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

AI Datacenter Boom Drives Massive Infrastructure Investment

Published September 27, 2026 11:41 PM · 9 views
AI Datacenter Boom Drives Massive Infrastructure Investment

As the AI sector intensifies competition, securing GPU capacity alone is no longer sufficient. The industry now requires vast spaces for tens of thousands of GPUs, high-speed networking infrastructure, substantial power supply, and advanced cooling systems.

The race has expanded into building larger, more powerful datacenters at a faster pace. Hyperscalers' 2025 capital expenditure (CAPEX) is projected to reach $130 billion, approximately 1,800 trillion won, with a significant portion flowing into datacenter construction.

Currently, the U.S. has 3,408 datacenters total: 1,301 operational and 2,059 planned. Operational facilities consume about 61,000 MW of power, while planned centers require approximately 378,000 MW.

Traditional datacenters housed thousands to tens of thousands of servers handling various tasks like search, payments, databases, and email. However, AI has transformed this model: instead of multiple CPUs processing different tasks sequentially, AI requires massive parallel computation using thousands to hundreds of thousands of GPUs working together on a single model.

For AI datacenters, the critical factor is not just GPU performance but how effectively they can be interconnected through high-speed networks. Data must constantly flow between GPUs; slow communication would render expensive hardware underutilized.

Datacenters are evolving into massive computing systems where tens of thousands to hundreds of thousands of GPUs are connected via ultra-fast networks, functioning as a single AI computer.

Four main groups are building these facilities: Big Tech and cloud providers (AWS, Microsoft, Google, Meta, Oracle), AI companies (OpenAI, xAI), specialized AI infrastructure firms (CoreWeave, Nebius), and datacenter development operators (QTS, Ventas, STACK).

Notably, datacenter owners don't necessarily purchase GPUs directly. For instance, AWS or Microsoft may buy GPUs themselves, while providers like QTS rent space to customers who then procure their own GPU infrastructure.

NVIDIA is no longer just selling GPUs; it's also offering complete server systems including memory, storage, networking, power supply, and cooling solutions. This expansion means the entire datacenter supply chain is growing alongside AI demand.

Key components seeing surging demand include HBM (High Bandwidth Memory) for faster GPU data access, storage solutions like NAND and high-performance SSDs, network infrastructure such as switches, NICs, optical transceivers, and fiber optics, power distribution equipment including transformers and UPS systems, and advanced cooling technologies.

According to IEA projections, AI server power consumption is expected to grow at an annual rate of about 30% through 2030, significantly outpacing the 9% growth for CPU-based servers. Global datacenter electricity use will more than double from around 415 TWh in 2024 to approximately 945 TWh by 2030, with the U.S. accounting for about 45-50% of this total.

Several bottlenecks threaten datacenter construction: power grid connections, shortages of transformers and switchgear, lengthy transmission line development (4-8 years), regulatory hurdles, cooling infrastructure limitations, and skilled labor shortages. JLL reports that 57% of 2025 datacenter projects face delays exceeding three months.

Delayed completion carries significant risks: revenue loss from delayed operations, reduced economic value due to lower present value of future cash flows, shortened GPU economic lifespan as newer models emerge, increased construction and financing costs, customer contract issues, and worsened investment recovery and debt burden.

Korean Source

This article is an English localization of a Korean-language crypto news report. Original headline: 1800조원은 어느 기업으로?…AI 데이터센터의 기업지도 [박신영의 월가 아나토미]