AI Data Center Cost Calculator

CAPEX (Capital Expenditure)

One-time upfront costs: GPU hardware, servers, racks, building construction, cooling systems, UPS, generators, networking, and installation. Paid once at the start.

OPEX (Operating Expenditure)

Recurring annual costs: electricity bills, staff salaries, hardware maintenance, software licenses, insurance. Paid every year for the lifetime of the facility.

PUE (Power Usage Effectiveness)

Ratio of total facility power to IT equipment power. PUE 1.0 = perfect efficiency (impossible). PUE 1.3 = 30% overhead goes to cooling/lighting. Lower is better. World-class AI facilities achieve 1.1-1.2.

TCO (Total Cost of Ownership)

CAPEX + (OPEX x years). The true cost of owning infrastructure over its full lifetime. Used to compare against cloud hosting costs for break-even analysis.

System Power (Watts per GPU)

Total server power draw per GPU slot — includes the GPU itself plus CPU, RAM, NVMe storage, power supply losses, and fans. Typically 1.5-2x the GPU TDP alone. An H100 at 700W TDP draws ~1400W at the system level.

Server BOM Multiplier (1.35x)

GPUs are ~74% of a server's bill of materials. The 1.35x multiplier adds CPU, motherboard, RAM, NVMe drives, power supplies, chassis, and management controllers on top of GPU cost.

Construction Cost per MW

The cost to build data center facilities per megawatt of IT load capacity. Varies dramatically by location: $5.5M/MW in India vs $16M/MW in Singapore. Includes civil works, building shell, electrical infrastructure, and fit-out.

Redundancy Tier (II / III / IV)

Uptime Institute classification. Tier II: single path, 99.741% uptime. Tier III: concurrently maintainable, 99.982%. Tier IV: fault tolerant, 99.995%. Higher tiers multiply power/cooling infrastructure costs by 1.35x-1.70x.

UPS (Uninterruptible Power Supply)

Battery-backed power system that provides instant failover during grid outages (holds load for 5-15 minutes until generators start). Sized at 125% of facility load for safety margin. Cost: ~$400/kW capacity.

Cooling Capacity (Tons)

Measured in refrigeration tons (1 ton = 3.517 kW of heat removal). Formula: kW facility load x 0.284. A 1MW facility needs ~284 tons of cooling — roughly equivalent to cooling 50 large homes simultaneously.

FP16 / FP8 TFLOPS

Trillion floating-point operations per second at 16-bit or 8-bit precision. FP16 is used for AI training; FP8 for inference. Higher = faster AI model training. B200 at 4500 FP16 TFLOPS is 2.3x faster than H100 at 1979.

HBM (High Bandwidth Memory)

Stacked memory chips soldered directly to the GPU die. HBM3e (used in B200/H200) delivers 8,000 GB/s bandwidth — critical for LLM inference where memory bandwidth is the bottleneck, not compute.

InfiniBand vs Ethernet

InfiniBand provides ultra-low latency GPU-to-GPU communication (1-2 microseconds) critical for distributed training across hundreds of GPUs. Costs $4,000/GPU vs $800-2,500 for Ethernet. Required for clusters above 256 GPUs doing LLM training.

Cloud Break-Even

The month when cumulative on-prem costs (CAPEX + running OPEX) become cheaper than renting equivalent cloud GPUs. Formula: CAPEX / (monthly cloud cost - monthly on-prem OPEX). Typically 12-24 months at 500+ GPU scale.

Contingency (10%)

Standard engineering safety margin added to total CAPEX. Covers unforeseen costs: supply chain delays, design changes, permitting issues, scope creep. Industry standard is 10-15% for data center construction.

Liquid vs Immersion Cooling

Liquid cooling pipes coolant to GPU cold plates (16 GPUs/rack, $8K/rack). Immersion submerges entire servers in dielectric fluid (32 GPUs/rack, $25K/rack). Both allow 2-4x rack density vs air cooling, reducing floor space and improving PUE.