KOSUNI CRYPTO
Korea's Crypto Pulse, in English
$BTC $ETH $XRP $SOL $DOGE
News
Neutral

AI Bottleneck Shifts to Power Supply…Energy Limits Loom Within Three Years

Published September 23, 2026 6:48 PM · 5 views
AI Bottleneck Shifts to Power Supply…Energy Limits Loom Within Three Years

AI’s biggest constraint is shifting from semiconductor performance to power supply, according to a recent analysis. As data center power capacity fails to keep pace with growing AI model demand, there is an urgent need to fundamentally redesign energy efficiency. During an September 21st interview on the All in podcast, Naveen Rao, co-founder and CEO of Unconventional AI, stated that "AI is one of the most transformative technologies humanity has created" and emphasized that hardware-based innovation is essential for creating true intelligence.

Rao explained that AI service energy demand is already approaching a critical threshold. Citing Google’s data, he estimated that processing over 3.2 trillion tokens monthly—just 10 lines per token—would require approximately 12 gigawatts of power. Given that global data center electricity consumption remains below 100 gigawatts, one company’s AI services alone could account for a significant portion of this total. "As models grow and demand increases, energy use will rise," he said. "My estimate is that within about three years, energy shortages will accelerate." He noted that the data center bottleneck has shifted from space, networking equipment, and GPUs to power contracts being the first consideration.

Rao stated that nearly half of the cost of providing AI tokens comes from energy. "Today, you must secure a power contract before monetizing every watt," he said, highlighting Unconventional AI’s business model, which claims a 1,000x improvement in power efficiency over conventional hardware. Drawing parallels to biology, Rao explained that the human brain operates on about 20 watts, moving roughly 1.6 billion bits per second through the cerebral cortex, while high-performance computing systems move nearly 30 trillion bits per second between memory and processors. "Most of computing energy is spent moving information," he said.

Rao pointed out that conventional computers have not significantly deviated from their 1940s-era architecture of transferring data between memory and processing units. With Moore’s Law improvements effectively reaching a limit, he argued that unnecessary abstraction must be reduced by more directly connecting semiconductor physics with neural networks. Unconventional AI is taking an integrated approach, handling everything from theoretical research to model development, circuit design, systems, and products within one company. Rao initially aimed for a 1,000x power efficiency improvement within five years but accelerated the target to three and a half years due to faster progress in solving scientific problems with AI.

Korean Source

This article is an English localization of a Korean-language crypto news report. Original headline: AI 병목은 전력…3년 안에 에너지 한계 온다