AI/HPC Data Center Design
AI/HPC workloads have flipped data center design. Atlas DC1 is air-cooled and traditional. AI clusters need liquid cooling, busway distribution, sub-millisecond network fabric, and 5-100× the power density. This section maps the gap.
What Makes AI/HPC Different
Atlas DC1 was designed for traditional cloud + enterprise workloads — distributed servers, mixed densities, web/database work. AI/HPC is fundamentally different. Training a single large language model can use 25,000+ GPUs in tightly-coupled clusters with sub-millisecond network coordination. The design constraints flip:
| Dimension | Traditional DC (Atlas DC1) | AI/HPC |
|---|---|---|
| Rack density | ~ 12 kW/rack | 30-100+ kW/rack (training); up to 200+ for inference |
| Cooling | Air (CRAH + containment) | Liquid (DLC, immersion) — mandatory |
| Power per row | 1-1.5 MW | 5-20+ MW |
| Network | 10-100 Gbps Ethernet, ms latency OK | InfiniBand or NVLink, sub-ms latency required |
| Workload pattern | Bursty (web requests come/go) | Constant (training runs for weeks at full load) |
| Failure tolerance | Application-level (web servers fail individually) | Cluster-level (one server failing can halt 1000-server training run) |
| Power continuity | UPS ride-through 5 min OK | Same — but checkpointing failures can cost days of training time |
| PUE target | 1.3-1.5 | 1.05-1.2 |
| Capital cost / MW | $15-22M/MW | $25-50M/MW (cooling + network premium) |
| Build timeline | 18-24 months | 24-36 months (custom mech, complex commissioning) |
The AI Compute Stack — What's Inside an "AI Cluster"
| Layer | Component | Power per unit |
|---|---|---|
| Accelerator | NVIDIA H100 (700W), H200 (1000W), B100/B200 (~ 1200W), AMD MI300 (750W), custom (Google TPU, AWS Trainium, Meta MTIA, Microsoft Maia) | 700-1200W per chip |
| Server (DGX-style) | 8 GPUs + 2 CPUs + memory + NICs | 10-12 kW per server (NVIDIA HGX H100 = 10.2 kW) |
| Rack | 4-8 servers per rack (with cooling) | 40-80+ kW/rack |
| Pod / Cluster | 16-128 racks tightly coupled by InfiniBand | 1-10+ MW per pod |
| SuperPOD / SuperCluster | Multiple pods coordinated for very large training (NVIDIA SuperPOD = 32-127 DGX systems) | 10-50+ MW per SuperPOD |
| Hyperscale AI campus | Multiple SuperPODs (xAI Memphis = 100,000+ H100 = 100+ MW) | 100 MW - 1+ GW |
The Network Fabric — Why It Matters for Power
AI training requires GPUs in different servers to share gradients millions of times per second. Standard Ethernet has too much latency. Two competing technologies:
| Technology | Use | Power impact |
|---|---|---|
| NVLink (NVIDIA) | GPU-to-GPU within and between servers — 900 GB/s per link, sub-microsecond latency | Switch racks (NVLink switches) consume 5-20 kW each |
| InfiniBand (Mellanox/NVIDIA) | Server-to-server within pod — 400 Gbps per port, microsecond latency | IB switches consume 1-3 kW each |
| Ethernet (RoCE) | Alternative for scale-out; emerging Ultra Ethernet | Lower than InfiniBand |
| Optical interconnect | Cross-pod cabling at 800 Gbps+ optical | Optical transceivers add 10-30W per port |
For a 10 MW AI cluster, the network fabric alone can consume 5-10% of total power — not negligible.
