The AI startups building beyond chatbots are choosing Amazon’s custom chips. Here's why.

Key takeaways AI startups building world models (which simulate physics, rather than generating text) are choosing AWS Trainium over other chips for training. Odyssey achieved 80% model flop utilization on Trainium, roughly double the industry average of 40-50%.

AWS offers both Trainium and Nvidia GPUs, letting customers choose the best infrastructure for their workload. Some of the biggest names in AI are already building on Amazon’s AI chips.

Anthropic trains and runs models on AWS Trainium chips, and OpenAI has committed to consume approximately 2 GW of future Trainium capacity as part of multi-year partnerships with Amazon. But a different kind of customer is emerging on the platform—one that reveals something about where AI is heading that the chatbot era alone can’t show.

Uber scales on AWS to help power millions of daily trips and train its AI models The ride-sharing giant expands its real-time infrastructure on AWS to speed up service for millions of daily riders and deliveries. A growing cohort of AI startups is choosing Trainium to train models that don’t generate text.

They generate physics, environments, and interactive simulations of the real world. These are called world models, and they represent one of the most compute-intensive frontiers in AI. What are world models and why do they need different infrastructure? World models are AI systems trained to simulate how the physical world behaves.

Rather than predicting the next word in a sentence, they predict the next frame of a scene—accounting for gravity, light, motion, and the interactions between objects. The applications range from robotics and autonomous vehicles to game engines and industrial simulation. Training these models requires enormous, sustained compute.

Unlike large language models, which can be trained in bursts, world models demand long, uninterrupted runs at high utilization—making the cost-per-useful-compute the defining metric for the companies building them.

How AI startup Odyssey achieved unprecedented compute efficiency on AWS Trainium Odyssey, a startup building world models that simulate physics, recently achieved 80% model flop utilization (MFU) on Trainium3 —a metric that measures how much of a chip’s theoretical peak performance is actually realized during a real workload.

In an industry where 40 to 50% MFU is considered well-optimized, 80% is exceptional. It means Odyssey extracts nearly twice the useful compute per dollar compared to typical infrastructure. 10 AI chip terms you should know AI runs on chips. Here’s what the most important terms mean and why they matter.

Ron Diamant—the vice president and distinguished engineer overseeing Amazon’s work on Trainium—called Odyssey’s team “very, very impressive,” noting their ability to optimize their world model on Trainium with minimal support from Amazon’s side. “They just went ahead and did it,” he said.

Why Amazon designed Trainium as a general-purpose AI accelerator Trainium wasn’t built for a single model architecture. Amazon’s chip team studied a range of workloads—transformers, vision encoders, diffusion models, and world models—then generalized the underlying compute primitives into a flexible instruction set.

“We're not building a transformer or world-model accelerator, that's not our approach. We study these workloads, work backwards to the primitives required to run them fast, and then generalize an instruction set for general-purpose compute that still accelerates these workloads exceptionally well", said Diamant.

That design philosophy is paying off as new customers arrive with novel architectures. Each world model has slight deviations from the last, and Trainium’s generalized approach means startups can achieve high performance without extensive custom optimization. “We're not building a transformer or world-model accelerator, that's not our approach.

We study these workloads, work backwards to the primitives required to run them fast, and then generalize an instruction set for general-purpose compute that still accelerates these workloads exceptionally well.”

Ron Diamant Vice President & Chief Architect, Trainium, Amazon Why sustained AI chip performance matters One of the less obvious advantages Diamant highlighted: Trainium can sustain 80% utilization over long training runs without overheating—a challenge that limits many competing chips.

Diamant explained that Amazon invests “across the stack, from the software to the thermal solution and the power delivery solution” to ensure Trainium can sustain high utilization over long inference or training runs, a challenge he says many competing chips can’t match.

For world model companies that need to scale compute to serve many customers cost-efficiently, this sustained performance translates directly to economics. How the next generation of AI researchers is using Amazon chips to accelerate discovery University researchers are using Amazon’s Trainium chip to push the bound