AI Infrastructure
Power Calculator
Calculate accurate power consumption, cooling loads, electrical infrastructure requirements, and operating costs for your AI GPU deployment. From single workstations to multi-rack data center configurations, plan your facility with confidence.
Executive Summary
Unprecedented Power Density
Modern AI GPUs consume 700W-1,100W each. An 8-GPU server can draw 10kW or more, creating facility challenges that traditional IT infrastructure never faced.
Significant Operating Costs
Energy is a large and recurring share of the cost of running accelerators, and unlike hardware it is paid every year. Sizing it early prevents budget surprises and shows where efficiency work pays back.
Infrastructure Requirements
GPU deployments require proper electrical circuits, cooling capacity, and rack power distribution. Underestimating leads to costly retrofits or capacity constraints.
Plan Before You Deploy
Use this calculator to determine power, cooling, and electrical requirements before procurement. Proper planning ensures smooth deployment and optimal operations.
Understanding AI Infrastructure Power
GPU Thermal Design Power (TDP)
TDP represents the maximum sustained power a GPU can draw under typical workloads. Modern AI accelerators have dramatically higher TDP than consumer GPUs:
- Consumer GPUs: 200W-575W (RTX 4090: 450W, RTX 5090: 575W)
- Workstation GPUs: 72W-350W (L4: 72W, RTX 6000 Pro: 350W)
- Datacenter GPUs: 350W-1,400W (H100: 700W, B200: 1,000W, B300: 1,400W)
- Superchips: 2,700W+ (GB200 NVL72 per tray)
Power Usage Effectiveness (PUE)
PUE measures total facility power divided by IT equipment power. It accounts for cooling, lighting, power distribution losses, and other overhead:
Utilization Impact
GPUs rarely run at 100% TDP continuously. Actual power consumption varies by workload:
- Idle: 20-40% of TDP
- Inference (Light): 50-70% of TDP
- Inference (Heavy): 70-85% of TDP
- Training: 85-100% of TDP
Plan for 85% average utilization for production inference workloads. Training clusters may sustain 95%+ utilization.
Conversion Factors
Essential conversions for power planning:
Power Requirements Calculator
Configure your GPU deployment and get instant power, cooling, and cost estimates
Configuration
Results
Power Consumption
Annual Operating Costs
Cooling Requirements
Electrical Requirements
GPU Power Specifications
Complete power specifications for AI accelerators from NVIDIA, AMD, and Intel
| GPU Model | TDP (W) | Memory | Category | Annual Cost* |
|---|---|---|---|---|
| NVIDIA Consumer | ||||
| RTX 4090 | 450W | 24GB GDDR6X | Consumer | $335/year |
| RTX 5090 | 575W | 32GB GDDR7 | Consumer | $428/year |
| RTX 5080 | 360W | 16GB GDDR7 | Consumer | $268/year |
| RTX 5070 Ti | 300W | 16GB GDDR7 | Consumer | $223/year |
| RTX 5070 | 250W | 12GB GDDR7 | Consumer | $186/year |
| NVIDIA Workstation & Edge | ||||
| RTX 6000 Ada | 300W | 48GB GDDR6 | Workstation | $223/year |
| L4 | 72W | 24GB GDDR6 | Edge/Inference | $54/year |
| RTX 6000 Blackwell Pro | 350W | 96GB GDDR7 | Workstation/Datacenter | $260/year |
| NVIDIA Datacenter (Ampere) | ||||
| A100 40GB | 400W | 40GB HBM2e | Datacenter | $298/year |
| A100 80GB | 400W | 80GB HBM2e | Datacenter | $298/year |
| NVIDIA Datacenter (Hopper) | ||||
| H100 PCIe | 350W | 80GB HBM3 | Datacenter | $260/year |
| H100 SXM | 700W | 80GB HBM3 | Datacenter | $521/year |
| H200 SXM | 700W | 141GB HBM3e | Datacenter | $521/year |
| NVIDIA Datacenter (Blackwell) | ||||
| B100 | 700W | 192GB HBM3e | Datacenter | $521/year |
| B200 | 1,000W | 192GB HBM3e | Datacenter | $744/year |
| B300 | 1,100W | 288GB HBM3e | Datacenter | $819/year |
| GB200 NVL72 | 2,700W/tray | Combined | Superchip | $2,010/year |
| AMD Instinct | ||||
| MI300X | 750W | 192GB HBM3 | Datacenter | $558/year |
| MI325X | 750W | 256GB HBM3e | Datacenter | $558/year |
| MI355X | 900W | 288GB HBM3e | Datacenter | $670/year |
| Intel | ||||
| Gaudi 3 | 900W | 128GB HBM2e | Datacenter | $670/year |
*Annual cost assumes 85% utilization, PUE 1.5, $0.10/kWh, 24/7 operation. Actual costs vary by configuration.
