Intermediate 15 minMini-PC

Power consumption of a local GB10 AI: our mesures

A monitor on a desk displays colored line charts.
Direct response

At idle, the GPU in our AI TOP ATOM draws 10.4 W. Under load, it draws 39 to 53 W on average, with peaks up to 95 W. A private assistant response uses 0.15 Wh, or 0.003 euro cents; a full LoRA training run uses 76.7 Wh, or 1.5 euro cents. These figures cover the GPU alone, read from the chip, during our per-task measurements from October 6 to 8, 2026; a continuous 30-day reading is in progress.

How much electricity does an AI running at home use? We recorded the GPU power draw of our AI TOP ATOM (NVIDIA GB10 chip, 128 GB), generously provided by GIGABYTE, during typical tasks, then continuously since the evening of October 6, 2026. These figures cover the GPU only, read from the chip, so treat them as indicative. This page will be updated when the 30-day recording ends.

By Mohamed Meguedmi·Update 2026-11-19·Measured on GIGABYTE AI TOP ATOM
i
Transparency
Hardware provided free of charge by GIGABYTE for this series of guides. The measurements and opinions are ours; GIGABYTE neither reviewed nor approved this content before publication. This page contains no affiliate links. Our data, infographics, and photos may be freely reused under the CC BY 4.0 license, with attribution to BestLLMfor.
Key figures
10,4W
consumed by the GPU at idle
0,15Wh
per private assistant response, GPU side
3centimes
of GPU electricity per 1,000 responses, at the EDF Base rate
95W
maximum peak across our tasks, for a 240 W adapter
The AI TOP ATOM held in one hand, shown from a three-quarter angle: the enclosure barely extends beyond the palm.
Handheld: the AI TOP ATOM weighs 1.2 kg and has a volume of about one liter. Photo of our unit, with the background neutralized. Enlarge ↗

#What exactly are we measuring?

This guide covers the GPU power of the GB10 chip, as reported by the chip itself. We read it with nvidia-smi, the NVIDIA tool: one reading per second during each task, then one reading every 10 seconds for the continuous log. NVIDIA specifies that this value represents the GPU only.

The measurements describe GPU power as reported by the chip. The CPU, memory, SSD, network, fans, and power supply losses are not counted.

i
GPU only
All watts and watt-hours on this page are for the GPU, as read from the chip. The whole machine draws more. Further down, wall-outlet measurements from other sources give a ballpark figure for the total.

The continuous 30-day logging run is in progress: it started on October 6, 2026 at 23:46. This page covers its first few days. It will be updated when the run ends, scheduled for November 5, 2026, and each update will be dated.

#How do you read watts, watt-hours, and euros?

Power is measured in watts (W): it is what the GPU draws at a given moment. Energy is measured in watt-hours (Wh): it is power multiplied by time. 1000 Wh make 1 kWh, the unit on your electricity bill. For example, 50 W for 2 hours makes 100 Wh. Memory sizes are in GB (1 GB = 1024 MB, as Linux displays it).

The rate is that of EDF's regulated Tarif Bleu plan, Base option, 6 kVA: 20.01 euro cents per kWh including tax, according to the rate schedule in effect on August 1, 2026. The subscription fee isn't counted for the ATOM, since you'd pay it without it. Your contract may have a different price.

The machine is a GIGABYTE AI TOP ATOM (128 GB of memory, NVIDIA GB10 chip, DGX OS 7.5.0). The per-task readings come from our test runs from October 6 to 8, 2026. We locked our protocol before the first measurement; the method and the data behind our tables are published on our methodology page.

#How much power does the GPU draw at idle?

At idle, with no tasks running, over 30 minutes, the GPU draws an average of 10.4 W, ranging from 10.2 to 11.0 W. That's roughly the power of a living-room LED bulb.

Over 24 hours, 10.4 W comes to 0.25 kWh, or 5 euro cents of electricity. Over 30 days, that's 7.5 kWh, or €1.50. These figures count the GPU only.

#How much power does the GPU draw under load?

