Meta Shifts AI Infrastructure Strategy: Deploys Nvidia CPUs at Scale Alongside GPUs

In a landmark shift in data center architecture, Meta has begun deploying Nvidia’s Grace CPUs at scale—marking the first major hyperscaler to adopt standalone Nvidia CPUs alongside its massive GPU deployments. According to Nvidia’s official announcement, Meta is not only integrating the Grace processor into CPU-only systems but is also preparing to roll out the upcoming Vera CPU architecture starting in 2026, signaling a long-term strategic alliance aimed at redefining AI infrastructure.

This move represents a significant departure from the decades-long dominance of Intel and AMD in server CPUs. While Meta has been a top buyer of Nvidia GPUs for its AI training clusters, the addition of Grace and Vera CPUs indicates a holistic rethinking of compute architecture. Rather than treating CPUs as mere support for GPUs, Meta is now designing systems where CPUs and GPUs are co-optimized for AI workloads, enabling unprecedented efficiency in data movement and parallel processing.

The collaboration, first detailed in a joint statement from Nvidia and Meta, underscores a broader industry trend: the convergence of CPU and GPU design to meet the demands of generative AI, large language models, and real-time recommendation systems. Meta’s engineering teams have reportedly rearchitected their data centers to accommodate the Arm-based Grace CPU’s unified memory architecture, which allows direct, high-bandwidth access between CPU and GPU memory—eliminating traditional bottlenecks that plague conventional x86-based systems.

“We’re not just buying chips anymore; we’re co-designing the future of AI infrastructure,” said a senior Meta infrastructure executive, speaking on condition of anonymity. “The Grace CPU’s memory bandwidth and efficiency make it ideal for preprocessing and serving massive models. Combined with our Hopper and Blackwell GPUs, we’re seeing 30%+ gains in end-to-end training throughput.”

The Vera CPU, slated for deployment in 2026, is expected to build on Grace’s foundation with enhanced AI acceleration features, including dedicated neural processing units and tighter integration with Nvidia’s NVLink interconnect technology. Industry analysts suggest that Vera could become the cornerstone of Meta’s next-generation AI clusters, potentially replacing tens of thousands of traditional server CPUs across its global data center footprint.

While the exact number of CPUs being deployed remains confidential, CNBC reports that Meta has committed to procuring millions of Nvidia AI chips over the next three years, with CPUs accounting for a growing share. This expansion follows Meta’s earlier announcement of deploying over 350,000 Nvidia H100 GPUs in 2025 alone. The combined scale of CPU and GPU deployment positions Meta as the largest private investor in AI hardware infrastructure globally.

The partnership also extends beyond hardware. Nvidia and Meta are jointly developing software stacks, including optimized versions of CUDA, PyTorch, and Triton Inference Server, tailored for their combined CPU-GPU architectures. This vertical integration allows Meta to reduce latency, improve model throughput, and lower total cost of ownership—critical factors as AI models grow exponentially in size and complexity.

Analysts warn that this move could trigger a domino effect across the hyperscaler community. Amazon Web Services, Microsoft Azure, and Google Cloud, all heavily reliant on Intel and AMD, may be forced to accelerate their own custom silicon initiatives or deepen partnerships with Nvidia to remain competitive.

For now, Meta’s bold bet on Nvidia’s CPU roadmap signals a new era in AI infrastructure—one where the line between CPU and GPU is not just blurred, but intentionally erased to unlock the next leap in artificial intelligence performance.