Nvidia Grace Servers Put CPUs Back in the AI Data Center Spotlight

As of July 23, 2026, Nvidia’s Grace CPU story looks more serious than a side note to its accelerator business. On July 21, Tom’s Hardware reported that Nvidia’s Ian Buck said the company has shipped “hundreds of thousands” of standalone Grace servers, following an earlier Nvidia disclosure of more than 2.5 million Grace CPUs shipped in total. That matters because these are not only CPUs hiding inside GPU-heavy systems; Nvidia is highlighting Grace as a server platform in its own right. The bigger signal is that AI data centers are starting to care more about the CPU layer again, especially when workloads involve data preparation, orchestration, inference pipelines, and agent-style software loops rather than only dense model training. (tomshardware.com)

The core Grace building block is a 72-core Arm Neoverse V2 CPU with 4x 128-bit SVE2 support, 64KB instruction cache and 64KB data cache per core, 1MB L2 cache per core, and 114MB of L3 cache in the single Grace CPU C1 configuration. Nvidia lists LPDDR5X memory options of 120GB, 240GB, and 480GB for Grace CPU C1, with memory bandwidth up to 512GB/s on the 120GB and 240GB options, and up to 384GB/s on the 480GB option. The dual-Grace CPU Superchip combines two Grace CPUs for 144 Arm Neoverse V2 cores, 228MB of L3 cache, LPDDR5X memory options up to 960GB, and NVLink-C2C bandwidth up to 900GB/s between the two CPUs. (nvidia.com)

View NVIDIA DGX Spark Personal AI Desktop Supercomputer on partner website

The AI hardware conversation has often centered on how many GPUs can be placed behind a host CPU, but agentic AI changes the rhythm of the system. Agents may call tools, run code, query databases, parse documents, manage context, or launch many smaller tasks around a model request. Tom’s Hardware noted that Nvidia sees some evolving workloads moving away from ratios such as eight GPUs per CPU toward more balanced CPU-to-GPU designs in certain cases. Nvidia’s own messaging around Vera, the upcoming successor CPU, says CPU execution can sit on the critical path for agent actions, reinforcement learning feedback, tool use, sandbox execution, and data pipelines. (tomshardware.com)

See ASUS GeForce RTX 5090 price

Grace does not try to look like a conventional Xeon or EPYC deployment with standard DDR memory and a familiar x86 software baseline. Its main difference is the package-level focus on bandwidth, power efficiency, and Nvidia’s own fabric design. Nvidia says Grace uses a Scalable Coherency Fabric with 3.2TB/s of bisection bandwidth, while its server-class LPDDR5X memory is designed to provide high bandwidth at lower power than traditional DDR memory. That makes Grace especially interesting for workloads where memory movement is a major part of the job, such as analytics, graph processing, microservices, storage, and CPU-side AI infrastructure tasks. It is not a drop-in answer for every x86 workload, but it gives hyperscalers another architecture to evaluate when CPU bandwidth and energy use matter. (nvidia.com)

Buy Crucial T705 13 GB/s NVMe SSD here

Grace is also important because it gives Nvidia a running start before Vera arrives. Nvidia announced Vera in 2026 as a custom server CPU built around its Olympus cores, with 88 cores, Spatial Multithreading, LPDDR5X memory bandwidth up to 1.2TB/s, and second-generation NVLink-C2C offering up to 1.8TB/s of coherent CPU-to-GPU bandwidth for Vera Rubin platforms. Nvidia says Vera systems are expected to be available from system builders and cloud partners starting in fall 2026, so Grace is the proven bridge between Nvidia’s first serious data-center CPU wave and its more custom CPU roadmap. The takeaway is simple: Nvidia still dominates the AI accelerator discussion, but Grace standalone servers show that the company wants the CPU side of the AI data center race too. (investor.nvidia.com)

View Lenovo Legion Pro 7i laptop with Core Ultra 9 and NVIDIA RTX 5090 24GB on partner website