Why the AI boom is about to hit a wall

AI Summary

Nate B. Jones argues that the perception of AI as purely a software business is fundamentally flawed; it is, in reality, an industrial enterprise heavily constrained by physical infrastructure. He highlights Microsoft's Q3 2026 earnings call, where Satya Nadella announced a $190 billion capex spend for the year, yet still expects the company to be capacity constrained. This constraint is not primarily about GPUs, but rather the underlying supply chain components like high bandwidth memory (HBM) and packaging, which are less understood in boardrooms. Hyperscalers like Microsoft, Meta, Amazon, and Google are investing hundreds of billions in building AI factories, necessitating a shift in how companies approach AI vendor contracts—they are now effectively supply contracts requiring allocation, capacity terms, and fallback plans. Jones emphasizes that developers and engineers need to be involved in procurement decisions, as token spend can be enormous (e.g., 500 million tokens in a week for his team). The physical bill of materials for AI includes chips, memory, packaging, networking, power, cooling, land, and data center construction, with HBM and packaging being critical bottlenecks. He notes that traditional data center construction timelines are no longer viable for large AI campuses. Despite falling serving costs and efficiency gains (like Microsoft's Copilot inference throughput increasing 40% in a quarter), demand for tokens continues to outpace these improvements, driving the massive capex. Jones concludes by urging businesses to ask critical questions about reserved capacity, routing to cheaper models, and hidden human supervision in AI workflows, asserting that the era of elastic, invisible cloud compute is over for AI, and companies must now think like industrial operators managing a production line.

Claims Extracted (11)