Baseten, an AI infrastructure provider, has achieved significant cost-performance improvements for AI inference. By utilizing Google Cloud's A4 virtual machines with NVIDIA Blackwell hardware and its Dynamic Workload Scheduler, Baseten offers enhanced efficiency. This breakthrough allows companies to affordably deploy advanced AI models into production, enabling new applications. Inference, the process of running AI models, is crucial for enterprise AI and has become a bottleneck due to increasing compute demands. Baseten addresses this by offering a scalable inference platform for various proprietary and open-source models, emphasizing performance and cost. Their strategy involves optimizing cutting-edge hardware with a sophisticated open software stack. This includes leveraging NVIDIA GPUs and advanced frameworks like NVIDIA Dynamo and TensorRT-LLM. Baseten's software optimizations boost throughput and reduce latency, positioning them as a top performer. Furthermore, their global infrastructure and use of Google Cloud's Dynamic Workload Scheduler ensure high availability and redundancy. This powerful combination of hardware and software unlocks new AI applications across industries at unprecedented scale and cost-efficiency.