AI Value Chain: How much does a GW of AI ASIC data center capacity actually cost?

Based on a series of conversations with industry experts, supply chain checks, and third party data, we extend our analysis of Nvidia Blackwell and Vera Rubin AI data center pod-level economics to include estimates for AMD and ASIC architectures. We revise our Vera Rubin estimates downward, driven by lowered estimates of per-rack NAND content and more conservative HBM 4 pricing, while also revising our estimates for networking content, cooling and power delivery based on Asia channel checks and reflecting the latest DRAM and NAND pricing. We estimate AMD's MI455X architecture to be relatively similar in cost to Vera Rubin, though AMD appears to earn lower margins on GPUs and thus has lower computing content despite higher CPU value, with networking content significantly higher due to up to 3 NICs per GPU. We find OpenAI's Jalapeno to be a radically different design that intends to scale-up the domain and hold data locally in HBM and DRAM, reducing dependence on scale-out and NAND, whil