World Models vs. Transformers: Two $1B+ Bets That the LLM Era Is Ending
Yann LeCun raised $1.03B for AMI and Mira Murati got 1 gigawatt of compute for TML. Both are betting the current AI paradigm needs fundamental challenges.
On the same day the Pentagon tried to ban an AI company, two of the most credible AI researchers alive bet more than 2 billion dollars that the current AI paradigm is wrong.
Yann LeCun, who won the Turing Award for his work on deep learning, raised $1.03 billion for AMI (Advanced Machine Intelligence) to build "world models" — AI systems that understand the structure and dynamics of the physical world, not just pattern-match on text.
Mira Murati, who left OpenAI last year to start Thinking Machines Lab, announced a multi-year partnership with NVIDIA for at least one gigawatt of compute starting early 2027, positioning herself to challenge OpenAI at frontier scale.
Both announced on the same day. Both are explicitly betting against the idea that large language models are the end state of AI. And both are backed by the kind of capital that makes their bets credible.
This is not fringe. This is the mainstream AI research community saying: We think there is a better path.
Yann LeCun's world model thesis
LeCun has been the most public critic of large language models' fundamental limitations. His argument, refined over the last two years: LLMs can predict the next token in a sequence. They cannot understand causality, physics, or the structure of the world.
A person can watch a video of a ball rolling off a table and predict where it will land. An LLM cannot. An LLM can summarize text about gravity. It cannot truly grasp why the ball falls. That gap is the difference between understanding and pattern matching.
LeCun's bet with AMI is that the next frontier is systems that build internal models of how the world works. World models that learn from raw sensory data (video, audio, images) and build causal understandings of physics, objects, agents, and interactions.
The commercial angle: If you can build an AI system that understands the physical world, you can deploy it on robots, autonomous vehicles, manufacturing systems, and any domain where physical understanding matters.
This is not an academic exercise. LeCun raised $1.03 billion at a $3.5 billion pre-money valuation. Investors include Khosla Ventures, Thrive Capital, and others betting on the world model thesis.
Mira Murati's compute-backed bet
Murati left OpenAI as CTO in September 2024, saying she wanted to "explore the field and work on frontier AI in a new way." She founded Thinking Machines Lab (TML) with three goals: build AI systems at frontier scale, operate outside the OpenAI ecosystem, and compete directly with the incumbent frontier labs.
On Monday, she announced that NVIDIA would commit at least one gigawatt of Vera Rubin compute capacity to TML starting early 2027. One gigawatt is a threshold only matched by OpenAI, Google, and Microsoft. Three companies. Now four, if you count TML.
What makes Murati's bet credible is compute. Frontier AI development requires massive compute resources. Murati just secured enough to compete with the incumbents. She is not building in the shadows. She is building at scale, with NVIDIA's explicit backing.
The message to the market is clear: we have the capital, the compute, and the talent to challenge the current frontier labs.
The deeper story: paradigm shifts
LeCun's world models and Murati's frontier labs are not attacking each other. They are attacking the status quo from different angles.
LeCun is saying: The current architecture (transformers optimized for text) is fundamentally limited. The next architecture needs to be built on different principles.
Murati is saying: The current frontier labs (OpenAI, Google, Anthropic) are concentrating too much power and moving in the wrong direction. We can compete by building frontier-scale AI that maintains independence.
Both bets require a belief that the current concentration — OpenAI's dominance — is fragile. That belief has gotten more credible in the last week as Microsoft, 37 researchers, and the entire AI industry lined up against the government instead of supporting the government's preferred vendor.
The vendor consolidation that seemed locked in two months ago is now uncertain.
What this means for enterprise AI buyers
If you locked into OpenAI as your AI platform two years ago, you were betting the transformer architecture would dominate. That bet is looking less certain.
It is not that LeCun or Murati will replace OpenAI in six months. Both are betting on 2027+ timelines. The world model thesis needs years to mature. Frontier-scale independent labs need time to prove they can match OpenAI's capabilities.
But the runway is real. Enterprise AI buyers who make long-term commitments to single vendors should be thinking about their exit strategy. If the AI paradigm shifts, the vendor you chose might become second-tier within five years.
That is not an argument to avoid OpenAI. It is an argument for optionality. Build your AI stack assuming that the current vendor choice might not be the optimal choice in 2028. That means:
Use APIs and standard interfaces, not custom integrations. Make it easy to swap vendors.
Invest in model-agnostic evaluation. Do not build your entire business logic around GPT-specific quirks.
Keep an eye on alternative architectures (world models, reasoning systems, embodied AI). As they mature, experiment.
The concentration erosion
One year ago, the AI market looked locked. OpenAI had GPT-5, Google had Gemini, Anthropic had Claude. The market was consolidating around three vendors.
Today:
- OpenAI is in a legal and political war
- Anthropic is fighting the government and winning industry support
- Google is quietly building alternatives
- Yann LeCun just raised $1B to build a different paradigm
- Mira Murati just secured 1 gigawatt to challenge from the frontier
The monopoly has cracks. The paradigm is uncertain. The vendors are diversifying.
For enterprises, that is good news. It means lock-in is less likely. It means switching costs are lower. It means you have leverage you did not have two months ago.
The LeCun and Murati announcements are not threats to today's AI market. They are bets on what comes next. And they just told the market: next year's AI landscape will look different from today's.
Frequently Asked Questions
Q: Is Yann LeCun saying that transformers are dead?
A: No. He is saying transformers are necessary but not sufficient. Transformers excel at pattern matching in sequences (like text). They cannot reason about causality or physics. World models are the next layer needed on top of or alongside transformers to handle those problems.
Q: Can Mira Murati actually compete with OpenAI?
A: With 1 gigawatt of compute and top-tier talent, she has the resources to build frontier-scale systems. The question is whether she can match OpenAI's research efficiency and product execution. Compute alone does not guarantee competitive advantage — but it is a necessary condition.
Q: Should I switch away from my current AI vendor because of these announcements?
A: No. Both LeCun and Murati are betting on 2027+ timelines. Today's AI decisions are still solid. But if you are making a multi-year commitment, ask your vendor about their research roadmap. Are they investing in world models? In reasoning? In alternatives to pure scaling? The vendors who are exploring alternatives will age better.