Accelerating AI innovation for over 25 million vehicles using AWS with BMW

BMW has worked alongside AWS for a long time to power the next generation of vehicles with AI and machine learning (AI/ML) capabilities. The company’s ConnectedDrive Division delivers AI-powered digital services and connectivity features to BMW’s vehicles and mobile apps. The division provides customers with vehicle-related services, entertainment features, and smart functionalities through cloud-connected solutions.

As AI use cases grew, teams across BMW were trying to establish consistent application development and deployment capabilities with advanced security features. They also needed to manage complex networking requirements and provision computational resources for large-scale AI workloads to create AI-supported product innovations of the future.

To overcome these challenges, BMW built the Connected AI Platform. The platform consists of a cloud-based AI development and deployment environment that helps internal teams develop, train, and deploy AI applications for in-vehicle systems and mobile apps without managing the underlying infrastructure. One example is the Charging Flap use case, which covers BMW’s full AI/ML lifecycle. Teams train, validate, compile, and sign AI/ML models in the cloud (with Connected AI Platform Offboard), and deploy to and infer the models in the vehicle (with Connected AI Platform Onboard).

Another use case is the Proactive User Recommender Engine (PURE-AI). Incorporated into BMW’s Intelligent Personal Assistant, the engine operates entirely in the cloud and uses state-of-the-art AI/ML models, including reinforcement learning methods. PURE-AI processes large volumes of interaction and context data from BMW’s Connected Customer Shadow, a comprehensive digital customer profile. This rich data source aggregates vehicle usage and behavioral patterns from the BMW fleet. As a result, PURE-AI can learn customers’ preferences and anticipate their needs, delivering highly personalized suggestions for in-car functions and apps. This enables customers to experience a more intuitive, more convenient, and less distracting interaction with their vehicle, as relevant features are proactively surfaced exactly when needed. To maintain effectiveness, BMW monitors both offline and online metrics, tracking the real-world usefulness of the recommendations that its customers receive.

To accelerate experimentation, drive improvement, and power its use cases, BMW turned to AWS infrastructure to train, deploy, and run AI/ML models at scale. BMW also adopted comprehensive AWS tools for building AI applications with both custom and foundation models. And by implementing robust security and privacy controls on AWS, the manufacturer could meet its stringent requirements.

BMW built its Connected AI Platform on Amazon Elastic Kubernetes Service (Amazon EKS)—a service that lets developers start, run, and scale Kubernetes without thinking about cluster management. The architecture emerged from a collaboration with the AWS team, including a hackathon. During this event, BMW’s team experimented with the open-source framework Ray and built a proof-of-concept cluster to test its distributed-training capabilities.

The platform comprises three architectural layers that separate infrastructure complexity from AI development work. The first layer handles the core computational needs. Amazon EKS manages cluster orchestration, automated scaling, and high availability across multiple data centers. The second layer, Orbit, manages BMW-specific requirements such as security protocols, compliance controls, and enterprise system integration.

The third layer provides AI capabilities through three specialized APIs: a large language model API for model access and routing, a workflow API for AI processing and training, and an inference API for near real-time and batch operations. Teams across BMW can access powerful computing resources and develop AI applications without needing deep technical knowledge of the underlying infrastructure. Domain experts, such as battery or electrical engineers, can train sophisticated AI models by using the same familiar tools they use for other work.

The Connected AI Platform operates through two core engines. The workflow engine manages AI model training, data processing, and distributed-computing tasks, using Ray to coordinate work across multiple GPUs and nodes. The inference engine deploys the trained models into production, automatically handling scaling, reliability, and performance optimization necessary to serve BMW’s global vehicle fleet. The solution also incorporates Karpenter for intelligent resource management, automatically provisioning GPU-powered nodes when workloads require them and spinning down resources when they’re no longer needed.

BMW’s new architecture demonstrated promising results in early testing. Previously, the company’s Intelligent Personal Assistant team needed to wait overnight for its AI models to finish training. But with access to the new distributed-computing capabilities, training times dropped dramatically. “The proof-of-concept cluster that we built on Amazon EKS helped us scale down to 30 minutes of distributed training by using distributed GPUs while staying under EUR €5 per training run,” says Thomas Riedl, ML engineer for the Connected AI Platform at BMW. “This shows how the project can continue to scale up and do more training with the help of the Connected AI Platform on AWS.”

BMW’s Connected AI Platform is currently in the proof-of-concept phase and is scheduled to be fully functional in a few months. Several of the platform’s components are already operational and support internal AI initiatives, including Charging Flap and PURE-AI. The company plans to fully roll out the solution to all teams, supporting more than 550 developers and powering over 60 use cases. As BMW prepares to unveil new AI-powered features in its next generation of vehicles, the Connected AI Platform positions the company to rapidly scale AI development and deployment across its global operations.

“We use Amazon EKS and open-source solutions wherever possible,” says Paul Weber, product manager Connected AI Platform for BMW. “We want a flexible approach so that we can create solutions for a global footprint.”