We logged GPU power during six typical tasks, with idle as the baseline: a private assistant answering questions about documents, indexing those documents, LoRA training, QLoRA training, reading 200 invoices, and a 2-hour service session for 8 users.

→
Model and tool names
gpt-oss 120B, Llama 3.3 70B, and Qwen3-8B are language models. bge-m3 is a model that prepares documents for search. vLLM is a model server and Unsloth is a training tool. LoRA is a lightweight fine-tuning method; QLoRA is the same method with weights compressed to 4 bits. NVFP4 is a compressed format from NVIDIA.
Bar chart of average GPU power by task (orange: private assistant and indexing; purple: training; blue: invoice reading and serving with vLLM; light blue: idle), with the peak in a pale shade and the duration on the right: indexing 5,000 pages 53.2 W (peak 64.1 W, 6.1 min), LoRA training 53.0 W (62.7 W, 86.8 min), private assistant with gpt-oss 120B 50.5 W (68.1 W, 18.0 min), QLoRA of Llama 3.3 70B 49.3 W (81.2 W, 10.8 min), reading 200 invoices 41.0 W (88.5 W, 20.2 min), 2-hour serving for 8 users 39.1 W (95.0 W, 120.8 min), idle 10.4 W (11.0 W, 30 min).
GPU power in watts by task. Solid bar (∅): average power. Pale section and triangle (▲): peak. Clock: duration in minutes. Icons: moon for idle, speech bubble for private assistant responses, documents for indexing, looping arrows for training, document for invoice reading, group of people for multi-user serving. Colors: orange for the private assistant and indexing, purple for training (LoRA and QLoRA), blue for invoice reading and serving with vLLM, light blue for idle. Enlarge ↗
GPU power read from the chip (nvidia-smi, one reading per second), October 6 to 8, 2026 (idle measured overnight from October 6 to 7). The private assistant and indexing use gpt-oss 120B and bge-m3.
SituationDurationAverage powerPointeEnergy
Idle, no tasks30 min10,4 W11,0 W—
Private assistant: 100 responses, 4 in parallel, 500-page knowledge base18.0 min50,5 W68,1 W15.2 Wh
Private assistant: 100 answers, 5,000-page knowledge base18.2 min50,7 W66,8 W15.4 Wh
Indexing 5,000 pages (701 files)6.1 min53,2 W64,1 W5.4 Wh
LoRA training on invoices (Qwen3-8B, 500 steps, or 85 min of training)86.8 min53,0 W62,7 W76.7 Wh
Reading 200 invoices, twice: base model (7.7 Wh), then trained model (6.1 Wh)20.2 min41,0 W88,5 W13.8 Wh
2-hour serving: Llama 3.3 70B NVFP4, 8 concurrent users (vLLM)120.8 min39,1 W95,0 W78.8 Wh
QLoRA of Llama 3.3 70B (Unsloth, including loading)10.8 min49,3 W81,2 W8.9 Wh

Under load, average GPU power ranges from 39.1 W (the 2-hour serving run) to 53.2 W (indexing 5,000 pages). The size of the document base makes almost no difference: 50.5 W with 500 pages, 50.7 W with 5,000 pages. Peaks stay below 100 W: 95.0 W at most, during the 2-hour serving run.

A 70-billion-parameter model serving eight people draws less power (39.1 W) than indexing or training (53 W). We didn't measure the cause. One possible explanation: this serving workload spends most of its time reading memory rather than computing, which puts less load on the GPU.

#How much energy and money per task?

A task's energy is its average power multiplied by its duration. The cost is based on the EDF rate of 20.01 euro cents per kWh. The table and chart show the result for each task, GPU only.

