Energy Analysis

The Hidden Cost: GPU Electricity Expense Per Training Run

An H100 draws 700W at peak. An 8-GPU cluster running 24/7 costs $4,900/year in electricity alone. But nobody tracks this cost — it doesn't appear on your cloud bill, and most on-premise teams can't attribute it to specific workloads.

Updated June 2026·10 min read

GPU Power Draw: The Numbers

Every GPU has a Thermal Design Power (TDP) — the maximum sustained power draw. But real power consumption depends on workload intensity, memory access patterns, and thermal conditions. Here are the actual TDP ratings for popular AI GPUs:

GPUTDP (W)Idle (~15%)FP16 TFLOPSMemoryArchitecture
H100 SXM700W105W98980 GBHopper
H200 SXM700W105W989141 GBHopper
A100 SXM 80GB400W60W31280 GBAmpere
A100 SXM 40GB400W60W31240 GBAmpere
RTX 4090450W68W16524 GBAda Lovelace
L40S350W53W36248 GBAda Lovelace
AMD MI300X750W113W1307192 GBCDNA3
AMD MI250X500W75W383128 GBCDNA2
T470W11W6516 GBTuring

Key insight: an idle GPU still draws 15% of TDP. An idle H100 consumes ~105W — costing $0.013/hr in electricity even when doing nothing. Across an 8-GPU cluster sitting idle for a month, that's $73 in electricity for zero compute.

Electricity Cost Per Training Run

We modeled energy consumption for common ML workloads using real TDP data and typical utilization profiles. All costs use the US average electricity rate of $0.12/kWh (European rates at $0.25/kWh shown for comparison).

WorkloadGPUsDurationkWhUS CostEU CostCO₂ (kg)
Fine-tune Llama 3 8B (LoRA)
LoRA fine-tune on custom dataset, 3 epochs
1x RTX 40904h1$0.16$0.340.5
Fine-tune Llama 3 70B (QLoRA)
QLoRA 4-bit, distributed across 4 GPUs
4x A100 SXM 80GB12h13$1.54$3.225.0
Pre-train 7B model from scratch
Full pre-training, 1 week on 8-GPU node
8x H100 SXM7d776$93.14$194.04298.8
Pre-train 70B model
Full pre-training, ~1 month, 8-node cluster
64x H100 SXM30d27,689$3322.63$6922.1410660.1
Inference server (vLLM, 24/7)
Serving a 70B model, typical production load
2x A100 SXM 80GB1 mo128$15.42$32.1249.5
Idle cluster (95th percentile reality)
8 GPUs sitting idle — the industry average
8x H100 SXM1 mo593$71.13$148.19228.2

The Idle Cost Problem

The last row in the table above tells the real story. Industry research consistently shows average GPU utilization of just 5%. An 8-GPU H100 cluster sitting idle for a month wastes over $73 in electricity alone — and over $23,000/month in cloud compute costs for GPUs doing essentially nothing.

The electricity cost of idling is small compared to the cloud rental cost, but it's a useful signal: if you can measure power draw, you can measure utilization, and if you can measure utilization, you can eliminate waste.

The Formula

GPU energy cost per hour follows a simple formula:

Energy Cost ($/hr) = (TDP_W × Utilization × GPU_Count) / 1000 × Electricity_Rate

Example: 8x H100 at 90% utilization, $0.12/kWh

= (700 × 0.90 × 8) / 1000 × $0.12 = $0.605/hr = $441/mo

Electricity vs. Cloud Compute: The Ratio

For cloud deployments, electricity is already baked into the instance price — you pay it indirectly. But for on-premise or colocation deployments, electricity is a direct, measurable cost:

  • H100 SXM: Electricity is ~2% of total cloud cost, ~20% of on-prem TCO
  • RTX 4090: Electricity is ~6% of total cloud cost (lower cloud price, high TDP)
  • T4: Electricity is ~2% of total cloud cost (low TDP, low price)

For cloud users, the bigger win is eliminating idle compute time. For on-prem teams, both idle compute and direct electricity savings matter.

Carbon Footprint

Every kWh of electricity produces CO₂ emissions depending on the local grid mix. Using the US average of 385g CO₂/kWh:

  • Fine-tuning Llama 3 8B: ~1 kg CO₂ (equivalent to driving 2.5 miles)
  • Pre-training a 7B model: ~1,700 kg CO₂ (equivalent to a cross-country flight)
  • Pre-training a 70B model: ~100,000+ kg CO₂ (equivalent to 15 cars for a year)

The EU AI Act (Article 13) requires GPAI model providers to report energy consumption during training. NemulAI provides automated energy metering and compliance-ready reports. Read our EU AI Act compliance guide →

How to Measure Real Energy Costs

The estimates above use TDP as a proxy, but real power draw varies with workload. The only way to know your actual energy cost is to measure it:

# Install the agent (works on NVIDIA, AMD, Intel, Apple Silicon)

pip install nemulai

# Start measuring

nemulai

# See energy cost per job

nemulai report --format csv --with-carbon

Try the Calculator

Model your own GPU electricity costs with our free GPU Cost Calculator. Includes compute costs, electricity, and CO₂ footprint across 9 providers.

Measure real energy costs, not estimates

NemulAI measures actual GPU power draw per job with 5-second resolution across 6 hardware platforms.