How Much Does a GPU Cost to Run? Power Draw and Monthly Electricity

2026年7月5日

I run Ollama LLM inference and ComfyUI image generation on my home Ubuntu PC almost every day, and one day I found myself wondering: “How much is this GPU actually costing me in electricity?"

My setup is a two-card build: an RTX 3090 (TDP 350W) and an RTX 3060 12GB (TDP 170W). “Sure, it must eat power," I figured — but I had no real idea how much it was actually using.

So I looked into it. In this article I pull together the power draw and electricity-cost math for a running GPU, based on manufacturer spec sheets (TDP) and user measurements. The bottom line: there’s about a 6× difference between idle and full load, and by picking the right GPU for the job and managing power, you can reportedly save more than ¥1,000 a month.

※ Prices and rates are as of April 2026.

What a watt meter is

When you want to actually measure your home power draw, the tool for the job is a watt meter. You just slip it between the outlet and the power plug, and it shows the power draw of the connected device in real time.

Using it really is simple: plug the watt meter into the wall outlet, then plug your PC’s power cable into the watt meter. That’s it — now you can see, as a number, how much electricity your PC is using right now.

The price is around ¥1,500–4,500. Think of it as identifying the “culprit" behind your monthly bill, and it’s a cheap investment.

Recommended products (comparison table)

Product	Reference price (as of April 2026)	Wi-Fi	Logging	Accuracy	Best for
Sanwa Supply TAP-TST8N	about ¥2,000	×	×	±1.5%	Those who want quick real-time readings. Easy-to-read display, ideal for beginners
Ratoc Systems RS-WFWATTCH1	about ¥7,000–9,000	○	○ (app)	±1%	Remote monitoring from your phone. Those who want long-term logs
ELPA EC-05EB	about ¥2,600–3,500	×	×	±2%	Those for whom simple features are enough. Bare-minimum display
SwitchBot Plug Mini	about ¥2,480	○ (BLE+Wi-Fi)	○ (app)	±2%	Those who want smart-home integration. Remote power on/off too

[kimono_product id="15852″]

How to read this table: Products with “○" for Wi-Fi let you check power draw in real time from a phone app, so you can monitor even while you’re out. Products with “○" for logging let you look back at power trends by day, week, or month as a graph. For accuracy, a smaller number is more precise; for everyday electricity-cost checks, even ±2% is plenty practical.

The standard pick is the Sanwa Supply TAP-TST8N (about ¥2,000, as of April 2026). The wattage shows on the LCD in real time, and one review I saw noted that it’s genuinely fun to watch the number jump the instant you load up the GPU. It’s simple, the LCD is easy to read, and for “I just want to measure power draw" it’s said to be more than enough.

If you’re choosing on simplicity, the ELPA EC-05EB (about ¥2,600–3,500). It has only basic features, but there’s no problem starting with it as your first watt meter. Conversely, if you want to “check power-draw logs from your phone" or “monitor while away from home," the Ratoc Systems RS-WFWATTCH1 (about ¥7,000–9,000) is Wi-Fi capable and lets you check logs from an app. The SwitchBot Plug Mini (about ¥2,480) also does smart-home integration and keeps a power-draw log, so if you want logs on a budget, it’s another good pick.

Recommended patterns

Pattern	Product	Cost (as of April 2026)	Best for
Just want to measure	ELPA EC-05EB	about ¥2,600–3,500	First-time measurers
Want to keep logs	SwitchBot Plug Mini	about ¥2,480	Those who want to check from a phone
Serious monitoring	Ratoc Systems RS-WFWATTCH1	about ¥7,000–9,000	Those who want 24-hour monitoring

How to read this table: Top to bottom, it runs “simplicity-first → features-first." When in doubt, the middle option, the SwitchBot Plug Mini, is a well-balanced recommendation. If you just want to glance at real-time numbers, the Sanwa Supply TAP-TST8N (about ¥2,000) is plenty.

Research data: how much power draw changes with GPU load

Below are power-draw estimates for a build matching my own (RTX 3090 + RTX 3060 12GB, CPU: Core i7-11700, 64GB RAM, Linux), organized from manufacturer spec sheets (TDP) and various reviews and user measurements.

Power-draw estimates (detailed)

State	GPU in use	Estimated GPU power (W)	Estimated whole-PC (W)
Idle (desktop left sitting)	—	—	80–120
Ollama 8B inference (llama3.1:8b, etc.)	RTX 3060	100–140	220–280
Ollama 14B inference (qwen2.5:14b, etc.)	RTX 3060	130–170	250–320
Ollama 32B inference (qwen3:32b, etc.)	RTX 3090	250–320	400–500
ComfyUI SDXL image generation	RTX 3090	280–340	420–520
ComfyUI FLUX image generation	RTX 3090	300–350	450–550
RTX 3060 only, lightweight-model inference	RTX 3060	130–170	250–350
Both cards at full load (3090+3060)	Both	480–520 total	600–700

How to read this data: The left column is the load state, and the “estimated whole-PC" column on the right is roughly the number you’d see on a watt meter. Because it includes the CPU, memory, storage, and more beyond the GPU, it tends to run higher than the GPU-only figure. You can see power draw climbing as model size grows (8B → 14B → 32B). Note that for GPU-only power draw, you can also check it in real time with the nvidia-smi command.