Bar chart of GPU energy by task (orange: private assistant and indexing; purple: training; blue: invoice reading and serving with vLLM), with the cost on the right: 1,000 private assistant responses 150 Wh (€0.030), 2-hour serving for 8 users 78.8 Wh (€0.016), LoRA training 76.7 Wh (€0.015), reading 200 invoices twice 13.8 Wh (€0.003), QLoRA of Llama 3.3 70B 8.9 Wh (€0.002), indexing 5,000 pages 5.4 Wh (€0.001), a single response 0.15 Wh (€0.00003).
GPU energy in watt-hours (bars) and cost in euros at the EDF Tarif Bleu Base rate (right-hand column), for each task. Each row shows the measured quantity: 1000 responses (× 1000), 2 hours of serving, 200 invoices read twice (200 × 2), 5000 pages indexed, a single response (× 1), barely visible at this scale: 0.15 Wh. Icons: speech bubble for private assistant responses, group of people for serving, looping arrows for training, document for invoice reading, documents for indexing. Colors: orange for the private assistant and indexing, purple for training, blue for invoice reading and serving with vLLM. Enlarge ↗
GPU energy and cost at the EDF Tarif Bleu Base 6 kVA rate (20.01 euro cents per kWh including tax, rate schedule of August 1, 2026). GPU only, subscription fee not included.
TaskGPU energyCost
One private assistant response0.15 Wh0.003 euro cents
1,000 private assistant responses150 Wh (0.15 kWh)3 cents
Indexing 5,000 pages5.4 Wh0.11 euro cents
QLoRA of Llama 3.3 70B8.9 Wh0.18 euro cents
Reading 200 invoices, twice13.8 Wh0.28 euro cents
LoRA training (85 min of steps, 86.8 min of logging)76.7 Wh1.5 euro cents
2-hour serving for 8 users78.8 Wh1.6 euro cents
Idle for 24 h250 Wh (0.25 kWh)5 cents

A private assistant response uses 0.15 Wh, measured with 4 questions processed in parallel: 0.003 euro cents. A thousand responses cost 3 cents. At 1,000 responses every day for a year, the energy for the responses alone would come to about €11. A year of GPU idle time (a constant 10.4 W, or 91 kWh) costs about €18. A full 85-minute training run costs 1.5 cents, and two hours of serving eight people costs 1.6 cents.

To put these amounts in perspective, on October 8, 2026, the AI TOP ATOM was listed at €6,346.27 (model 9001, official AORUS store) and €7,999.95 (model 9000, LDLC). GB10 prices change every week, but the order of magnitude is clear: GPU electricity costs very little compared with the purchase price. Even counting the whole machine, electricity stays modest: at a constant 36.5 W, the idle power ServeTheHome measured at the wall, a month costs about €5.

#How do you estimate your own bill?

Your usage isn't ours. Here's a four-step method, with two examples calculated from our measurements. These are calculations, not measurements, and they only count the GPU.

  1. 01
    Break the month into tasks and idle time
    Write down how many hours per month the machine is working, and on which task. The rest of the time, it's idle.
  2. 02
    Take the average power of each task
    Use the table in the load section: 39.1 W for serving 8 people, 50.5 W for a private assistant, 53.0 W for training, 10.4 W at idle.
  3. 03
    Calculate the energy
    Multiply the power in W by the hours to get Wh. Add up the tasks and idle time.
  4. 04
    Convert to euros
    Divide by 1,000 to get kWh, then multiply by €0.2001 (the rate used on this page) or by your plan's price. Add the share for the rest of the machine, which only a wall-outlet measurement can give you.

Example 1, a small team. Eight people continuously use a 70-billion-parameter model for 2 hours per business day, 22 days per month. The machine sits idle the rest of the time, i.e., 720 h minus 44 h. Serving: 78.8 Wh × 22 = 1,734 Wh. Idle: 10.4 W × 676 h = 7,030 Wh. Total: 8.76 kWh, or €1.75 per month.

Example 2, light use. The private assistant answers 100 questions a day, 22 days a month, and the machine stays on all the time. Responses: 15.2 Wh × 22 = 334 Wh. Idle: 10.4 W × 713.4 h = 7,419 Wh. Total: 7.75 kWh, or €1.55 per month. In this scenario, idle accounts for more than 95% of the GPU's energy.

The lesson from both examples: for everyday use, idle time matters more than the tasks themselves. Idle power is the first number to check.

#What do the first days of the 30-day log show?

Since October 6, 2026 at 23:46, we have been logging GPU power roughly every 10 seconds, regardless of what the machine is doing. For each day, the table shows the time actually covered. Here are the first two days.