Cooling Requirements
Heat Generation Basics
All electrical power consumed by GPUs converts to heat that must be removed. Use the formula: BTU/hr = Watts × 3.412
- GPU Power: 8 × 700W = 5,600W
- Heat Output: 5,600 × 3.412 = 19,107 BTU/hr
- Cooling Tons: 19,107 ÷ 12,000 = 1.6 tons
Cooling Technology Options
Rack Density Planning
Calculate cooling needs by rack configuration:
| Configuration | Power/Rack | BTU/hr | Cooling Method |
|---|---|---|---|
| 2× 8-GPU H100 servers | ~12 kW | ~41,000 | Air (Hot/Cold Aisle) |
| 4× 8-GPU H100 servers | ~24 kW | ~82,000 | Air + RDHx |
| 4× 8-GPU B200 servers | ~40 kW | ~136,000 | Direct Liquid Cooling |
| DGX GB200 NVL72 | ~120 kW | ~409,000 | Liquid Cooling Required |
Electrical Infrastructure Requirements
Voltage Requirements
Higher distribution voltage means lower current for the same power, which means smaller conductors. Common arrangements include:
- 208 V three-phase: Common in North American data centers
- 240 V single-phase: Common in smaller deployments
- 400 V or 415 V three-phase: Common in Europe and increasingly used elsewhere
The current figures this calculator produces are single-phase. They divide power by the voltage. Three-phase current is smaller for the same power, because the line-to-line voltage is multiplied by the square root of three in the denominator. If your distribution is three-phase, the figures above are conservative rather than correct, and the actual conductor and breaker sizing must come from an electrical engineer working to your local code.
NEC Compliance (80% Rule)
Per National Electrical Code, continuous loads (running 3+ hours) must not exceed 80% of circuit rating. This means circuits must be sized at 125% of the load:
Common Server Power Requirements
| Server Type | Power Draw | Amps @ 208V | Circuit Needed |
|---|---|---|---|
| 4× RTX 6000 Pro Server | ~2.0 kW | ~10A | 15A circuit |
| 8× A100 Server | ~4.0 kW | ~19A | 30A circuit |
| 8× H100 SXM Server | ~6.0 kW | ~29A | 40A circuit |
| 8× B200 Server | ~10.0 kW | ~48A | 60A circuit |
| DGX H100 | ~10.2 kW | ~49A | 60A circuit |
PDU and Rack Power
Plan Power Distribution Units (PDUs) based on total rack load:
- Basic PDU: Power distribution only, no monitoring
- Metered PDU: Displays total power consumption
- Monitored PDU: Per-outlet monitoring, remote access
- Switched PDU: Remote power cycling capability
For AI deployments, use monitored or switched PDUs for visibility and management. Plan for N+1 redundancy on critical workloads.
Electricity Rates by State
Average price of electricity to ultimate customers, from the US Energy Information Administration. Put the rate that applies to you into the calculator above rather than using an average.
