AI Supply Chain · From my video notes
Agents run on more than GPUs.
The resource list behind my video on the agent-era supply chain. Agents burn CPU cycles, memory bandwidth and network capacity around the clock. This page tracks the layers under the GPU. For the financing layer, see the AI Infra Financing Risk Map in this library.
By Koki · October 2026
The layers under the GPU
Vendor pages for specs. Independent analysis for the map.
CPU vs GPU · DigitalOcean
A clear explainer of what CPUs and GPUs each do well, and why production AI systems need both. Start here if the terms blur together.
Vendor · CPUNVIDIA Grace CPU
NVIDIA's Arm-based data center CPU: 72 cores and LPDDR5X memory, built to sit next to its GPUs in GH200 and GB200 systems.
Vendor · CPUAMD EPYC for AI
AMD's server CPU line for AI infrastructure, with its published benchmarks on AI data center workloads.
Vendor · CPUIntel Xeon 6+
Intel's current data center CPU family, positioned for agent hosting density, with up to 288 efficiency cores per socket.
AnalysisSemiAnalysis · Datacenter Model
The research firm's data center model: capacity, power and accelerator deployments tracked facility by facility. The supply chain, quantified.
Open source · ServingSGLang
A high-performance open-source serving framework for LLMs and multimodal models, widely used in production inference stacks.
Open source · ServingvLLM
The open-source inference engine from UC Berkeley's Sky Computing Lab. PagedAttention made high-throughput serving a commodity.