Uber bets on Amazon's custom chips to boost AI efforts

Uber has accelerated its use of Amazon's Graviton4 processors as well as Trainium3 AI chips in an effort to improve its ride-matching, delivery optimizations, and dynamic pricing through the expanded partnership with AWS. Uber is already processing over 30 million trips daily through AWS, with Graviton managing driver/customer connections in milliseconds, while Trainium is used for training predictive models based on trillions of historical ride data. Initial benchmarks show 40% cost reductions versus GPU alternatives for AI workloads.

Deployment spans Uber's Trip Serving Zones architecture, dynamically scaling compute during surge periods across 10,000 cities. Trainium3 clusters enable continuous model retraining on live data, improving arrival estimates 22% and reducing empty miles 15% through refined dispatch algorithms. Graviton4 powers non-AI workloads like payment processing and mapping at 25% lower power draw.

Amazon positions the chips as Nvidia alternatives for hyperscale inference and training, with Uber joining Anthropic and OpenAI as marquee Trainium adopters. AWS claims Trainium delivers 30-50% better price-performance on transformer models powering recommendation engines and fraud detection. Uber's move coincides with Q1 earnings showing AI investments driving 18% gross booking growth to $37.8 billion.

The strategic shift here is indicative of an evolution in cloud economics as a result of custom silicon being able to meet the demand for Nvidia's product while increasing enterprise lock-in to AWS.

Uber's engineers report that they can train models in only two days, rather than waiting weeks for Nvidia's GPUs to be available. This compressed train cycle enables Uber's engineers to iterate their work much faster than their competitors at Lyft (agentic routing) or DoorDash (predictive logistics).

The implications of this are also seen at all of the hyperscalers (Azure vs. OpenAI inference API, Google Cloud vs. Anthropic's Claude, and AWS, courting independents, via chip specialization). Uber has committed to running approximately 70% of its production workload on AWS and signals the readiness of Trainium for use in enterprises beyond frontier model labs.

The roadmap for Trainium includes integrating with Trainium4 before the end of Q4 to support new multimodal models that will combine rideshare data with weather, event, and social signals. Additionally, Uber has been able to maintain margin (22% adjusted EBITDA) while experiencing headcount growth of approximately 25% in AI teams due to cost discipline.