Power, Copper & Data Centers — The Foundation of AI
agent37.club
The absolute foundation. Without energy, nothing above exists.
Power is the binding constraint on AI scaling. Not talent, not algorithms, not capital — electricity.
projected global DC power by 2026
(up from 49 GW in 2023)
A single GB200 NVL72 rack draws ~120 kW. A data center with 10,000 such racks needs 1.2 GW — the output of a large nuclear plant.
This is just one facility for one company.
top 4 hyperscaler capex in 2025. Projected $602B across top 5 in 2026.
| Metric | Value |
|---|---|
| Build cost per MW (IT capacity) | $10–15 million |
| PUE (Power Usage Effectiveness) | Best: 1.1 / Typical: 1.3–1.5 |
| Design life | 15–20 years |
| Time to build (greenfield) | 18–36 months |
| Staff per 100 MW facility | ~50–100 |
Total facility power ÷ IT equipment power. PUE of 1.3 means 30% overhead for cooling, lighting, networking. Google claims PUE of 1.1. AI racks are pushing PUE higher because of cooling demands.
GPUs dominate upfront cost. But over 15 years, electricity becomes 30–40% of total cost of ownership. That’s why power pricing matters so much.
| Region | $/kWh | 100 MW Annual Cost | Notes |
|---|---|---|---|
| Nordic countries | $0.03–0.05 | $26–44M | Hydro, cool climate |
| Quebec | ~$0.04 | ~$35M | Cheap hydro |
| US (avg) | $0.08–0.12 | $70–105M | Varies wildly by state |
| Northern Virginia | $0.07–0.09 | $61–79M | Fiber hub, but grid maxed |
| Singapore | $0.15+ | $131M+ | Land + power constrained |
A 2x difference in electricity price = $40M+/year difference for a single 100 MW facility. Over 15 years, that’s $600M.
Big tech signed 10 GW+ of new US nuclear capacity in the past year.
| Source | Capacity Factor | $/MWh | Carbon | AI Suitability |
|---|---|---|---|---|
| Natural Gas | 60–80% | $40–60 | Medium | Good — dispatchable |
| Solar | 25–30% | $30–50 | Zero | Intermittent — needs storage |
| Wind | 30–45% | $30–60 | Zero | Intermittent — needs storage |
| Nuclear | 90%+ | $60–90 | Zero | Best for baseload |
| Geothermal | 90%+ | $40–80 | Near-zero | Excellent — limited locations |
AI workloads run 24/7. Solar and wind alone can’t do this without massive battery storage. Nuclear and gas provide the baseload reliability AI data centers need.
of water per kWh for evaporative cooling
Arizona, Texas, Nevada — popular data center locations that are water-stressed. Some municipalities are already pushing back on new data center permits.
Liquid cooling (direct-to-chip) reduces water use by 80%+ but increases capex and complexity.
| Workload | Power/Rack | Cooling |
|---|---|---|
| Traditional server | 5–10 kW | Air |
| GPU (A100 era) | 20–30 kW | Air (pushed) |
| GPU (H100 era) | 40–60 kW | Air + rear-door |
| GB200 NVL72 | 120–140 kW | Liquid required |
Direct-to-chip: cold plates on GPU/CPU, water loops. Most practical upgrade.
Immersion: entire server submerged in dielectric fluid. Best thermal performance but requires complete redesign.
Liquid-cooled racks: 46% of market in 2024, projected 50–76% by end 2026. No longer optional.
Data centers are copper-intensive:
DCs projected to consume 1.1M tonnes/year by 2030 (~3% of global demand). Copper prices +35% in 2025, above $11,000/tonne.
Data centers are fighting EVs and grid electrification for the same copper. An EV uses 4x the copper of an ICE vehicle.
New copper mines take 10–15 years from discovery to production. Supply can’t respond quickly. Prices at all-time highs.
| Technology | Bandwidth | Use Case | Key Player |
|---|---|---|---|
| NVLink | 900 GB/s | GPU-to-GPU (within node) | NVIDIA |
| InfiniBand | 400 Gb/s | GPU-to-GPU (across nodes) | NVIDIA (Mellanox) |
| Ethernet (800G) | 800 Gb/s | General + AI training | Broadcom, Arista |
| NVSwitch | N/A (fabric) | All-to-all GPU connectivity | NVIDIA |
A 100,000 GPU training cluster needs ~$1B in networking alone. InfiniBand has lower latency but creates NVIDIA lock-in. Ethernet is catching up.
average wait time for new grid interconnection in many US regions
It’s not just generation — it’s transmission. You can build a solar farm or a nuclear plant, but getting the power to the data center requires grid infrastructure that doesn’t exist yet.
This is why companies are exploring on-site generation (natural gas turbines, SMRs, fuel cells).
| Material | Used For | Supply Concentration |
|---|---|---|
| Neodymium | Magnets in generators, fans, drives | China: ~60% mining, ~90% processing |
| Lithium | Backup batteries (UPS systems) | Australia, Chile, China |
| Silicon | Semiconductor wafers, fiber optics | China: ~70% polysilicon |
| Germanium | Fiber optics, infrared sensors | China: ~60% |
| Copper | Power, cables, cooling, networking | Chile, Peru, Congo |
China controls processing for most critical materials. In 2023, China restricted gallium and germanium exports. Material access is a geopolitical weapon.
Upfront, GPUs dominate. Over the full lifecycle, electricity is the #1 cost. This is why every hyperscaler is obsessed with power contracts.
| Sector | Companies | Thesis |
|---|---|---|
| Power Utilities | NextEra, Vistra, Southern Company | Growing power demand = growing revenue |
| Nuclear | Constellation, Cameco, NuScale | AI driving nuclear renaissance |
| Copper Miners | Freeport-McMoRan, Southern Copper | Structural supply deficit |
| Cooling | Vertiv, Modine, nVent | Liquid cooling transition |
| Networking | Arista Networks, Broadcom | $1B+ networking per 100K GPU cluster |
Vertiv stock: up ~8x since 2023. Vistra: up ~5x. Constellation: surging on nuclear PPAs. The market is waking up to the infrastructure layer.
Next week: Build Sessions — from understanding to building
agent37.club