OUCH: The AI economy could crash on mounting chip costs — and those token costs won’t help.

Hardly a week passes without news of another hyperscaler spending billions of dollars on AI chips. A single moderate-to-large data center today uses AI chips costing billions of dollars. A single Nvidia Blackwell GPU in a modern AI chip cluster could cost as much as a new Tesla Model 3. Non-AI chip costs have also risen sharply, with both total spending and unit costs for CPU and memory chips at unprecedented levels. All of this has significant implications for the economy.

The primary reason chip costs are increasing is excessive demand. Proliferation of AI, the Internet of Things, and electric vehicles has increased the overall demand for chips. In particular, chip demand for AI has exploded, supporting both the training of AI models and their deployment across applications. Historically, AI model quality scaled with the volume of compute used to build it — more chips meant better outputs. But the demand driver now is shifting from training to inference. Goldman Sachs forecasts a 24-fold increase in token consumption by 2030, reaching 120 quadrillion tokens per month, as agentic AI systems replace single-prompt interactions with multi-step tasks that consume orders of magnitude more compute per query. Meanwhile, chips must still be replaced every few years simply to remain cost-competitive, compounding demand pressure from both ends.

The enterprise reality is already arriving. Microsoft recently canceled most of its direct Claude Code licenses after discovering that employee AI usage had grown so large that, in the words of one Nvidia executive, “the cost of compute is far beyond the costs of the employees.”

If that last line isn’t indicative of a bubble, I’m not sure what else it might be.

On the consumer side, a low-to-mid-tier 2TB SSD like the one I bought 18 months ago for $89 now goes for $299.