Deux cartes, pour le 7 et le 8 octobre 2026. Le 7 octobre, 20,7 h couvertes : la moitié du temps le GPU consomme 10,5 W ou moins, la moyenne est de 21,6 W, 28 % du temps au-dessus de 20 W, 447 Wh. Le 8 octobre, 11,3 h couvertes : 10,1 W, 21,4 W, 28 %, 242 Wh.
The first two days of continuous logging. First line (50%, ≤): half the time, the GPU draws at most this much power (median); the dotted line marks idle, 10.4 W. Second line (∅): average power. Third line: share of time above 20 W. Lightning bolt: GPU energy over the covered period. Clock: period covered by the log, in hours. Moon: idle. Colors: light blue for idle and the median, orange for the average and for the share of time above 20 W. Enlarge ↗
Continuous GPU measurement, one reading every 10 seconds, since October 6, 2026, at 23:46 (first few days).
JourTime period coveredMoyenneMedianGPU energyShare of time above 20 W
October 7 (intensive measurement day)20,7 h21,6 W10,5 W447 Wh28 %
October 8 (until 11:18 a.m.)11,3 h21,4 W10,1 W242 Wh28 %

La médiane vaut 10,5 W le 7 octobre et 10,1 W le 8: elle est proche de la valeur de repos (10,4 W). La moitié du temps, le GPU est donc au repos. La moyenne, 21 W environ, est plus haute à cause des tâches. Le GPU ne dépasse 20 W que pendant 28 % du temps.

Even the intensive measurement day on October 7 used only 447 Wh on the GPU side over the 20.7 hours covered: 9 cents of electricity.

Here are two projections, as ballpark figures. At a constant 10.4 W, 30 days comes to 7.5 kWh, or €1.50. At 21.5 W, the average of our first two days (measurement days, with 28% of the time above 20 W), 30 days comes to 15.5 kWh, or €3.10. These aren't measurements for the month: they'll be replaced by the full log, GPU only.

#Is the 240 W power supply enough?

The AI TOP ATOM ships with a 240 W USB-C adapter, according to GIGABYTE's spec sheet. In our per-task measurements, the highest peak is 95.0 W, and the continuous measurement also stayed below 100 W. NVIDIA lists 140 W as the thermal design power (TDP) of the GB10 chip, which combines the GPU and the CPU.

Rear of the AI TOP ATOM held in hand: PWR button, USB-C DC-IN input, three USB-C ports, an HDMI output, a 10G LAN port, and the two QSFP ports on the ConnectX-7 card.
The rear I/O: power button, USB-C power input, three USB-C ports, one HDMI output, a 10 GbE network port, and the two QSFP ports on the ConnectX-7 card. Photo of our unit, with the background neutralized. Enlarge ↗
Scale from 0 to 240 W with six markers: GPU idle 10.4 W, 2-hour serving for 8 users 39.1 W, indexing 5,000 pages 53.2 W, maximum peak in our per-task measurements 95.0 W, GB10 chip thermal design power 140 W, and included adapter 240 W. The green zone runs from 0 to 95.0 W.
The power levels from our logs plotted on a scale from 0 to 240 W, the adapter's rating (distances are proportional): averages for idle, serving, and indexing, and the peak for 95.0 W. Icons: moon for idle, group of people for serving 8 users, documents for indexing, lightning bolt for the maximum peak, chip for thermal design power, plug for the adapter. Green zone: from 0 to the maximum peak in our per-task logs. Hatched zone: up to the chip's thermal design power (140 W, NVIDIA). Dark zone: the rest of the 240 W adapter's capacity. Enlarge ↗

So there's still headroom. The adapter's 240 W covers the chip and the rest of the machine: memory, SSD, networking, and fans.

#What do others measure at the wall?

The measurements published by ServeTheHome and Peter Falkingham cover the whole machine at the wall. Their methodology differs from our GPU readings, so they give an order of magnitude, not a direct comparison.