| State | Commercial | Industrial | Why it appears here |
|---|---|---|---|
| Texas | 8.26 | 6.33 | Lowest commercial rate among the states shown |
| Iowa | 10.58 | 6.62 | Low industrial rate, established wind generation |
| Virginia | 10.84 | 10.53 | Largest concentration of US datacenter capacity |
| Georgia | 10.88 | 7.04 | Growing datacenter market |
| Washington | 11.72 | 7.24 | Hydro-heavy generation mix |
| Arizona | 12.15 | 7.54 | Growing datacenter market, high cooling load |
| U.S. total | 13.54 | 8.71 | National average, for reference only |
| Indiana | 13.88 | 9.08 | Growing datacenter market |
| Ohio | 13.89 | 9.87 | Growing datacenter market |
| New York | 22.40 | 9.55 | Wide gap between commercial and industrial tariffs |
| California | 24.10 | 20.20 | Highest rates among the states shown |
Which sector rate applies to you depends on how the utility classifies the load, not on what the facility is called. Large deployments are often on negotiated or industrial tariffs rather than a published commercial rate, and transmission, demand, and capacity charges may sit outside the energy rate entirely. Treat these as a starting range and confirm the tariff with the utility.
Power Efficiency Best Practices
Optimize PUE
Reducing PUE from 1.6 to 1.3 saves 18.75% on total power costs. Invest in efficient cooling, use free cooling when possible, and maintain hot/cold aisle containment.
Right-Size Deployments
Match GPU selection to workload requirements. Using RTX 6000 Pro for inference instead of H100 saves 50% power per GPU while often meeting latency requirements.
Use Power Capping
NVIDIA GPUs support power capping via nvidia-smi. Reducing H100 from 700W to 500W often provides 80% of performance at 70% of power consumption.
Schedule Workloads
Run batch training during off-peak hours when electricity rates are lower. Many utilities offer time-of-use rates with 30-50% savings at night.
Monitor Continuously
Use DCIM tools and GPU monitoring (nvidia-smi, AMD SMI) to track actual power consumption. Identify idle GPUs and optimize utilization.
Consider Location
Electricity costs vary 4× between regions. For large deployments, colocating in Texas or Nordic countries can save millions over 5 years.
Frequently Asked Questions
What exactly does this calculator include?
Accelerator board power only, multiplied by your utilization and then by PUE. It does not add CPUs, memory, storage, network adapters, fans, or power-supply losses, so the facility figure it shows is lower than your real facility load. Use it to get a starting range, then size the plant from the OEM system specification for the configuration you intend to order.
Why is board power not enough to size cooling?
Because the accelerator is only part of the chassis. Server input power adds everything else in the box plus supply losses, rack design load adds diversity and headroom, and facility input power adds cooling and electrical losses on top of the IT load. Each is larger than the one before it, and the mechanical plant is sized from the last one. Board power is useful for comparing accelerators and not much else.
Are the current and circuit figures safe to build from?
No. They are single-phase arithmetic: power divided by voltage, then multiplied by 1.25 to reflect the general principle that a continuous load should not exceed 80% of a circuit rating. If your distribution is three-phase the real current is lower, and actual conductor and breaker sizing depends on code adoption in your jurisdiction, conductor temperature rating, bundling, and the specifics of the installation. Conductor sizing is work for a licensed electrical engineer.
What is PUE and how should I set it?
Power usage effectiveness is total facility power divided by IT power, so a PUE of 1.5 means half a watt of overhead for every watt of IT load. It multiplies straight into the energy figure. PUE is an outcome of a whole facility rather than a property of a cooling technology, so if you have a measured figure for your site, use it. If you do not, treat the result as a range across plausible values rather than a single number.
Where do the electricity rates come from?
The state table is from the US Energy Information Administration's Electric Power Monthly, Table 5.6.A, for the period and retrieval date shown in the table caption. EIA labels those values preliminary estimates based on a sample. A rate you are actually offered will differ: large deployments are often on negotiated or industrial tariffs, and transmission, demand, and capacity charges may sit outside the energy rate entirely. Confirm the tariff with the utility.
Does the calculator account for redundancy?
No. There is no N+1 on power or cooling, no UPS or generator sizing, no diversity factor, and no headroom for growth. A real design carries all of those, and they increase both the capital and the provisioned capacity above what this page shows.
Formulas, assumptions, and limits
Every figure this calculator produces is an estimate derived from the inputs you set. It is planning arithmetic to get you a starting range, not a design.
