Where will the next AI bottleneck emerge?

The next AI bottleneck is most likely to emerge in physical infrastructure: power, advanced AI-chip capacity, and high-bandwidth memory, followed by networking, cooling, electrical systems, and data-center buildout. Demand remains strong, so the limiting factor is increasingly the ability to manufacture, connect, cool, and energize complete AI systems rather than a lack of customers or training data.

The strongest evidence points to a physical-infrastructure bottleneck rather than weak demand or immediate data scarcity. Epoch AI’s forward-looking analysis identifies power availability and chip-manufacturing capacity as the principal constraints on continued training-scale growth. The industry sources add system-level detail: Broadcom reports rapidly rising demand for custom AI accelerators and networking; Micron describes tight memory conditions and sharply expanding DRAM, NAND, SSD, and HBM demand; and Vertiv is accelerating capacity expansion to serve data-center demand. Nvidia’s system-level framing reinforces that usable compute now depends on integrating chips with interconnect, cooling, power, and other data-center components. Microsoft’s strong cloud growth and contracted obligations suggest customers are still demanding AI capacity. The likely next bottleneck is therefore the ability to manufacture and deploy complete, powered AI systems—especially advanced chips and memory, followed by networking, electrical infrastructure, cooling, and data-center construction. The supplied Weixin source could not be substantively verified and does not change that conclusion.