Where're All The AI Chips?

Newsletter · Watch/Read · (gathered 2026-09-23)

• The AI industry is in a massive speculative bubble, with hyperscalers and neoclouds misleading investors about operational AI data center capacity; NVIDIA’s record revenues are largely due to speculative purchases of GPUs years in advance, which are sitting in warehouses or unpowered data centers, not actively generating revenue.
• Microsoft, as a prime example, claims gigawatts of new capacity but only has ~2GW of AI-specific operational capacity, while warehousing an estimated $50-$100 billion in GPUs. This “weasel wording” about “capacity” is an industry-wide problem, conflating total power secured with actual IT load.
• Across the industry, over $200 billion in GPUs are estimated to be uninstalled within $374 billion+ of “Construction In Progress” assets. This overbuild is exacerbated by OpenAI and Anthropic’s massive, but largely speculative, compute commitments, which distort the perception of “insatiable demand.”
• Investors and the public are conflating GPU sales with actual AI compute demand, setting the stage for brutal depreciation charges, write-offs, and potential financial destruction for many companies when the vast excess capacity eventually comes online without sufficient demand.