Is AI running out of compute?

AI demand is shifting from model training to inference, and agentic tools are sharply increasing the amount of compute each user consumes. That surge is colliding with tight supply in chips, power, and data center capacity, creating a more constrained market than many investors expected.

In this environment, pricing power should favor providers of scarce inputs such as advanced semiconductors, energy infrastructure, and data center operators. Hyperscalers remain important because they are building the “toll roads” for future demand, but their costs may rise before productivity gains fully show up.

For businesses using AI, unchecked usage could pressure margins if compute costs climb faster than revenue benefits. Investors may want to broaden exposure beyond the largest AI platforms and toward infrastructure names that stand to benefit from years of buildout.