One Rupee Buys You 1.8 Hours of AI: What Running a Model at Home Actually Costs in Electricity
Measured GPU power draw across four local models on a 4 GB laptop. Running one costs about 3 rupees per million tokens — and the model that sips the least power is the most expensive to run.
On a laptop with a GTX 1650 Ti, generating a million tokens costs between 3.1 and 5.0 Nepali rupees in GPU electricity, depending on the model. At the US average electricity price of 19.7 cents per kWh the same generation costs 5 to 8 US cents. One rupee buys about 1.8 hours of continuous generation.
Get new posts in your inbox
No spam, unsubscribe anytime.
About the Author
Sujit KarkiFinance Researcher & Market Analyst
Independent finance researcher and market analyst with expertise in macroeconomics, equity markets, and personal finance. I help regular investors make better-informed decisions through rigorous, data-driven analysis.
Generating a million tokens on a 4 GB laptop GPU costs 3.1 to 5.0 Nepali rupees in electricity, or 5 to 8 US cents at the US average electricity price.
One rupee buys roughly 1.8 hours of continuous generation.
The lowest-wattage model was the most expensive per token: it drew 22% less power but ran 58% slower.
At US electricity rates that is 98x to 489x cheaper than cloud API pricing — but the comparison is electricity against a retail price, and the local models are far weaker.
Idle GPU draw is 3.51 W; under load it rises to roughly 45 W.
Writing about running AI at home usually compares graphics cards, or quotes a
GPU's rated wattage off a spec sheet. Neither answers the question a person
actually has, which is: if I use this thing, what happens to my electricity bill?
So I measured it. A background sampler polled the GPU's power draw four times a
second while four models generated text, on the same laptop used for the
numeracy benchmark and the tax
test on this site — an Acer
Nitro 5 with a GTX 1650 Ti.
The number
At the Nepal Electricity Authority domestic rate of 12 rupees per kilowatt-hour:
Measured GPU energy and cost, five generations per model
Model
Tokens/sec
Mean draw
Wh per 1k tokens
NPR per 1M tokens
qwen2.5:3b
52.2
45.5 W
0.26
3.12
gemma2:2b
52.4
44.7 W
0.26
3.17
llama3.2:3b
46.9
45.9 W
0.29
3.45
phi3.5:3.8b
22.2
35.7 W
0.42
5.01
A million tokens is a great deal of text — roughly seven or eight full-length
novels. Producing it costs between three and five rupees.
Put another way: one rupee buys about 1.8 hours of continuous generation. The
electricity is not the constraint. It is not close to being the constraint.
Cost per million tokens, cheapest model0NPR, GPU electricity only, at 12 NPR/kWh
The same numbers in US cents
Rupees are the wrong currency for most readers of this site, so here it is in
dollars. Energy per token does not depend on where you live — only the price
of that energy does, and the US price is published: the Bureau of Labor
Statistics tracks the average price of electricity per kilowatt-hour as series
APU000072610. As of July 2026 it stands at 19.7 cents.
Same measurement, US city average electricity price (19.7 c/kWh, BLS, July 2026)
Model
kWh per 1M tokens
USD per 1M tokens
qwen2.5:3b
0.260
$0.051
gemma2:2b
0.264
$0.052
llama3.2:3b
0.287
$0.057
phi3.5:3.8b
0.417
$0.082
Five cents per million tokens. American electricity is about two and a half
times the Nepali domestic rate — 19.7 cents against roughly 7.9 cents equivalent
— so the same computation costs more in Ohio than in Kathmandu. It is still five
cents. Nothing about that changes the conclusion.
It is worth noting which direction that cuts, though. The economics of running a
model at home are better in a country with cheap electricity, and Nepal has
some of the cheapest domestic power in the region. If local inference ever
becomes genuinely competitive on capability, the places that benefit most are the
ones where a kilowatt-hour is cheap and a cloud subscription is expensive in
local wages.
The model that sips power costs the most
Look at the last row of that table again, because it inverts the intuition the
whole "wattage" framing encourages.
phi3.5:3.8b draws 35.7 watts under load — about 22% less than the other
three, which sit near 45. On a power meter it is the gentle one. Yet it is the
most expensive model per token, by a margin of 61%.
