Introduction

On May 9, 2025, NVIDIA officially began shipping its long-awaited Grace Blackwell super chip platform, pairing its Grace CPU with the Blackwell GPU in a tightly integrated module designed for AI-heavy data center workloads. Announced last year, the Grace Blackwell duo now powers the first commercial systems from Supermicro, HPE, and Dell, targeting customers running massive AI models across science, enterprise, and government.

“This is the highest-performance computing module we’ve ever built,” said Ian Buck, VP of Hyperscale and HPC at NVIDIA. According to NVIDIA benchmarks, ¹ Grace Blackwell connects CPU and GPU over a 900 GB/s coherent NVLink, enabling 2.5x better performance per watt than traditional x86-plus-discrete-GPU servers.

Unlike conventional server boards, Grace Blackwell unifies memory access for CPUs and GPUs via LPDDR5X, eliminating many bandwidth and cache bottlenecks plaguing training workloads.

Why it matters now

• LLM training infrastructure demands tighter CPU-GPU integration to manage trillion-parameter scale.
• Traditional x86 platforms struggle to keep pace with AI-specific compute needs.
• NVIDIA’s vertical stack threatens legacy server vendors and chipmakers.

Call-out: CPUs can no longer be dumb I/O hubs—Grace is now the brain beside the brain

NVIDIA reports 45% faster training throughput and 3x energy efficiency on GPT-class models when using Grace Blackwell versus standard dual-socket Xeon systems with A100s.

Business implications

Hyperscalers and research labs have new reasons to consolidate onto NVIDIA’s full-stack platform. While Grace Blackwell won’t replace x86 universally, it will likely become the default for AI-specific buildouts in public and private clouds.

Procurement and IT teams must reconsider cost-per-token and performance-per-watt benchmarks that favor tighter integration. Meanwhile, system OEMs without direct access to Grace Blackwell may struggle to keep pace as AI infrastructure standardizes around NVIDIA’s design.

Looking ahead

AWS, Microsoft Azure, and Oracle are expected to launch Grace Blackwell-based compute instances in Q3. NVIDIA has also begun shipping early developer units for its DGX GB200 systems and plans to release open architecture specs for OEMs later this year.

Gartner predicts that by 2028, 40% of enterprise AI compute will use CPU-GPU fusion designs like Grace Blackwell, as chip disaggregation reconfigures server architecture.

The upshot: With Grace Blackwell, NVIDIA isn’t just winning the GPU war—it’s reshaping how CPUs fit into the AI stack. Grace’s moment has finally arrived as AI eats more of the enterprise workload pie.

––––––––––––––––––––––––––––
¹ Ian Buck, NVIDIA HPC keynote, May 9, 2025.