IA · 18 June 2026 · 3 min read
The Era of World Models: How Physical Simulation is Redefining AI and XR
In brief: AI startup Odyssey has raised $310 million at a $1.45 billion valuation to develop 'world models' that simulate physical reality in 3D, bypassing the limits of standard 2D video generators. A strategic alliance with Amazon includes model optimization for Trainium chips, highlighting the growing trend of hardware diversification in the enterprise AI space.
by Team Mocchi's
A New Dimension for Artificial Intelligence
The landscape of artificial intelligence is moving rapidly beyond the generation of two-dimensional text and images. While large language models and image generators have shown extraordinary capabilities in manipulating symbols and pixels, the new frontier lies in understanding and accurately simulating the physical world. This transition is driven by "world models"—systems capable of understanding not just how a scene looks, but how it works and how it reacts to the laws of physics.
Confirming this trend, one of the most innovative startups in the sector, led by pioneers of autonomous driving systems, has just closed a major $310 million funding round, reaching a valuation of $1.45 billion. The deal saw participation from major tech infrastructure and chip players, demonstrating how the industry is betting heavily on this technology to unlock the next generation of applications in custom software, robotics, and virtual and extended realities.
What is a "World Model" and Why It Changes Everything
Until now, most AI systems capable of generating video relied on diffusion algorithms that "guess" the next pixel to create a coherent visual sequence. The limit of this approach is obvious: while producing spectacular visuals, the AI has no real understanding of gravity, object solidity, or geometric proportions. Consequently, visual hallucinations frequently occur, such as objects clipping into each other or physically impossible movements.
World models address this issue from an opposing perspective. Instead of merely imitating training images, these systems reconstruct a three-dimensional, mathematically consistent representation of space, integrating real physics.
To train these models, the development team adopted a radical, boots-on-the-ground approach: sending human operators equipped with specialized high-tech backpacks—loaded with advanced sensors, high-resolution cameras, and spatial mapping systems—to scan the real world with millimeter precision. This massive collection of high-fidelity data allows the algorithm to learn the actual geometry of our planet, creating interactive 3D simulations that strictly respect physical rules, such as light, gravity, and collisions.
The Hardware Battle and Chip Diversification
Training and running world models require unprecedented computational power, even higher than that needed for traditional language models. It is no surprise, then, that the recent funding was backed by prominent silicon manufacturers and cloud computing giants.
A crucial aspect of this evolution involves hardware diversification. The startup has designated the cloud infrastructure of a major industry leader as its preferred partner, optimizing its models to run on proprietary AI training chips known as Trainium. This move represents a cornerstone in a broader industrial strategy aimed at reducing dependence on the standard graphics processors that dominate the market, proving that training enterprise-grade AI can succeed on alternative, optimized hardware architectures.
Implications for Extended Reality, Robotics, and Custom Software
The practical applications of this technology are revolutionary and directly impact the sectors in which we operate as a software agency.
- Extended Reality (XR) and Gaming: Creating virtual worlds will require a fraction of the current time. Developers will be able to generate entire interactive, physically accurate 3D environments from simple text descriptions or real-world footage, eliminating manual 3D modeling bottlenecks.
- Industrial Robotics and Drones: Before deploying a robot or drone in the real world, companies can train it inside virtual simulations indistinguishable from reality, eliminating physical collision risks and speeding up development by months or years.
- Digital Twins and Enterprise Software: For industrial enterprises, the ability to create ultra-realistic "digital twins" of their facilities will allow them to plan logistics, predictive maintenance, and emergency scenarios with unprecedented geometric and physical accuracy.
The transition from language models to world models marks the beginning of a new era in computing. It is no longer just about processing information and text, but about giving technology the tools to interpret and simulate with absolute precision the complex reality in which we live.