Event date · · NVIDIA
Blackwell Ultra Launch: Inference-Time Scaling Reshapes AI Infrastructure
NVIDIA announced Blackwell Ultra, Dynamo, and the AI Factory platform for inference agents at GTC 2025.
Compute demand has shifted from centralized training to sustained, high-concurrency, long-chain inference, making inference throughput and data center coordination new bottlenecks.
GB300, Dynamo, networking, and storage are all optimized around test-time scaling and agent workloads.
Chip value continues to extend to complete systems, networking, software, and power infrastructure.
Investment assessment should shift from GPU shipments to per-unit inference output, deployment cycles, and data center cash flow.
Monitor inference revenue share, cluster utilization, energy constraints, and alternative accelerators.