Fidelity's Director of Global Macro Jurien Timmer called the AI trade 'dead money for more than three months,' citing flat-to-declining token expenditure and GPU rental rates for H100 and A100 chips. Morgan Stanley put numbers to the rot: more than half of GPU servers sold between 2026 and 2028 may have nowhere to plug in. Jefferies adds that power shortages are only the start. Labor, transformers, cooling equipment, and backup generation are all bottlenecks. Between $200 billion and $300 billion of NVIDIA hardware sold since 2022 is sitting in warehouses or unplugged in data centers.
The core argument here is not just that chips are idle. It is that GPU sales figures have been misread as proof of real AI demand. Microsoft CEO Satya Nadella has admitted chips are sitting unplugged. OpenAI and Anthropic account for 70 to 80 percent of all AI compute revenues, and they received over $217 billion in investment in nine months, mostly from Google, Amazon, Microsoft, and NVIDIA themselves. The 'demand outstripping supply' narrative collapses when two customers are generating most of that demand with money their suppliers gave them.
The full piece traces exactly how hyperscaler capex creates a circular illusion: companies buy chips years early, manufacturing a backlog that looks like scarcity, while the 'customers' driving revenue growth are the same entities funding the whole loop. Friday's follow-up will interrogate generative AI's effect on productivity, treasury yields, inflation, and what breaks first when the cycle stops. If you have information on Anthropic, OpenAI, or related companies, contact Signal at ezitron.76.
[READ ORIGINAL →]