What’s worth noting is the gap between idle and the RTX 3090 at full load. 80W versus 500W is about a 6× difference. Many user reports say “the instant I start image generation, the watt meter number shoots up," and seeing it as a number really drives home the feeling of “I’m using power."

Another interesting point: for the same AI task, if the RTX 3060 can handle it, the whole PC stays around 250–350W. Just by choosing between “a job for the 3090" and “a job the 3060 can handle," power draw changes considerably.

By GPU: monthly electricity-cost simulation (scatter plot data)

Let’s compare “how much does mine cost per month?" side by side.

How to read this graph

In the scatter plot below, the horizontal axis is the GPU’s TDP (maximum power draw) and the vertical axis is monthly electricity cost (¥); further right means a more power-hungry GPU, and higher up means a bigger bill. Because a GPU is often said to run at about 70% of TDP in real use, this is calculated as “TDP × 0.7 × 8 hours × 30 days × ¥30/kWh."

Monthly electricity cost by GPU TDP
Calculation = TDP × 0.7 (average load factor) × 8 hours/day × 30 days × ¥30/kWh

RTX 3060     TDP: 170W  Monthly cost: ¥858
RTX 4060 Ti 16GB  TDP: 165W  Monthly cost: ¥832
RTX 3090     TDP: 350W  Monthly cost: ¥1,764
RTX 5070 Ti  TDP: 300W  Monthly cost: ¥1,512
RTX 5080     TDP: 360W  Monthly cost: ¥1,814
RTX 4090     TDP: 450W  Monthly cost: ¥2,268
RTX 5090     TDP: 575W  Monthly cost: ¥2,898

Scatter plot data (table version)

GPU	TDP (W)	Average load (W)※	Monthly usage (kWh)	Monthly cost (¥)
RTX 3060	170	119	28.6	about ¥858
RTX 4060 Ti (16GB)	165	116	27.7	about ¥832
RTX 5070 Ti	300	210	50.4	about ¥1,512
RTX 3090	350	245	58.8	about ¥1,764
RTX 5080	360	252	60.5	about ¥1,814
RTX 4090	450	315	75.6	about ¥2,268
RTX 5090	575	403	96.6	about ¥2,898

※ Average load = TDP × 0.7. Formula: average load (kW) × 8h × 30 days × ¥30/kWh

How to read this table: These are the “GPU-only" electricity costs, calculated from the GPU’s power alone. In reality the whole system — CPU, memory, and so on — adds to the power draw, so for the whole PC it comes out about ¥1,000–2,000 higher than the values in this table. An RTX 3060-class card is under ¥1,000/month; an RTX 5090 is about ¥3,000. Use it as a rough electricity-cost guide when choosing a GPU.

Monthly electricity-cost simulation

Using the power-draw estimates above, let’s compute the monthly electricity cost. The rate is the national average of about ¥30/kWh (as of April 2026, roughly equivalent to the Juryo-dento B plan). It moves up or down depending on your region and contract, but as a rough guide this figure is fine.

Running AI on an RTX 3090 for 8 hours a day

Average whole-PC power draw: about 450W (0.45kW)
Daily usage: 0.45kW × 8 hours = 3.6kWh
One month (30 days): 3.6kWh × 30 days = 108kWh
Cost: 108kWh × ¥30 = about ¥3,240/month

Running lighter AI on an RTX 3060 for 8 hours a day

Average whole-PC power draw: about 300W (0.30kW)
Daily usage: 0.30kW × 8 hours = 2.4kWh
One month (30 days): 2.4kWh × 30 days = 72kWh
Cost: 72kWh × ¥30 = about ¥2,160/month

Left on at idle 24 hours a day

Average whole-PC power draw: about 100W (0.10kW)
Daily usage: 0.10kW × 24 hours = 2.4kWh
One month (30 days): 2.4kWh × 30 days = 72kWh
Cost: 72kWh × ¥30 = about ¥2,160/month

Cloud AI vs local AI: monthly cost comparison

Here’s an answer to the question “How does local AI’s electricity cost compare to cloud AI’s monthly subscription?"

How to read this table

Below is a comparison putting cloud AI services’ monthly fees next to the whole-PC electricity cost of running local AI 8 hours a day. Local AI’s electricity cost includes not just the GPU but the whole system — CPU, memory, and so on. It does not include the initial investment (buying the GPU).

