GW-Scale Data Centers: The Power Grid is the Next Bottleneck in the AI Compute Race

When the annual electricity consumption of a data center exceeding that of a medium-sized city with 500,000 population becomes the norm, and when the term "Gigawatt (GW)" from the power generation industry begins to appear frequently in data center planning, we are forced to confront a fundamental question: Is AI's computational demand sprinting at a speed that exceeds the carrying limits of physical infrastructure?

According to TMTPost's report "AIDC Enters the GW Era," the total power of a single campus can reach 1,000 Megawatts (MW), dozens of times that of traditional large data centers (typically tens of MW). The impact of this number far exceeds the technical parameters themselves—it means data centers have officially leaped from "power hogs" to "regional energy hubs." Previously, we discussed server density and cooling efficiency; now, the core variable has become "Can you secure GW-level power supporting facilities?"

I. Short Term: Power Supporting Facilities Become the First Threshold for AI Infrastructure Construction

From 2024 to 2026, capital expenditures of major global tech giants (Microsoft, Google, Amazon, Meta) will concentrate on AI infrastructure. According to Dell'Oro Group's Q4 2024 report, global data center capex is expected to exceed $350 billion in 2025, with over 60% allocated to AI training and inference clusters. The construction cycle for GW-level campuses has stretched from the traditional 18-24 months to 36-48 months, with the core bottleneck not being chips or racks, but power approval and substation upgrades.

Key Data Comparison:

Dimension	Traditional Large Data Center	GW-Level AI Data Center
Single Campus Power	30-50 MW	500-1000 MW
Power Approval Cycle	6-12 months	18-36 months
Annual Electricity Consumption (Ref.)	Approx. 0.5-1 TWh	5-10 TWh (Equivalent to a medium-sized city)
Main Cooling Method	Air/Water Cooling	Liquid Cooling + Waste Heat Recovery

In the short term, GW-level data center construction will exhibit obvious "regional polarization." Site selection strategies shift from "close to network cores" to "close to energy sources." According to IDC data, 27% of new global data centers in 2024 were sited directly near wind/solar bases or nuclear plants, up from just 12% in 2023. This means data centers are no longer just internet infrastructure but are directly embedded into national energy strategic maps.

Changes in Competitive Landscape:

• Cloud Giants Building Own Power Resources: Microsoft has signed multiple 10GW-level renewable energy Power Purchase Agreements (PPAs) and plans nuclear backup capacity in Virginia, USA.
• Third-Party IDC Operators Accelerating Integration: Domestic leaders like GDS and Chindata are beginning to increase single campus design capacity from 50MW to over 200MW, but there remains a generational gap to GW levels.
• Power Companies Moving Upstream: State Grid and China Southern Power Grid intervene at the site selection stage, offering "dedicated substations + green power direct supply" solutions, which was almost unimaginable five years ago.

II. Long Term: Bidirectional Reconstruction of Computing Networks and Power Networks

From a long-term perspective (2027-2030), GW-level data centers will no longer be isolated giant server rooms but nodes in "Computing-Power" composite infrastructure. There are two core trends here:

1. From "Data Center" to "Computing Hub + Energy Storage Hub"

When single campus power reaches GW levels, its load curve impact on regional grids cannot be ignored. AI training loads themselves have intermittency (training tasks can pause/resume), naturally matching the volatility of renewable energy. Therefore, future GW-level data centers will inevitably incorporate large-scale energy storage systems (lithium batteries, flow batteries, or even hydrogen storage). According to BloombergNEF forecasts, global data center-associated energy storage installed capacity will exceed 120 GWh by 2030, equivalent to 15% of global energy storage shipments in 2024. The role of data centers shifts from "pure consumers" to "demand response nodes," actively participating in grid peak shaving.

2. Liquid Cooling Becomes the "Only Solution" Rather Than "Optional"

Traditional air cooling cannot support high-density deployments above 30kW per rack, while the average power per rack in GW-level campuses is expected to reach 50-80kW. According to ICT research reports, liquid cooling penetration in data centers will break 40% in 2025, with immersion cooling dominating GW-level scenarios. This brings major supply chain changes: Coolants (fluorinated liquids/mineral oils), liquid cooling pipes, and CDUs (Coolant Distribution Units) will become fixed asset investments alongside servers, with maintenance costs significantly higher than traditional air cooling.

III. Open Judgment: Is the GW Era an Opportunity or a Trap?

From a business logic perspective, GW-level data centers are essentially the ultimate answer to computing demand—only ultra-large clusters can meet the training needs of next-generation foundation models (GPT-5, Llama-4 level). But risks are equally clear:

• Asset Concentration Risk: Investment in a GW-level campus is approximately $5-8 billion, far exceeding the hundreds of millions for traditional data centers. Once a technological disruption occurs (e.g., algorithmic efficiency improvements reducing computing demand), massive assets may face impairment.
• Policy Lock-in Period Risk: Power supporting facilities usually bind PPAs for 20+ years, requiring enterprises to have extremely strong long-term demand forecasting capabilities. Historically, "Bitcoin mines" in 2018 saw massive asset abandonment due to sudden policy tightening. Will GW-level AI data centers repeat this mistake?
• Energy Supply Ceiling: Global annual new renewable energy installed capacity is about 500 GW, and a single GW-level data center consumes one-thousandth of that. If 10 giants each build 5 GW-level campuses, the required power equals 10% of global new renewable energy. Is this ratio sustainable?

Leaving a question for discussion: When AI's "intelligence" is priced in watts, do we dare bet that this market will eventually grow large enough to absorb GW-level energy consumption? Or is this merely the prelude to a bubble driven by capital, ignoring physical limits?

Original Link: https://www.tmtpost.com/8067651.html