Infrastructure sequencing
The early AI investment cycle was dominated by graphics processors. Attention then moved to advanced packaging, high-bandwidth memory, networking, and power. These constraints remain relevant, but the composition of AI workloads continues to change.
Training and inference require different systems
Model training rewards highly parallel computation, which favours GPUs. Inference requires a broader orchestration layer: data movement, tool calls, memory management, security, storage, and general-purpose processing all become more important as AI applications move into production.
Agentic systems increase that complexity further. An agent does not simply produce a single answer; it coordinates multiple processes and tools. That can increase the relative requirement for CPUs and supporting infrastructure alongside continued GPU demand.
The investment question is not whether AI spending continues. It is which component becomes the next constraint on system deployment.
Physical supply cannot respond instantly
Semiconductor capacity is capital intensive and slow to expand. Foundry capacity, packaging equipment, and data-centre construction are committed well in advance. When demand changes faster than the supply chain can respond, pricing and earnings can move sharply for the businesses positioned at the constraint.
How Priori approaches the theme
- Map the entire system rather than treating AI infrastructure as one homogeneous trade.
- Separate durable capacity requirements from short-lived market narratives.
- Underwrite the earnings impact, competitive position, and valuation of each beneficiary.
- Review exposure as the binding constraint shifts across the supply chain.