SIA and Deloitte just did a virtual teardown of a leading AI server rack and found 4,500 packaged chips inside, with semiconductors claiming more than 95 percent of the rack's content value. The number that matters for supply chain planning is the projection: $1.2 trillion in annual AI chip revenue by 2028, a nearly tenfold increase over four years, and more than 50 percent above total global semiconductor sales from 2025 across all end uses.
The more useful part of the report is not the headline figure. It is the chip category breakdown. The teardown confirms that AI infrastructure is not a GPU story. It is a full-stack story. Advanced logic (accelerators, ASICs, FPGAs, CPUs, DPUs, networking chips) plus memory (HBM, DRAM, SRAM, NAND) plus analog and foundational chips (power regulators, transceivers, controllers, sensors) are all load-bearing. The analog and foundational segment is the one most underrepresented in the AI build-out conversation, and also the one with the longest lead times and least fab optionality.
For hardware teams sizing their AI infrastructure supply chain or designing custom AI rack solutions, the $2.8 trillion in projected semiconductor spend through 2028 is not background noise. It is the forward signal that every chip category in a rack is now a capacity constraint, not just GPUs and HBM. The teams that win are the ones that have already mapped their bill of materials across all 4,500 chips, not just the headline accelerators.