World model funding surges past $2 billion（約3200億円） in weeks, but a May 25 proof reveals the dominant approach has fundamental limits in real-world conditions.

What happened: World Labs raised $1 billion（約1600億円） in February at a reported valuation near $5 billion（約8000億円）, and AMI Labs raised $1.03 billion（約1600億円） weeks later at a $3.5 billion（約5600億円） pre-money mark, described as the largest seed round in European history. On May 25, researchers including Yann LeCun (executive chairman of AMI Labs) published a proof showing that the leading statistical method for building world models—systems trained to predict how the physical world evolves—recovers true reality only when variables follow a bell curve and drift in a gentle way; most physical systems break both conditions.

Why it matters: The capital influx reflects investor conviction that world models (AI systems that learn to predict and reason about physical environments) represent what comes after large language models. However, the timing gap is stark: a billion dollars moved on this thesis in the same season a primary proof drew a hard line under the math. The proof, authored by one of the funded founders, identifies a risk factor now cited in a primary source but not yet priced into venture deals. Separately, a May 20 benchmark found that one leading model planned correctly about half the time in clean conditions, then dropped to about 12 percent when the agent changed color and to about 6 percent when the background changed—suggesting that demo-fidelity is a poor proxy for actual planning capability.

What to watch: The field is racing toward domain-specific reliability rather than general-purpose systems. Synthetic-data world models like NVIDIA Cosmos are production-grade today for generating training data in robotics and autonomous vehicles; generative-3D systems like Marble and Genie 3 are production-grade for content and simulation. However, a model you trust to reason about your physical plant and act on it over long horizons is not yet available to buy. The most defensible position is durable accuracy over long rollouts inside one physical domain—autonomous driving, robotics, or industrial simulation—rather than trying to out-general the well-capitalized giants.