The reason is arithmetic. Energy per token is power divided by throughput,
and phi3.5 generates at 22.2 tokens per second against roughly 52 for the
others. Drawing 22% less power while taking 2.3 times as long is a losing trade.
It holds the GPU for longer, and duration beats wattage.
Pro Tip
This is why a wattage figure on its own tells you nothing about running cost. A
model that idles the card at low power for a long time can cost more than one
that runs it hard and finishes. The number to compare is energy per unit of work,
not watts.
There is a sharper edge to this. In the tax
benchmark, phi3.5:3.8b was the
only model of the four to score above zero — 51% correct where the others got
nothing right. So the one model worth using is also the one that costs the most
to run. On this hardware, accuracy is not free; it is about 60% more expensive
per token. It is still five rupees per million.
Against the cloud, honestly
Taking the US electricity price above, the cheapest model costs $0.051 per
million tokens to run:
Local electricity cost against published API pricing, at US electricity rates
Option
Per 1M output tokens
Ratio
qwen2.5:3b on this laptop
$0.051
baseline
Claude Haiku 4.5
$5.00
98x
Claude Sonnet 5
$10.00
196x
Claude Opus 5
$25.00
489x
At the Nepali rate the same ratios are roughly two and a half times larger again —
243x, 486x and 1,216x — because the electricity is cheaper while the API price is
the same everywhere. That is the quiet part of global API pricing: a dollar is a
dollar in Kathmandu and in California, but a kilowatt-hour is not.
Those ratios are real, and they are also the most misleading numbers in this
post if left without qualification. Three things they do not account for:
It is electricity against a retail price. The API figure covers datacentre
hardware, model training, serving capacity, bandwidth, support and profit. The
local figure covers none of those, because you already paid for the laptop. A
fair comparison would amortise the machine — and at these volumes the hardware
dominates completely.
The models are not comparable. A 3B model quantised to fit 4 GB is not
Claude Opus 5, and the two previous experiments on this site measured exactly
how badly these models fail at ordinary money arithmetic. Three of the four
scored zero on a tax calculation. Cheap output you cannot trust is not a bargain.
It is GPU-only. CPU, RAM, display and power-supply losses are real and I have
no wall-plug meter to capture them, so the true figure is somewhat higher. It is
a floor, not a total.
What this does and does not tell you
The defensible claim is narrow: once you own the hardware, the marginal cost of
running a model on it rounds to nothing. Three rupees per million tokens is not
a budget line. It is not worth thinking about.
That matters because it removes running cost from the decision entirely. If you
are choosing between a local model and an API, the real questions are capability,
privacy, and whether you already own a machine that can run it. Electricity is
not one of the questions, and any argument that leans on it — in either direction
— is arguing about a number too small to matter.
The corollary is less comfortable. If electricity is free and the hardware is
already bought, the only remaining cost of a wrong answer is the wrong answer
itself. On this hardware, that turns out to be the expensive
part.
Methodology
The raw power measurements are published with this post as
energy.json. The measurement script is
research/measure_energy.py; the run
reproduces on any machine with an NVIDIA GPU and Ollama.
Hardware. Acer Nitro 5, NVIDIA GTX 1650 Ti (4 GB VRAM), 8 GB system RAM,
Ollama 0.33.2. All four models ran fully on the GPU.
Sampling.nvidia-smi reports instantaneous power draw. A background thread
polls it every 250 ms for the duration of each generation, so each figure is a
mean over that specific request rather than a nameplate rating. Five generations
per model, each producing 608 to 768 tokens.
Marginal, not total. The laptop draws power whether or not it is inferring.
The idle baseline — 3.51 W, measured before the run — is subtracted, so the
headline figures are the additional energy inference costs. Total draw is
recorded alongside in the dataset and is roughly 8% higher.
Rates. Electricity at 12 NPR/kWh. Currency converted at the Nepal Rastra Bank
published rate for 7 September 2026 (USD sell 151.48). API prices are Anthropic's
published per-million-token output rates.
Limits. GPU only, as described above. One machine, one GPU, four small
models — throughput and therefore cost per token will differ on other hardware.
The script accepts any tariff via --rate, so the calculation reproduces at your
own electricity price rather than mine.
Disclosure. The models are the subject being measured, not the author. The
commentary was drafted with AI assistance; the experiment design, the measurement
method and the conclusions are mine.
Sujit Karki||7 min read