Cloud vs Local AI: Work Out Your Own Break-Even in 15 Minutes
By the end of this post you will have your own break-even number: the month in which a machine you own becomes cheaper than paying per token. You need nothing more than a spreadsheet or Python, and the prices and speeds below. Cheapest place to start: the laptop you already have.
What you need
- Your rough daily AI workload: how many requests, and how long each prompt and answer are. A guess within 2x is fine for a first pass.
- Ten minutes with the price tables below.
- Optional: Python 3 to run the calculator. Any AI assistant can also do the arithmetic for you if you paste the tables in.
Step 1: Write down what the cloud charges today
Prices are per million tokens, in USD, as published in September 2026. A token is about three quarters of an English word, and closer to half a word in Spanish.
| Model | Input / 1M tokens | Output / 1M tokens | Source |
|---|---|---|---|
| GPT-5 | $1.25 | $10.00 | OpenAI pricing |
| GPT-5 mini | $0.25 | $2.00 | OpenAI pricing |
| Claude Sonnet 5 | $2.00 | $10.00 | Anthropic pricing |
| Claude Haiku 4.5 | $1.00 | $5.00 | Anthropic pricing |
Two things to notice. Output tokens cost 4 to 10 times more than input tokens, so workloads that write a lot (reports, code, long answers) are the expensive ones. And the small models are cheap: a mini or Haiku class model at a few dollars a month is a real option, and this post will not pretend otherwise.
Step 2: Write down what local hardware costs and delivers
| Hardware | Price (EUR) | Runs comfortably | Speed we measured or sourced | Power |
|---|---|---|---|---|
| The laptop or PC you already own (16 GB RAM, CPU only) | 0 | 3B models, 7B slowly | 5 to 10 tokens/s (CPU) | already paid |
| NVIDIA Jetson Orin Nano Super | ~250 | 3B models | ~12 to 18 tokens/s (Edge AI Vision) | 25 W |
| Mac mini M4, 24 GB | ~920 | 7B to 14B models | ~35 tokens/s on Qwen 2.5 7B (Compute Market) | 30 W |
| Our own workstation (NVIDIA GB10, 128 GB unified memory) | far more than you need | 35B+ models | 33.2 tokens/s on Qwen 2.5 Coder 7B, measured 2026-09-09 | shared with 50 other services |
The last row is there for honesty, not as a recommendation. The numbers that matter to a small business are the first three rows.
Electricity for a Mac mini running all day, every day: 30 W × 24 h × 30 days = 21.6 kWh a month. At EUR 0.15 per kWh that is about EUR 3.25 a month. A Jetson is under EUR 3. That is the entire running cost once the box is on your desk.
Step 3: Turn your workload into tokens
Pick the row that looks like your business, or build your own from the same arithmetic.
Workload A: document processing. 200 documents a day, each sent with 3,000 tokens of context and returning a 500-token summary. Per month (22 working days): 13.2 million input tokens, 2.2 million output tokens.
Workload B: an internal assistant. 20 people, 50 questions a day each, 4,000 tokens of context per question, 300-token answers. Per month: 88 million input tokens, 6.6 million output tokens.
Step 4: Run the calculator
# break_even.py — months until owned hardware beats per-token billing
PRICES = { # USD per 1M tokens (input, output), Sept 2026
"GPT-5": (1.25, 10.0),
"GPT-5 mini": (0.25, 2.0),
"Claude Sonnet 5":(2.00, 10.0),
"Claude Haiku 4.5":(1.00, 5.0),
}
HARDWARE_EUR = 920 # Mac mini M4 24 GB; use 250 for a Jetson, 0 for your own PC
POWER_EUR_MONTH = 3.25 # 30 W, 24/7, EUR 0.15/kWh
IN_M, OUT_M = 88, 6.6 # workload B in millions of tokens per month
for name, (pin, pout) in PRICES.items():
cloud = IN_M * pin + OUT_M * pout # we treat 1 USD as 1 EUR: favours the cloud
months = HARDWARE_EUR / (cloud - POWER_EUR_MONTH) if cloud > POWER_EUR_MONTH else float("inf")
print(f"{name:18s} cloud {cloud:7.2f}/month break-even {months:5.1f} months")Run it with python3 break_even.py. Change the three numbers at the top to yours. If you prefer, paste the two tables into any AI assistant and ask it: “Calculate my monthly cloud cost and the break-even month for a EUR 920 machine at EUR 3.25 a month to run.”
What the numbers say
| Workload | Model | Cloud per month | Mac mini break-even | Jetson break-even |
|---|---|---|---|---|
| A: 200 docs/day | GPT-5 | $38.50 | 26 months | 7 months |
| A: 200 docs/day | GPT-5 mini | $7.70 | never worth it on cost alone | 56 months |
| A: 200 docs/day | Claude Haiku 4.5 | $24.20 | 44 months | 12 months |
| B: 20-person assistant | GPT-5 | $176.00 | 5.3 months | 1.4 months |
| B: 20-person assistant | Claude Sonnet 5 | $242.00 | 3.9 months | 1.0 month |
| B: 20-person assistant | GPT-5 mini | $35.00 | 29 months | 8 months |
| B: 20-person assistant | Claude Haiku 4.5 | $121.00 | 7.8 months | 2.1 months |
Read it honestly:
- At low volume, the cheapest cloud models win on cost. If you send a few hundred short requests a day and a mini-class model does the job, EUR 8 a month is hard to beat. Choose local for other reasons: the data never leaves the building, it works when the internet does not, and the bill can never surprise you.
- The moment several people use it daily, owned hardware pays for itself inside a year against every frontier model, and inside two months against the flagship tiers.
- The board the 3B model runs on costs less than one month of a busy team’s frontier bill. That is the zero-barrier entry point: start on a EUR 250 device or the PC you have, measure your real token counts for two weeks, then decide whether a bigger machine is justified.
What local cannot do, so you plan for it
A 7B model on a Mac mini will not match a frontier model on hard reasoning, long multi-step agents, or obscure knowledge. The practical pattern is hybrid: route routine, repetitive, private work to the local machine, and keep a metered cloud key for the 5 percent of tasks that need the biggest model. Your break-even calculation then applies only to the routine share, which is usually where the volume is anyway.
Next steps
- Size the box properly: Edge AI hardware guide 2026.
- Fit a bigger model into less memory: Quantization explained.
- Turn the break-even into a full business case: Local AI ROI framework and calculator.
- Install everything step by step: How to deploy AI locally in your business.
Work with us
We run this exact calculation with clients before recommending any hardware, using their real token counts rather than estimates. If you want a second pair of eyes on your numbers, get in touch or see how our consulting works.