Third-party measurements at the wall (power meter). Sources: ServeTheHome, March 5, 2026; Peter Falkingham, September 18, 2026.
SourceSituationMeasured power
ServeTheHome (AI TOP ATOM)Idle, network connected36 à 37 W
Peter Falkingham (AI TOP ATOM)Everyday use40 à 60 W
ServeTheHome (AI TOP ATOM)Sustained load102 to 158 W, peaks just under 200 W

ServeTheHome also notes two fans, each rated at 3 W. ServeTheHome records peaks just under 200 W. All of these values stay within the power supply's 240 W.

#What these measurements don’t tell you

The GPU alone
We don't measure the CPU, memory, SSD, network, or power supply losses.
Scope of the measurements
Sensor readings describe the GPU alone. Total power draw at the wall falls under a separate protocol.
Only the first few days
The 30-day measurement is in progress. The projections on this page are calculations, not measurements for the month.
Typical tasks
Our tasks don't cover every use case. Your hours of use and your models will change the total.
A single machine
We measured a single AI TOP ATOM. We give no figures for other GB10 machines.
The rate
The EDF rate is the one from August 2026. Your plan, such as Heures Creuses (off-peak), may give a different result.

Updates and corrections: this page will be revised at the end of the 30-day measurement, or if a new version of DGX OS changes these results; each correction will be dated.

#Who ATOM is the right choice for

What follows comes from our GPU measurements, along with their limitations.

The ATOM is the right choice if you want a local AI that's cheap to power
At idle, the GPU draws 10.4 W. A thousand private assistant responses cost 3 euro cents in GPU energy.
… if you leave it on all day
Our first two days were measurement days (21.5 W on average): 30 days at that rate comes to 15.5 kWh, or €3.10. At idle alone, 30 days comes to €1.50.
… if you share it within a team
Eight users for 2 hours: 78.8 Wh, or 1.6 euro cents.
… if you train your own models
One 85-minute LoRA training run: 76.7 Wh, or 1.5 euro cents.
… if you want a simple power setup
A 240 W adapter, and a GPU whose highest peak is 95.0 W in our per-task measurements.

If you need to justify an energy budget down to the euro, add a plug-in power meter: it will give you the whole machine's total, which our GPU readings don't.

#FAQ

How much electricity does the AI TOP ATOM's GPU use at idle?+
10.4 W on average over 30 minutes (from 10.2 to 11.0 W), as reported by the chip. That comes to 0.25 kWh per day, or 5 cents. The whole machine draws more: ServeTheHome measured 36 to 37 W at the wall, at idle with the network connected.
How much do 1,000 responses from a private assistant cost?+
3 cents of GPU energy, at the EDF rate of 20.01 cents per kWh. One answer uses 0.15 Wh. We measured it with gpt-oss 120B, 4 questions at a time, and a knowledge base of 500 or 5,000 pages: 50.5 W and 50.7 W average power.
How much does a month of use cost?+
We don't know yet: the 30-day monitoring run is still in progress. As a ballpark, 30 days at idle comes to €1.50 of GPU power, and 30 days at the rate of our first two days (21.5 W) comes to €3.10. This page will be updated when the monitoring run ends.
How is GPU power measured?+
The readings use NVIDIA telemetry: the power reported by the chip covers the GPU only. For the machine's total consumption, the wall-outlet measurements from ServeTheHome and Peter Falkingham are cited separately.
Is the 240 W power supply enough?+
Yes, based on our readings: across our tasks, the GPU's highest peak is 95.0 W, and NVIDIA lists a 140 W TDP for the entire chip. At the wall, ServeTheHome measured 102 to 158 W under sustained load, with peaks just under 200 W: everything stays within the adapter's 240 W.
How much power does the whole machine draw, not just the GPU?+
Wall-outlet measurements published by third parties complement the GPU readings: ServeTheHome measured 36 to 37 W at the wall at idle and 102 to 158 W under sustained load, with peaks just under 200 W; Peter Falkingham measured 40 to 60 W in everyday use. At a constant 36.5 W, a month comes to 26 kWh, or about €5.
In the same series
Did this guide help you?

Feedback, an error, or a clarification? Let us know—it improves the guide for everyone.

Prices in euros (€) are French market prices including VAT, as checked by BestLLMfor. US prices differ: the Amazon buttons show the current US price.