Vipul Ved Prakash, Together AI ｜ RAISE Summit 2025

Generative AI infrastructure is shifting toward a supercomputing-as-a-service model, where massive data centers function as single, unified computers to handle the immense horsepower required for modern workloads. Together AI CEO Vipul Vedprakash highlights the transition from consumer-grade GPUs to high-density clusters like the Grace Blackwell 200, which delivers 1.4 petaflops of compute per rack. This evolution enables AI-native companies and enterprises to move beyond basic SaaS models toward agentic, reasoning-heavy applications that utilize test-time compute to reduce hallucinations. Open-source models serve as a critical foundation for this growth, allowing developers to fine-tune and adapt 200+ available models for specific use cases in robotics, healthcare, and media. To maximize the utility of expensive hardware, specialized AI clouds must integrate sophisticated orchestration and fault-tolerance layers, enabling a seamless transition between real-time inference and background reinforcement learning.