AI infrastructure decisions now carry multi-year, multi-million commitments: reserved GPU capacity, rack-scale systems that need liquid cooling, data center power that takes years to secure, and platform choices that shape every AI product built on top. This course gives technology executives the judgment to make those decisions without getting lost in specifications. It starts from workloads (training, fine-tuning, batch and real-time inference, agents) and how to size demand in tokens, concurrency and latency targets, then compares frontier APIs, hyperscaler model platforms, open-weight models, cloud GPU instances, neoclouds, reserved capacity, colocation and owned clusters.
A hardware module reads the current roadmap at decision level, including NVIDIA Blackwell, Blackwell Ultra and the Vera Rubin generation, AMD Instinct and rack-scale Helios systems, and custom silicon such as Google TPU and AWS Trainium, along with the networking and storage costs that sit beyond the GPU. Further modules cover power density, liquid cooling and grid lead times; TCO models built on utilization, depreciation and cost per token; supply, obsolescence, sovereignty and security risk; and the internal AI platform of gateways, routing, observability and guardrails. The course ends with a three-year roadmap you can take to the board. Product facts are stated as of October 2026.