Power Distribution Architecture for AI/HPC
| Element | Atlas DC1 (traditional) | AI/HPC equivalent |
|---|---|---|
| Service voltage | 12.47 kV utility | Same OR higher (138 kV for hyperscale campuses) |
| Service transformers | 2 × 2,500 kVA | Multiple 5-30 MVA transformers (per pod) |
| Distribution voltage | 480Y/277V to 415Y/240V | 480V or 415V → some hyperscale exploring 800V DC for direct-to-server feed |
| Per-row distribution | RPP panelboard (400 A) | Bus duct (2,000-4,000 A) |
| Per-rack delivery | 30-60A branch circuits | 100-225A plug-in disconnect from busway |
| UPS ride-through | 5 minutes | Same OR shorter (some designs use rotary UPS for inertia + ride-through) |
| Redundancy | 2N (dual-fed servers) | 2N OR distributed redundant (4N3) at hyperscale; some accept N+1 at module level |
| Cooling power | ~ 30% of IT | ~ 5-15% of IT (DLC much more efficient) |
Worked Example — A 10 MW AI Pod
Example · NVIDIA HGX H100 SuperPODSizing electrical for a typical AI training cluster
The cluster
Compute
128 × HGX H100 servers (8-GPU each) = 1,024 H100 GPUs
Server power
10.2 kW each × 128 = 1,306 kW (just GPUs + CPUs)
Network
36 InfiniBand switches at 1.5 kW = 54 kW
Storage
Parallel filesystem (Weka, Lustre): 100 kW
Total IT load
~ 1,460 kW = 1.46 MW
Sized facility load
- IT load × 1.25 (continuous):1,460 × 1.25 = 1,825 kW
- Mech load (DLC + facility):10% of IT (vs 30% for air-cooled) = 146 kW + facility overhead 50 kW = ~ 200 kW
- Total facility demand:1,825 + 200 = ~ 2,025 kW
- PUE achieved:2,025 / 1,460 = ~ 1.39 (could be lower with optimized DLC)
Electrical infrastructure
Service transformer
2,500 kVA pad-mount (or 2 × 1,500 if 2N)
UPS
2 × 1,250 kVA online double-conversion (2N for IT)
Generators
2 × 2,500 kW Tier 4 diesel
Power distribution
480Y/277V busway (4,000 A) down each row
Per-rack feed
100 A plug-in disconnect (10 kW × 1.25 = ~ 30 A safety margin built-in)
Cooling
Direct liquid cooling (CDUs serving multiple racks); chilled water 30°C supply
i
Why this single pod is bigger than half of Atlas DC1
Atlas DC1 = 2.5 MW total. This single AI pod = 1.46 MW IT (2.0 MW total facility). One pod consumes more power than HALF of Atlas DC1's design capacity. Modern hyperscale AI campuses might have 50-100 of these pods coordinating on a single training run.
The Frontier — Coming Architectures
| Trend (2026) | Implication |
|---|---|
| 800V DC distribution | Eliminates AC-DC-AC conversion at every PSU. Pioneered by Open Compute Project (OCP). Adopted by hyperscale. |
| Battery backup IN the rack | Replace centralized UPS with batteries at each rack — eliminates UPS losses, simplifies redundancy |
| Microgrid + on-site generation | Pair AI campus with on-site PV + ESS + gas turbines. 100+ MW microgrids becoming common. |
| Submersion / two-phase immersion | Pushing rack densities to 200-400 kW/rack |
| Heat reuse to district heating | Datacenter waste heat (50-80°C with DLC) feeds neighboring buildings or even municipal heat grids (Helsinki, Stockholm) |
| Modular AI pods | Factory-built pods shipped to site; deploy in 6 months instead of 24 |
| Co-location with renewables | Build AI campus next to wind/solar farms; long-term PPAs lock in low-cost clean power |
If You See THIS, Think THAT
| If you see… | Think / use… |
|---|---|
| "AI/HPC data center" | 30-100 kW/rack · DLC mandatory · sub-ms network · single training run uses 1000s of GPUs |
| "NVIDIA HGX H100" / "DGX" | NVIDIA's reference 8-GPU server. ~ 10 kW. Standard AI building block. |
| "SuperPOD" | NVIDIA terminology for 32-127 DGX systems coordinated by InfiniBand |
| "InfiniBand" | Required for tight GPU coordination. Higher cost than Ethernet but required for training. |
| "NVLink switch" | NVIDIA's GPU-to-GPU interconnect within and between servers |
| "800V DC" | Open Compute Project standard. Direct DC to server. Hyperscale-only currently. |
| "Liquid cooling" in 2026 context | Almost certainly DLC (cold plates), increasingly immersion. See §35. |
| "PUE 1.1" or lower | DLC or immersion. Air-cooled cannot achieve this. |
| "Hyperscaler" | AWS, Google, Microsoft, Meta, Apple, Alibaba, Tencent. Operate own DCs. |
| "Cloud GPU on-demand" | End-user accesses these AI clusters via cloud APIs. The DC is hyperscaler's; the GPUs are rented by hour. |