Data Center / Infrastructure
By Calvin Alexander | Digital Dojo 360 | Nov 7, 2026
Key takeaways
- Power access, not buildings or chips, now sets the pace of AI expansion.
- Electrical systems take up to 45% of construction cost. The building shell takes only 10% to 15%.
- Racks above roughly 35 kW need liquid cooling, and AI racks are heading past 100 kW.
- Support for data centers falls from 93% nationally to 35% locally.
- Newer GPUs cost more per hour but can cost less per token.
Most people see AI through software: a browser tab, a chat window, an app. The real contest is happening in heavy industry. The computers behind generative AI live in massive facilities that consume huge amounts of power, land, structural steel, and raw materials.
The scale is historic. According to JLL, global data center power capacity is projected to nearly double, from 103 gigawatts (GW) in 2025 to roughly 200 GW by 2030. Driven by hyperscale tech giants and private equity firms, total global spending on data center construction and tenant IT infrastructure could approach $3 trillion over the same period.
As the buildout speeds up, developers, utilities, and hardware operators are running into physical, thermodynamic, and social limits. Here are five realities shaping the infrastructure supercycle behind AI.
1. The new bottleneck: time-to-energize replaces time-to-build
The binding constraint on AI expansion is no longer constructing buildings or buying chips. It is securing high-voltage utility grid connections and long-lead electrical equipment. In core data center markets, operators now face grid connection waits of up to 10 years.
Supply chains add to the squeeze:
| Equipment or step | Reported wait |
|---|---|
| Utility grid interconnection (core markets) | Up to 10 years |
| Generator step-up transformers (over 50 MVA) | 100 to 150+ weeks (up to 34 months) |
| 15kV medium-voltage switchgear | Sold out through 2028 in many channels |
| Average U.S. electrical hardware lead time | 42 weeks |
The result: 48% of contractors name power availability as the single biggest obstacle to finishing on schedule, and 57% of data center projects in 2025 were delayed by at least three months.
2. The inverted building: electrical systems are the main event
In conventional commercial real estate, the structure, foundation, and exterior envelope dominate the budget. Data centers flip that. In high-density builds, the structural shell and core are only 10% to 15% of total construction cost.
Heavy power infrastructure takes 40% to 45%. That covers utility interconnections, switchgear, uninterruptible power supply (UPS) systems, and standby generators. Mechanical cooling claims another 15% to 25%.
Average global shell-and-core costs reached $11.3 million per megawatt (MW) in 2026, up 47% from $7.7 million per MW in 2020. Fully fitted for high-density AI, total deployment costs run $20 million to $37 million per deliverable MW.
3. The end of air: liquid cooling becomes mandatory
Air cooling has hit a thermodynamic limit. It efficiently supports about 20 to 35 kilowatts (kW) per server rack. Modern AI clusters regularly need 30 to 50+ kW per rack, and the newest high-density deployments exceed 100 kW.
Hardware drives this shift. A single NVIDIA Blackwell GPU (DGX B200) draws up to 1,000 watts, compared with 700W for the previous-generation Hopper H100. That concentrated heat is more than air systems can manage, so deployments are moving to direct-to-chip (D2C) liquid cooling.
Liquid heat-transfer fluids have a volumetric heat capacity roughly 3,200 times greater than air. D2C cooling can cut cooling subsystem energy use by up to 90% and bring Power Usage Effectiveness (PUE) down from historical averages of 1.5 to 2.0 to between 1.03 and 1.20. JLL projects that 80% of new data center builds will include liquid cooling by 2030, even though those facilities cost 7% to 10% more upfront.
4. The acceptance paradox: a 58-point local support gap
National enthusiasm for AI remains high, but local communities are pushing back on the buildings that support it. JLL research shows a 58-point "acceptance paradox": 93% of survey respondents support data centers at a national level, while local community support drops to just 35%.
Concerns about noise, land use, grid strain, and environmental impact are changing site selection. Land is also scarce in primary hubs like Northern Virginia, where vacancy dropped to 0.3%. Developers are moving into rural and frontier markets across Texas, Louisiana, Indiana, and North Carolina.
To win approval, operators often commit to off-site civil works. For example, AWS committed up to $7 million for local road infrastructure as part of an $11 billion project package in St. Joseph County, Indiana.
5. The GPU price paradox: pay more per hour to spend less per token
Hourly rental rates can paint a misleading picture. Across cloud providers, B200 GPUs carry a 25% to 75% hourly premium over H100s, averaging about $5.95 per GPU-hour versus about $4.76. Measured per unit of finished work, though, the math changes with the precision mode.
FP8 versus FP4. At on-demand rates, Spheron measured standard 8-bit floating point (FP8) inference on a B200 at roughly 51% more expensive per token than on an H100. But Blackwell's native 4-bit floating point (FP4) mode, together with 1.8 TB/s NVLink 5 bandwidth, doubles inference throughput compared with FP8. That makes FP4 inference on a B200 about 26% cheaper per token than on an H100.
System-level gains. In MLPerf Inference v5.0 benchmarks, an 8-GPU B200 system delivered 3x the throughput of an 8-GPU H200 system on Llama 2 70B. At rack scale, liquid-cooled GB200 NVL72 deployments deliver up to 30x faster real-time LLM inference and up to 25x lower energy use and cost per token than H100 air-cooled clusters.
Conclusion: the grid-gated future
The path of AI is no longer set only by software or chip design. Its pace depends on physical constraints: land rights, structural steel, copper, fluid dynamics, and megawatt power contracts. As money flows into giga-campus projects worldwide, the main gating factor is speed to power.
The question for the next decade is simple. Can regional grids, equipment supply chains, and local approval processes adapt fast enough to keep up with the AI supercycle?
- JLL: global data center capacity outlook, cost per MW, liquid cooling adoption, and the acceptance paradox research
- Spheron: GPU rental pricing and cost-per-token measurements
- MLCommons: MLPerf Inference v5.0 benchmark results
- AWS: St. Joseph County, Indiana data center project commitments
Want more plain-English breakdowns on AI and IT service management? Visit digitaldojo360.com.
