At its global meeting held on September 17, 2026, AWS presented a vision for expanding the use of AI across infrastructure layers that include Trainium for computing, Bedrock for inference, and AgentCore for building agents. Matt Wood, AWS’s chief AI and technology officer, said the company wants to distribute AI tools and make them broadly available, in the same way cloud computing helped disseminate computing resources.
The meeting was not limited to presenting the vision; the startup Odyssey gave a live demonstration of world models, systems that learn physics, cause-and-effect relationships, and human behavior by observing large amounts of video, rather than relying entirely on manually written rules.
World Models Go Beyond Game Simulation
Odyssey unveiled Odyssey-3, its latest foundational world model, which was trained and run on Trainium3. According to the presentation, the model can use its learned understanding of the world to control robots, operate humanoid robots, drive vehicles, guide drones, and train other AI systems, in addition to running video games.
The company also announced Agora-2, a multi-agent model that places several intelligent entities inside the same simulated world simultaneously. Wood and Oliver Cameron, Odyssey’s chief executive, entered the simulation directly using two controllers, while attendees watched them interact and fight inside an AI-generated environment. Agora-2 supports up to 20 participants in a single simulation, without using a game engine in the background.
Cameron believes world models may reach a turning point similar to what GPT-3 represented for language models, with potential uses in robotics, autonomous driving, science, healthcare, education, gaming, defense, and energy. This vision is based on the claim that, after learning how the world works, the model can acquire new tasks with less task-specific data than previous systems.
Trainium Attracts AI Startups
AWS said Odyssey-3 runs on Trainium3 with a utilization rate exceeding 80% of the model’s computational operations, which the company described as nearly twice the industry average. These models require continuous computing for weeks, a workload AWS says poses a challenge for general-purpose hardware.
The startups building on Trainium include DeCart AI for real-time interactive environments, Neura Robotics for physical AI, Reactor for world-model infrastructure, Splash Music for music generation, and Poolside for code generation.
What Is Changing in Practice?
The material shows that AWS is trying to present itself not merely as a provider of computing resources, but as an integrated layer extending from chips and infrastructure to inference and agents. This could matter to companies that need to train models over long periods or build systems that interact with dynamic environments. However, most of the performance indicators and figures cited came from AWS or the participating companies, so they should be treated as provider claims until confirmed by independent tests. The presentation alone also does not establish that world models are ready for production use in sensitive sectors, nor does it specify the cost of making these systems available or their commercial terms.