Nvidia Unveils Next-Gen Blackwell Ultra Architecture as Enterprise AI Infrastructure Capex Tops $200B
SAN FRANCISCO — In a blockbuster keynote opening this year’s Global Tech & Enterprise Summit, NVIDIA unveiled its next-generation Blackwell Ultra architecture, delivering a tenfold leap in FP4 reasoning throughput and sub-millisecond inference for autonomous generative models.
Hyperscale infrastructure providers and venture-backed AI labs have already committed an annualized $200 billion in capital expenditures toward high-density accelerator clusters, underscoring an unprecedented paradigm shift across enterprise software, sovereign compute, and industrial automation.
Key Enterprise Milestones
| Architecture Tier | Compute Density (FP4) | Power Efficiency | Commercial Availability |
|---|---|---|---|
| Blackwell Ultra B300 | 28 PFLOPS / Node | 3.8x vs Previous Gen | Q1 2027 |
| Grace-Blackwell NVL72 | 1.4 ExaFLOPS Cluster | Liquid-Cooled 100kW | Available Now |
Industry consensus indicates that hardware constraints are rapidly evolving from memory bandwidth to power grid interconnects. Early pilot customers report achieving 70% reductions in training epochs for multi-modal reasoning models.
댓글목록
등록된 댓글이 없습니다.