Summary: Uber has extended its AWS agreement to utilize Amazon’s Graviton4 processor for its real-time ride-matching infrastructure and is testing AI model training on Trainium3, joining companies like Anthropic, OpenAI, and Apple as part of a growing list of clients benefiting from Amazon’s custom silicon strategy.
Uber operates on extremely tight timeframes, processing over 40 million trips daily across 72 countries by 2025. Its system, known as Trip Serving Zones, rapidly identifies suitable drivers and ensures quick service. Announced on April 7, 2026, Uber is transitioning more of this workload to AWS using Graviton4 processors and initiating a pilot program to train AI models on Trainium3. This move signifies a key win for Amazon’s custom silicon initiative.
Uber’s Transition and Rationale
The Graviton4 processors will support Trip Serving Zones, enhancing the system’s ability to handle high ride volumes efficiently without delays. Meanwhile, Uber is piloting AI model training with Trainium3, leveraging data from its extensive trip history. The company’s database includes over 13.567 billion trips from 200 million monthly active users, providing a rich source of data for AI applications. Economic advantages make utilizing Trainium3 an attractive option.
Kamran Zargahi, Uber’s VP of engineering, explained the importance of rapid operations: “Uber operates at a scale where milliseconds matter. Transitioning more tasks to AWS allows faster rider-driver matches and handles demand spikes smoothly.” Rich Geraffo from AWS highlighted Uber’s real-time requirements: “Uber represents one of the most demanding real-time applications globally, and we’re proud to be a critical part of their infrastructure.”
Uber’s Cloud Strategy
This AWS partnership is part of Uber’s broader cloud strategy, which already includes agreements with Oracle Cloud and Google Cloud to replace its own data centers. By engaging multiple cloud providers, Uber seeks to prevent dependency on a single vendor and to select the best platform for different workloads. This approach grants Uber leverage in negotiations and flexibility in resource allocation. The Graviton4 addition highlights AWS’s current competitive position in high-frequency, latency-sensitive tasks, while the Trainium3 pilot evaluates its competitiveness against existing GPU-based solutions.
Details on the Trainium3 Chip
Amazon’s Trainium3, an advanced AI training accelerator, stands out with 2.517 petaflops in MXFP8 precision, along with 144 GB of HBM3e memory and 4.9 terabytes per second of memory bandwidth. This chip operates at 30-50% of the cost of competing Nvidia hardware. The UltraServer setup can network 144 accelerators, achieving about 362 MXFP8 petaflops, suitable for large-scale model training.
While Nvidia hardware becomes standard for production environments, where interoperability is crucial, Trainium3 offers cost benefits for controlled training contexts, as the expenses accumulate over numerous experiments. Recent advancements in Trainium tooling make it an increasingly viable option.
Expanding Clientele for Trainium
With Uber now part of its client list, Amazon strengthens its Trainium offering. Leading companies like Anthropic, OpenAI, and Apple have capitalized on Trainium’s cost efficiency for substantial AI workloads. The increasing demand for sustainable compute costs across AI-dependent organizations makes Trainium attractive. For Amazon, each new Trainium client not only validates the chip but also enhances its ecosystem, easing future adoptions. Uber’s training use case stretches Amazon’s capabilities further, broadening their application scope.
Implications for Nvidia
Amazon’s focus on custom silicon is partly driven by dissatisfaction with GPU reliance, of which Nvidia is a major provider. Uber’s pilot may indicate a shift, showcasing real-world trials of alternatives to Nvidia’s ecosystem. Nvidia’s NVLink Fusion strategy aims to bridge third-party silicon with their network, allowing integration while maintaining their ecosystem’s dominance. Whether Uber fully transitions its AI training to Trainium depends on pilot outcomes and Amazon’s advancements toward matching Nvidia’s CUDA ecosystem. This testing represents a serious exploration of Nvidia alternatives at scale, a significant development given the industry’s reliance on Nvidia thus far.