Timeline
The cloud is older than it looks.
Scrub through a century and a half. The campus below changes with the era, from a raised-floor computer room to a liquid-cooled AI factory.
2024
The AI factory as a service.
- Building
- Modular pods, retrofit shells, rapid expansion modules
- Rack density
- ≈40–150 kW/rack
- Cooling
- Direct-to-chip liquid
- Power
- Leased utility capacity, on-site gas turbines or bridging power
- Networking
- Dedicated per-tenant fabrics or partitioned shared fabrics; dark fiber and carrier diversity
- Software
- GPU cloud control plane: reservation, scheduling, health checks, tenant isolation
- Business model
- Reserved clusters and on-demand GPU-hours sold to model builders and enterprises
- Bottleneck
- GPU allocation, energized power and the speed of capital
- Geography
- Anywhere with power available quickly: retrofits, colocation, build-to-suit and remote sites
- Environment
- Speed-to-power can mean on-site fossil generation; choices vary widely by operator
2024Rack-scale, liquid-cooled systems
Liquid coolingBlackwell-generation racks join 72 GPUs in one NVLink domain at well over 100 kW per rack; direct-to-chip liquid cooling becomes the norm for frontier clusters.
| System | GB200 NVL72: 72 GPUs, 36 CPUs | Vendor spec |
|---|---|---|
| Rack power | ≈120–132 kW | Reported |
| GPU | B200: 208B transistors, TSMC 4NP, 192 GB HBM3e | Vendor spec |
Key components
- Compute trays
- NVLink switch trays
- Copper spine
- CDUs
- Cold plates
Inspect in 3D