CPUs return to AI spotlight as agentic workloads reshape data center demand; Perplexity plans to adopt Nvidia's Vera

The rise of agentic AI workloads is forcing a fundamental re-architecture of data center hardware. CPUs are regaining strategic relevance due to their superior control-flow handling and low-latency orchestration—capabilities GPUs inherently lack. This shift accelerates adoption of chiplet-based designs, CXL memory pooling, and heterogeneous compiler stacks, favoring next-gen x86 and ARM server CPUs with AMX or SVE extensions. Geopolitical constraints amplify supply chain risks for firms overly reliant on a single GPU vendor, exposing them to both export controls and software lock-in. While NVIDIA leverages its Vera platform to deepen ecosystem moats, rivals like AMD, Intel, and Taiwan, China’s Ampere Computing are capitalizing on open ISAs and customizable IP to capture edge inference segments. Within 18 months, CPU-GPU co-design will become the de facto standard for LLM deployment, marginalizing accelerator-only vendors lacking full-stack integration.