2.1b Single-phase vs three-phase power and why AI racks need three-phase¶
🌐 Context Introduction¶
Before we talk about GPUs, AI models, or cooling systems, we need to understand the very first thing that makes an AI data center work: electricity. Power comes into a facility from the utility grid, and how that power is delivered—single-phase vs three-phase—determines what kind of equipment you can run. For AI workloads, the power demands are enormous, and that's where three-phase power becomes essential. This section breaks down the difference in simple terms for new engineers.
⚡ What Is Single-Phase Power?¶
Single-phase power is the standard electricity you find in most homes and small offices. It uses two wires: one "hot" wire and one neutral wire.
Key characteristics: - Voltage typically ranges from 120V to 240V (depending on region) - Power delivery is pulsed — it drops to zero 120 times per second (in 60 Hz systems) - Suitable for low-power devices: lights, small appliances, desktop computers - Maximum practical load is around 10–15 kW per circuit
Think of it like: A single-lane road. Traffic flows, but if too many cars try to use it at once, everything slows down or stops.
⚡⚡⚡ What Is Three-Phase Power?¶
Three-phase power uses three "hot" wires (plus a neutral in some configurations). The three waveforms are offset by 120 degrees from each other.
Key characteristics: - Voltage typically ranges from 208V to 480V (common in data centers) - Power delivery is continuous — at least one phase is always at peak voltage - Suitable for high-power devices: large motors, industrial equipment, AI servers - Can handle loads from 50 kW to over 1 MW per circuit
Think of it like: A three-lane highway. Traffic flows smoothly because vehicles can use all lanes simultaneously, and there's always room for more.
📊 Single-Phase vs Three-Phase — Comparison Table¶
| Feature | Single-Phase | Three-Phase |
|---|---|---|
| Number of wires | 2 (hot + neutral) | 3 or 4 (3 hot + optional neutral) |
| Voltage (typical) | 120V / 240V | 208V / 480V |
| Power delivery | Pulsed (drops to zero) | Continuous (always delivering) |
| Efficiency | Lower (more energy lost as heat) | Higher (less energy lost) |
| Typical load capacity | Up to ~15 kW | 50 kW to 1 MW+ |
| Common use cases | Homes, small offices | Industrial, data centers, AI racks |
| Cost of equipment | Lower | Higher (but more efficient at scale) |
🧠 Why AI Racks Need Three-Phase Power¶
AI racks are not like standard server racks. A single AI server with 8 GPUs (like an NVIDIA DGX system) can draw 6–10 kW or more. When you fill a rack with multiple such systems, the power demand quickly exceeds what single-phase can provide.
Here's why three-phase is non-negotiable for AI:
- Higher power density — A single AI rack can consume 40–80 kW or more. Single-phase circuits simply cannot deliver that much power safely or efficiently.
- Reduced current draw — Three-phase delivers the same power with lower current. Lower current means thinner cables, less heat, and safer operation.
- Balanced load — Three-phase distributes power evenly across all three phases, preventing overload on any single circuit.
- Efficiency at scale — Three-phase power transmission loses less energy as heat compared to single-phase, which is critical when you're running hundreds of racks.
- Future-proofing — As AI models grow, power demands will only increase. Three-phase infrastructure can scale to meet those demands.
Real-world example: - A single NVIDIA DGX H100 system draws approximately 7 kW - A rack with 4 such systems = 28 kW - A rack with 8 such systems = 56 kW - Single-phase max capacity per circuit ≈ 15 kW — you'd need 4 separate circuits - Three-phase at 480V can deliver 50+ kW on a single circuit
📊 Visual Representation: Single-Phase vs. Three-Phase Power Delivery¶
This diagram contrasts single-phase power delivery (pulsed and capacity-limited) with three-phase power delivery (continuous and highly scalable) which is necessary for modern high-density AI clusters.
🛠️ Practical Takeaway for New Engineers¶
When you walk into an AI data center, look at the power distribution units (PDUs) and the cables feeding the racks. If you see three thick cables (plus a ground) and a large connector, that's three-phase power. If you see only two wires and a smaller plug, that's single-phase — and it's likely powering non-AI equipment like networking gear or management servers.
Remember this rule of thumb: - Single-phase = small loads (monitors, switches, lights) - Three-phase = big loads (AI servers, GPU clusters, cooling systems)
🕵️ Quick Check — Can You Spot the Difference?¶
| Scenario | Power Type Needed | Why? |
|---|---|---|
| A single desktop workstation | Single-phase | Low power draw (~500W) |
| A rack of 8 NVIDIA H100 servers | Three-phase | ~56 kW total load |
| A network switch | Single-phase | Typically under 1 kW |
| A row of 20 AI racks | Three-phase | 1 MW+ total load |
| Office lighting | Single-phase | Very low power per circuit |
✅ Summary¶
- Single-phase power is for low-power equipment and is limited to about 15 kW per circuit.
- Three-phase power delivers continuous, high-capacity electricity needed for AI racks.
- AI servers draw massive power (6–10 kW per server), making three-phase the only practical choice.
- Three-phase reduces current, improves efficiency, and allows for balanced power distribution.
As a new engineer, understanding this distinction is your first step toward designing, operating, or troubleshooting AI infrastructure. When in doubt: if it has GPUs, it needs three-phase.