AI Infrastructure as an Industrial Supply Chain
Nate B Jonesgo watch the original →
the gist
AI is no longer just software; it is an industrial manufacturing process where capacity constraints in memory, packaging, and power dictate your ability to ship, requiring CFOs and developers to treat vendor agreements as supply contracts.
The Industrial Reality of AI
AI is not a standard software business with a backend, but an industrial manufacturing system. When hyperscalers like Microsoft report being "capacity constrained," they are not referring to a shortage of GPUs, but to the physical bottlenecks in the production of integrated compute modules. Intelligence at scale requires a massive physical bill of materials, including high-bandwidth memory (HBM), advanced packaging, power, cooling, and data center construction.
The Supply Chain Bottleneck
The primary constraint is not logic die production, but the integration of compute and memory. According to Epic AI, the four largest AI chip designers consumed 90% of global chip packaging capacity and 90% of HBM supply in 2025, while utilizing only 12% of advanced logic die production. This means the industry is limited by the ability to assemble chips into functional, cooled, and powered systems rather than the ability to design the chips themselves.
Procurement and Operational Strategy
Because AI vendors are now dependent on hyperscaler allocation, software contracts must be treated as supply contracts. Organizations should move away from forecasting based on seats or users and instead forecast based on token consumption per workflow, accounting for agent loops, concurrency, and latency requirements.
- Evaluate Allocation Terms: Determine what percentage of vendor spend is reserved capacity versus best-effort allocation, and establish a formal fallback plan for supply disruptions.
- Implement Routing Layers: Deploy routing logic to shift tasks to smaller, cheaper models when high-end model performance is not required, measuring savings against user experience impact.
- Monitor Utilization: Treat token utilization as a core operating metric, as depreciation on hardware occurs regardless of whether the system is serving tokens.
- Involve Engineers in Procurement: Ensure technical staff evaluate whether allocated capacity is actually usable for specific high-volume workloads, such as agentic coding assistants.