Are AI models running out of power? | The Economist

TL;DR

AI firms face a severe compute shortage because GPU supply, data center construction, and chip manufacturing can't keep pace with exponentially growing demand.

Key Points

• 1.AI companies are already rationing compute access. Anthropic changed its terms of service to discourage peak-time usage, and OpenAI shut down its Sora video tool to redirect scarce GPU resources toward more profitable products.
• 2.The shortage spans the entire hardware stack, not just GPUs. Legacy components like transformers and switches have lead times of 3–5 years, while data center construction faces local opposition over electricity, land, and water use, further delaying capacity.
• 3.Two critical chokepoints dominate the semiconductor supply chain. Nvidia controls over two-thirds of global AI processing power and its chips are effectively sold out, while TSMC is the sole manufacturer of most AI chips — raising its capex by $60 billion this year but still falling short of demand.
• 4.The supply crunch threatens to slow AI development and raise prices. Inference costs, which have been falling every six months, could reverse if shortages persist, potentially slowing adoption — with some analysts calling it a 'natural brake' on AI spending.