.svg){kind=spacer}
Details
- Runway announced a partnership with NVIDIA to deploy its Gen-4.5 video generation model on Vera Rubin, described as the first video generation model to run on NVIDIA's most advanced accelerator before its official release.
- The announcement highlights video generation and world models as ushering in a new era of AI that understands and interacts with the physical world.
- Vera Rubin refers to NVIDIA's H100 NVL GPU, a next-generation accelerator based on the Hopper architecture with dual GPUs connected by NVLink.
- H100 NVL specs include 94GB HBM3 memory per GPU (188GB total), 3.9TB/s memory bandwidth per GPU (7.8TB/s combined), and up to 3,341 TFLOPS FP8 Tensor Core performance.
- The H100 NVL is optimized for demanding AI inference, delivering up to 12x faster inference than the prior A100 for generative AI workloads like large language models.
- Partnership enables early access, positioning Runway to leverage NVIDIA's cutting-edge hardware for advanced video generation ahead of competitors.
Impact
Runway's early deployment of Gen-4.5 on NVIDIA's H100 NVL accelerator ahead of its release gives it a head start in video generation, where real-time performance and model scale are critical differentiators from rivals like OpenAI's Sora or Google's Veo, which rely on similar but potentially less optimized hardware stacks. This partnership lowers inference costs through H100's 30x speedup over A100-era GPUs and massive HBM3 bandwidth, accelerating adoption of world models that simulate physics for applications in robotics, gaming, and autonomous systems. It aligns with the GPU shortage narrative by showcasing efficient scaling via NVLink, pressuring cloud providers to prioritize H100 allocations and steering R&D toward on-accelerator video diffusion. Over the next 12-24 months, expect this to widen the performance moat for frontier video AI, funneling more venture funding into multimodal generation while highlighting NVIDIA's lock-in on the AI hardware trajectory.