Item	Monthly cost	Notes
ChatGPT Plus	¥3,000/month	Access to the latest GPT models (as of April 2026)
Claude Pro	¥3,000/month	Extended Claude usage (as of April 2026)
RTX 3060 local AI	about ¥2,160/month	Whole PC about 300W × 8h × 30 days × ¥30/kWh
RTX 3090 local AI	about ¥3,240/month	Whole PC about 450W × 8h × 30 days × ¥30/kWh
RTX 5090 local AI	about ¥5,040/month	Whole PC about 700W × 8h × 30 days × ¥30/kWh

[Bar chart data] Cloud AI vs local AI monthly cost comparison

RTX 3060 local:  ¥2,160 (green)
ChatGPT Plus:    ¥3,000 (blue)
Claude Pro:      ¥3,000 (purple)
RTX 3090 local:  ¥3,240 (orange)
RTX 5090 local:  ¥5,040 (red)

How to read this graph: A shorter bar means lower cost. Running local AI on an RTX 3060 comes out cheaper than ChatGPT Plus or Claude Pro. On an RTX 3090 it’s about the same. At the RTX 5090 class, cloud AI is the better value. That said, local AI has advantages the cloud doesn’t: privacy, no usage caps, and no internet required.

The comparison with ChatGPT Plus is interesting

Using AI 8 hours every day on an RTX 3090 costs about ¥3,240/month in electricity. Meanwhile, ChatGPT Plus is ¥3,000/month (as of April 2026).

Nearly the same amount.

Of course, local AI carries the up-front GPU investment, and there are plenty of situations where the latest models available on ChatGPT Plus are smarter. Still, considering local AI’s upsides — privacy, freedom to customize, and running without the internet — the electricity cost seems acceptable.

Personally, once I frame it as “the electricity for my home AI setup is covered by the price of a ChatGPT subscription," it’s not a bad deal at all.

Tips for cutting the electricity cost

Based on what I found, here are power-saving habits I actually practice in my own setup.

1. Put the GPU to sleep when you’re not using it

On Linux, NVIDIA GPUs automatically enter a low-power mode when idle, but with something like Ollama resident, the GPU can struggle to fall back to idle. Simply unloading the Ollama model when a task finishes (ollama stop model-name) helps.

2. Send light tasks to the secondary GPU (RTX 3060)

For jobs where an 8B-class lightweight model is enough, handing them to the RTX 3060 lowers power draw substantially. I recommend the split of using the 3090 only when running big models like 32B or 70B (quantized).

3. Turn the PC off when you’re not using it

Obvious, but on paper even leaving it idle costs more than ¥2,000 a month. I have a soft spot for the idea of “an AI server running 24/7," but considering the electricity, it’s more realistic to boot it only when needed.

Handy extras to have on hand

If you’re going to measure with a watt meter, it’s worth getting the power-side environment in order at the same time. Because a GPU at full load pulls large surges of power momentarily, using a surge-protected power strip gives peace of mind.

• Power strip (surge-protected): Sanwa Supply TAP-SP2110-1 (about ¥2,500, as of April 2026). Protects equipment from lightning surges. With two GPUs installed, you’ll want power-side safety measures in place
• Extension cord: Elecom T-ADR5-2620WH (about ¥2,600–3,500, as of April 2026). Slotting in a watt meter tends to crowd the outlet area, so an extension cord makes things easier to route

You can measure without either, but if you’re going to run a GPU setup seriously, having them on hand makes life more comfortable.

[kimono_bar title="Whole-PC power draw by GPU load (rough guide from aggregated reviews)" unit="W" color="#1565c0″ highlight="6,7″ max="750″ note="Assumes RTX 3090 + RTX 3060 + i7-11700 + 64GB RAM"]
Idle (left alone)|100
Ollama 8B (RTX 3060)|250
Ollama 14B (RTX 3060)|285
Ollama 32B (RTX 3090)|450
ComfyUI SDXL (RTX 3090)|470
ComfyUI FLUX (RTX 3090)|500
Both cards at full load|650
[/kimono_bar]

[kimono_bar title="Cloud AI vs local AI: monthly cost comparison" unit="¥" color="#1976d2″ max="6000″ note="Local AI is electricity only (8h/day, ¥30/kWh). Excludes upfront GPU cost."]
RTX 3060 local|2160
ChatGPT Plus|3000
Claude Pro|3000
RTX 3090 local|3240
RTX 5090 local|5040
[/kimono_bar]

Summary: the value of “making it visible" for a few thousand yen

Before I looked into it, my understanding was a vague “GPUs apparently eat power." But once I organized it into numbers, it became concrete: there’s a 6× difference between idle and full load, and the monthly bill can top ¥3,000.

If you want to check your own home’s numbers, a single ¥1,500–4,500 watt meter makes all of this visible. Just “making it visible" changes how conscious you are of your usage, and it becomes motivation to pick the right GPU for the job and to manage power.

For anyone who uses local AI daily, or is about to start, I recommend getting a watt meter along with the GPU.