Uber is expanding its AWS partnership to run more ride-sharing infrastructure on Amazon's custom Trainium chips, marking another win for Amazon's strategy to compete with Nvidia through purpose-built AI silicon. The expanded contract moves additional Uber workloads off Oracle and Google Cloud onto AWS's custom hardware, though specific workloads and financial terms weren't disclosed.
This matters because it's validation that Amazon's multi-billion dollar bet on custom chips is actually working in production. While everyone obsesses over Nvidia's GPU monopoly, Amazon has been quietly building an alternative with Trainium for training and Inferentia for inference. Uber's expanded commitment suggests these chips can handle real-world ML workloads at scale—not just AWS marketing demos. It's also a strategic blow to Oracle and Google, who've been trying to win back enterprise workloads with their own AI infrastructure plays.
The move fits Uber's broader pattern of consolidating on fewer cloud providers while demanding better economics for AI workloads. Uber processes massive amounts of real-time data for pricing, routing, and matching—exactly the kind of inference-heavy workloads where custom silicon can deliver cost advantages over general-purpose GPUs. What's unclear is whether Uber is using Trainium for training new models or just running inference on existing ones.
For developers, this signals that Amazon's custom chips are production-ready for demanding ML workloads. If you're building on AWS and dealing with high inference costs, Trainium and Inferentia instances might be worth testing. But the real story is infrastructure consolidation—betting on one cloud provider's full AI stack instead of mixing and matching across vendors.