The honest local-AI break-even engine
Will it local?
Your rig, your usage, and the cloud model you'd actually otherwise pay for, run through one auditable formula. The verdict is OWN, RENT, or HYBRID, with the exact month owning wins. And yes, we'll happily tell you not to buy the GPU.
We get paid the same whether the verdict says buy or rent — full honesty policy →
The test chamber
Four questions. One verdict.
No account, no email, nothing leaves your browser until you choose to share it.
Pick your rig
Street prices, realistic sustained wattage, not marketing peak numbers. Or go custom.
loading the hardware table …
Community reality check
Pick your workload
Monthly tokens and how many hours a day the machine actually runs. Honest idle hours included: idle watts are real watts.
loading the usage table …
Pick the cloud model you'd actually use
This choice moves the verdict more than any hardware decision. Be honest: compare against what you'd really pay for, not the most expensive thing on the menu.
loading the pricing table …
Electricity & horizon
Your rate decides whether idle hours are a rounding error or a second subscription.
Decoder ring: the jargon, one line each
VRAM: the GPU memory that decides which models fit at all. Quant / Q4: quantization, a smaller, faster copy of a model (Q4 = 4-bit) with a slight quality cost. M tokens: a million tokens, roughly 750,000 words in or out. Idle watts: what the machine burns while sitting ready and doing nothing, real watts, so we bill them.
Go deeper Compare, stress-test, and dig into the numbers
The bottom line
$0
Effective cost per million tokens
The cloud baseline is the whole game
Same rig, a second cloud
| Same rig vs… | Cloud/mo | Break-even | Verdict |
|---|
The other lever
Same cloud & usage, a second rig
| Same cloud vs… | Build | Local/mo | Break-even | Verdict |
|---|
Run the ratio backwards
How much use would make this worth owning?
Stress-test it
Drag any lever. The tipping point under each is the exact value where the verdict would flip — computed from the same formula, not eyeballed off the chart.
Sovereignty: /10 — what money can't say, never mixed into the dollar math above.
Some links are affiliate/referral links. The verdict above was computed before these were rendered, and that ordering is the whole point.
The data flywheel
Community-measured, not vendor-claimed
Real tokens-per-second and real wall-plug watts from real rigs. Every row is a person who measured instead of guessed. Submissions are human-reviewed before publishing.
| Rig | Model | Quant | tok/s | Watts | Samples |
|---|---|---|---|---|---|
| No published measurements yet. Be specimen № 1. | |||||
Your plan
For when the verdict is OWN (or stubbornly HYBRID)
The Build Plan
Your verdict, turned into a shopping list — generated from the exact scenario you just ran. Delivered instantly, printable, yours forever.
- Parts list at street prices
- Quantization settings that fit your VRAM
- Expected tokens-per-second
- The power-limit tweak that pays for itself
- Your break-even math to argue with
Show your work
Four formulas. Zero mystery.
The entire model is public and auditable: amortization, honest idle-watt accounting, and the break-even division. Data tables are refreshed quarterly and stamped: hardware & cloud prices last verified 2026-07-05. If a number looks stale, it links to the live source so you can catch us.
The real electricity cost of a 70B coding assistant in 2026
Watts at the wall, idle hours included, and why your state changes the answer more than your GPU.
Read → Field note № 2Local vs cloud: when it actually saves money (and when the internet lies to you)
The cloud baseline nobody puts in their calculator, and the 2026 price collapse that changed the honest answer.
Read → Field note № 3Does a used RTX 3090 still pay for itself?
The people's GPU, re-audited for 2026: three cloud baselines, three completely different verdicts.
Read →