Finding
Rack power has risen roughly 15-fold in one GPU cycle, so AI halls, megawatt charging depots and BESS sites now hit thermal limits before electrical ones
A DGX H100 rack drew about 41 kW; a GB200 NVL72 draws about 132 kW; NVIDIA's Rubin Ultra NVL576, announced for 2027, is expected near 600 kW. Air stops working around 40-50 kW per rack, so every step above that moves cooling to the chip. At device level, a single Blackwell GPU dissipates about 1,000-1,400 W from a package a few centimeters across.
The same limit shows up across electrified sites. Megawatt Charging System (MCS) connectors carry up to 3,000 A and need liquid-cooled cables and inlets; chargers roll back current when connector temperature climbs. Liquid-cooled BESS containers hold cell-to-cell temperature spread within a few degrees to protect life and safety, and C-rate is capped when they cannot. Gigafactory dry rooms couple humidity control and process heat. In each case nameplate electrical capacity overstates usable capacity on the hottest day with one cooling unit out.
The first signals are measurable: routine derating on hot afternoons, hotspots despite adequate average capacity, and pump or fan power climbing faster than load. See Heat Rejection Architectures for the responses and Thermal Failure Modes for what happens when they are ignored.
Concepts
Core concepts
| Concept | Meaning | Why it matters | Shows up first in |
|---|---|---|---|
| Power density | Power per area or volume (kW per rack, kW/m2, MW/acre) | Cooling scales nonlinearly past thresholds | AI halls, DCFC yards, dense process zones |
| Heat flux | Heat per unit area at the source (W/cm2) | Sets the minimum viable cooling method | GPUs, power modules, chargers |
| Approach temperature | Gap between reject temperature and ambient | Decides capacity on hot days | Towers, dry coolers, economizers |
| Coolant delta-T | Supply-to-return rise across the load | Sets flow rate and pumping power | Liquid loops, CDUs |
| Thermal headroom | Margin between load and rejection capacity | Prevents throttling at peaks and in degraded states | Mission-critical and fast-growing sites |
Where it hits
Where density limits hit first
| System | Pressure point | Symptom | Response |
|---|---|---|---|
| AI data centers | Rack density plus fast load swings | GPU downclocking, hotspot alarms | Direct-to-chip, buffering, stronger controls |
| Fleet DC fast charging and MCS | Power cabinets, cables, connectors | Charger rollback, connector overtemperature | Liquid-cooled cables, load scheduling |
| BESS sites | Cell temperature uniformity under cycling | C-rate limits, reduced availability | Liquid cooling, zoning, redundant HVAC |
| Semiconductor fabs | Process stability and cleanroom utilities | Yield drift, tool downtime | Redundant chilled water, heat recovery |
| Gigafactories | Dry rooms, formation, process heat | Throughput caps, humidity excursions | Zoned thermal plants, buffering |
Signals
Threshold signals and metrics
| Signal | What it indicates | Next move |
|---|---|---|
| Routine derating in hot weather | Rejection capacity reached at worst-case ambient | Add headroom; hybridize rejection; more exchanger area |
| Hotspots with adequate average capacity | Heat flux or distribution problem | Move cooling to the source; rebalance manifolds |
| Pump or fan power rising faster than load | Brute-forcing density with flow | Raise delta-T; cut pressure drop; re-architect loop |
| Expansion needs a full redesign | Non-modular thermal plant | Standard cooling blocks and headers |
| Water becomes a board-level risk | Wet cooling misaligned with region | Dry or hybrid; reclaimed water; heat reuse |
Sources
Sources and scope
- Scope
- Thermal Density Limits
- Sources
- NVIDIA product specifications and GTC 2025 roadmap; CharIN MCS specification
- Updated
- October 2026