Power and memory, not chips, now limit AI buildout

The bottleneck holding back artificial intelligence is no longer the processor. It is everything else — memory, power, and the geopolitical scaffolding required to move chips across borders.

A new analysis from GlobalData, published August 5, finds that AI infrastructure spending has entered a phase in which capital investment, supply-chain resilience, and trade policy are determining deployment speed more than raw compute demand. The shift is visible in this earnings season's results: SK Hynix ↘ 3.12% posted net income growth of 397.5% year-on-year, Samsung 486.7%, and Micron ↘ 2.22% a 1,398.3% surge — all driven by a high-bandwidth memory shortage that has sold out DRAM capacity through 2026.bloomberg+2

The $700 Billion Capital Cycle

Hyperscaler capital expenditure is the clearest signal of the new era. According to Bloomberg, the four largest US cloud companies now plan to spend as much as $725 billion this year on capital expenditures, primarily on AI data-center equipment. Microsoft ↗ 0.52%, Alphabet ↘ 3.58%, and Meta ↗ 1.02% each posted capex growth above 65% year-on-year. Alphabet raised its 2026 guidance to $175–$190 billion, Meta to $125–$145 billion, and Microsoft toward $120–$190 billion. GlobalData calls it "the largest peacetime capital cycle in corporate history."it-online+1

Nvidia's ↘ 1.47% net income more than tripled on data-center demand, while TSMC Taiwan Semiconductor Manufacturing Company Limited ↘ 1.20% grew net income 77.4% even as advanced-node capacity runs short of orders. Constellation Energy's ↗ 0.16% 1,247.5% net income jump underscores what GlobalData analyst Murthy Grandhi calls the real binding constraint: "Power, not chips, is now the binding constraint — a point hyperscalers have made explicitly on recent earning calls".technologyforyou+1

Geopolitics Baked Into the Numbers

Trade policy has become a permanent cost variable. In January 2026, the Trump administration imposed 25% tariffs on select advanced AI chips under Section 232. TSMC, Samsung, and SK Hynix lost blanket Validated End-User exemptions for their China fabrication plants and now require annual US licenses. The Pax Silica pact, signed by two dozen nations in December 2025, formalizes a trusted semiconductor supply chain outside China's orbit.technologyforyou+1

IDC projects worldwide AI infrastructure spending will reach approximately $497 billion in 2026 — up roughly 53% from 2025 — and exceed $1 trillion by 2029. "The risk isn't slowing AI demand," Grandhi said, "but power, memory and licensing constraints capping deployment speed".stocktitan+1