun proposed in 2022, called JEPA (Joint Embedding Predictive Architecture). Instead of predicting exactly what every pixel of the next video frame will look like, JEPA learns a compressed, abstract representation of reality and predicts within that space.

Think of it as the difference between memorizing every word of a conversation versus actually understanding what was being discussed. Action-conditioned versions then let an AI simulate the consequences of its own actions before taking them.

Meanwhile, two weeks before LeCun's raise, Fei-Fei Li closed $1 billion for World Labs on the same thesis. Same investors, Nvidia, Sea, and Temasek backed both.

This is a coordinated conviction.

Now, what if he's right?

Every company building physical-world products on top of language models may have bought the wrong foundation entirely, not a weak foundation. Language models cannot learn from sensor data the way the physical world produces it. You cannot patch your way out of an architectural mismatch.

If JEPA works at scale, the next generation of AI companies won't fine-tune language models at all. They'll train world models on real-world sensor data from scratch. And the $100 billion already poured into scaling language models will look like the most expensive wrong turn in the history of technology.

The architecture war is just starting. And that's worth watching closely.

In 2024, Sam Altman called ads "uniquely unsettling." In 2026, ChatGPT got ads. However, OpenAI is projected to lose $14 billion this year alone. Ads were never the plan. They became the only plan.

Meanwhile, Perplexity, who tried ads first, made only $20,000 all year. Anthropic saw all this coming and bought Super Bowl spots to say: "Ads are coming to AI. But not to Claude."

The whole industry just picked a side. Reach vs. trust - both can't win the same way!