Estimate wattage, monthly electricity cost, and BTU/hr for any local LLM rig
Running a local LLM is not free — even with no API fees, the GPU pulls real watts. This tool turns your hardware + usage pattern into a monthly electricity bill, a heat load in BTU/hr, and a friendly “light-bulb equivalent.” Built for home-lab users sizing their power budget before buying a card.
Monthly cost
$7.63
kWh / day
1.59kWh
kWh / month
47.7kWh
Annual cost
$92
Heat output
981BTU/hr
100W bulb equiv.
11.5bulb·h/day
Note: Inference draw assumes a 50% duty cycle on the RTX 5090 TDP (575W). Small models (≤8B) under-utilize big GPUs; large models (70B+) saturate them. Math is performed in USD; currency selector changes the display symbol only.
How it works
1. Inference watts = TDP × duty cycle. Sustained LLM inference draws ~50% of TDP on small models (the GPU is bandwidth-bound) and ~85% on 70B+ models (compute-saturated).
2. Daily kWh = (inference_W × hours + idle_W × (24 − hours)) ÷ 1000. Idle draw matters: 20W × 24h still adds up over a month.
3. Monthly cost = kWh/day × 30 × $/kWh. US average residential rate is $0.16/kWh (EIA 2025).
4. BTU/hr = watts × 3.412. This is the heat your GPU dumps into the room — useful for sizing AC or planning a winter heater replacement.
When you’d use this
Sizing your home-lab power budget
Before you buy a 575W RTX 5090, find out if your 15A circuit can carry it alongside your monitor and PC.
Deciding if local is cheaper than API
Pair with the API Cost vs Local calculator to see whether your local rig actually beats GPT-4o at your token volume.
Planning cooling and room temperature
A 700W H100 dumps ~2,400 BTU/hr into the room — about the same as a small space heater. Plan AC accordingly.
Related Runyard tools
VRAM Calculator →
See whether the model you want will actually fit on the GPU you priced out.
API Cost vs Local Cost →
Compare your local electricity + hardware bill against monthly OpenAI / Anthropic / Google API spend.
Local LLM Cost Savings →
Build the full local-vs-cloud cost story for sustained, high-volume usage.
More on the tradeoff: read our blog deep-dives on GPU power, thermals, and cost.
FAQ
It depends on your GPU and how many hours you run it. An RTX 4090 (450W TDP) running a 70B model for 4 hours/day at 80% duty cycle uses about 45 kWh/month — roughly $7 at the US average rate of $0.16/kWh. An H100 running 24/7 at full tilt can exceed 500 kWh/month or $80+.
Usually about the same or slightly less. Modern AAA gaming pushes a GPU to 90–100% of TDP. LLM inference typically sits at 60–85% — a 70B model on an RTX 4090 draws roughly 360W vs 430W in Cyberpunk 2077. The bigger difference is duration: people often run inference longer per session than they game.
The RTX 5090 has a 575W TDP. Running a 70B model like Llama 3 at ~80% duty cycle, it draws about 460W, which equals 1,570 BTU/hr of heat. That is comparable to a 500W space heater — enough to noticeably warm a small room.
For light usage (a few prompts a day), ChatGPT Plus at $20/month is cheaper than local. For heavy usage (millions of tokens), local wins fast — even after factoring in $30–50/month of electricity, you avoid per-token API fees that scale with use. Use the API Cost vs Local calculator on Runyard to see your exact break-even.
Monthly cost = (inference_watts × active_hours + idle_watts × idle_hours) × 30 days ÷ 1000 × your $/kWh rate. The hard part is the inference watts: a GPU rarely runs at full TDP. For LLMs assume ~70% of TDP for medium models and ~85% for 70B+ models. This calculator does that math for you.
Power draw scales with how hard the model pushes the GPU. A 7B model barely warms up an H100 (compute-bound at maybe 40% TDP) but maxes out an RTX 4060. The duty cycle in this calculator approximates that — smaller models on big GPUs are less efficient per watt because much of the silicon sits idle.