The formulas
accelerator power kW = board power W x count x utilization % / 1000facility figure kW = accelerator power kW x PUEannual energy kWh = facility figure kW x hours per day x days per yearannual cost = annual energy kWh x rate per kWhheat BTU per hour = accelerator power kW x 1000 x 3.412cooling tons = BTU per hour / 12000current A = facility figure kW x 1000 / voltage(single-phase)circuit A = current A x 1.25
What the figures cover, and what they leave out
This matters more than any single number on the page. The calculator starts from accelerator board power only. It does not add CPUs, memory, storage, network adapters, fans, or power-supply losses.
- The facility figure understates real facility load. It is accelerator power multiplied by PUE, so everything else in the chassis is missing from it. In a typical 8-accelerator server that omitted overhead is hundreds of watts per chassis, and it scales with the number of servers.
- The heat and cooling figures cover the accelerators only. The heat you actually have to reject is the full server input power, which is larger.
- Board power is a single value here. Several accelerators publish a configurable range rather than a figure, and the design value comes from the OEM system. Where a range exists, this tool uses one number from it.
- The current figures are single-phase. If your distribution is three-phase, the real current for the same power is lower, so the conductor and breaker sizes implied here are conservative rather than correct.
- Utilization is applied as a flat multiplier on board power. Real draw varies through a job and does not scale linearly with a duty-cycle percentage.
- Nothing about redundancy is modelled. No N+1 on power or cooling, no UPS or generator sizing, no diversity factor, and no headroom.
To size the mechanical and electrical plant properly, start from the OEM system specification for the exact configuration you intend to order. What drives that design, and the four distinct power figures that get confused with each other, are covered in the cooling requirements guide.
On the NEC 125% factor
The 125% figure reflects the general principle in the US National Electrical Code that a continuous load should not exceed 80% of a circuit's rating. It is reproduced here because it is the arithmetic the tool applies, not as a compliance statement. Code adoption, amendments, derating for conductor temperature and bundling, and the treatment of a specific installation all vary by jurisdiction and by circumstance.
What this does not replace
The OEM system requirements, local electrical and mechanical code, a site survey, or stamped electrical and mechanical engineering. Do not procure, order conductors, or commit to a facility design from the output of this page.
Sources and basis
What the figures on this page rest on.
State electricity prices, commercial and industrial, in cents per kilowatthour.
EIA labels these preliminary estimates based on a cutoff model sample. State averages are not offered tariffs, and large loads are frequently on negotiated or industrial rates with separate demand and capacity charges.
Accelerator board power used as the starting point for every figure.
Several parts publish a configurable board-power range rather than a single figure; this tool uses one value from it. The design figure comes from the OEM system specification. Consumer and Intel entries in the dropdown are not covered by the database and carry no check date.
Every kW, kWh, BTU per hour, cooling ton, current, and cost figure the calculator outputs.
Calculated from your inputs with the formulas shown above. Covers accelerator power only, excludes server overhead and redundancy, and computes current single-phase.
The unit conversions: watts to BTU per hour at 3.412, and BTU per hour to cooling tons at 12,000.
Standard definitional conversions.
The 125% circuit-sizing factor.
Reflects the general US National Electrical Code principle that a continuous load should not exceed 80% of a circuit rating. Reproduced as the arithmetic the tool applies, not as a compliance statement. Code adoption, amendments, and derating vary by jurisdiction and installation.
The absence of an electricity share of total cost of ownership, per-server annual cost figures, and international electricity rates.
The previous version published a 30 to 40% TCO share, dollar costs per server, and rates for Iceland, Norway, and Germany, none with a source. The TCO share is an output of a specific model rather than a general fact, and SLYD has no primary source for the international rates.
Need Help Planning Your AI Infrastructure?
From hardware procurement to facility planning, SLYD provides end-to-end support for your AI deployment
GPU Hardware
Source H100, H200, B200, and MI300X systems from tier-1 OEMs including Dell, HPE, Supermicro, and Lenovo.
Explore Hardware →Equipment Financing
Financing options to spread GPU infrastructure cost over time rather than funding it from capital. Terms depend on the structure and the counterparties.
Learn About Financing →Cloud Compute
On-demand GPU compute from SLYD marketplace providers, without a hardware procurement cycle.
Browse Compute →Consulting Services
Expert guidance on infrastructure design, facility planning, and deployment optimization.
Contact Us →