Will It Local? / Methodology

Last verified 2026-07-05

Every formula, every assumption

This is the audit page. It contains the four formulas the site runs on, the thresholds that turn numbers into verdicts, where every price comes from, and the things the model deliberately leaves out. The whole site is only worth what this page can defend.

The four formulas

These are copied verbatim from the header of engine.mjs — the one file that computes everything on this site. There is no fifth formula hiding anywhere.

1. Hardware amortization

Hardware_amortized/mo = build_cost / (amort_years * 12)

Straight-line, nothing clever. A $1,200 rig over the default 3 years costs $33.33 a month whether you use it or not. You can set the horizon anywhere from 2 to 5 years; shorter horizons are honest if you expect to want new hardware sooner, and they make ownership look worse, not better. We default to 3 because that is roughly how long a GPU stays comfortable in this hobby.

2. Electricity, with honest idle accounting

Electricity/mo = [(active_kW * active_hrs) + (idle_kW * idle_hrs)] * 30 * $/kWh

Two terms, and the second one is where most calculators cheat. Term one is inference: whole-system draw under load, using a realistic ~70–80% of GPU TDP plus the rest of the platform — not the sticker peak, not a flattering minimum. Term two is idle: the hours the box sits powered on, drawing 50-odd watts while you're at lunch. That is real money and we count it. A used-3090 rig at 450 W for 3 hours plus 50 W for 5 hours burns 48 kWh a month — $9.04 at the national average rate of $0.1883/kWh (EIA, verified 2026-07-05). Your state can double that. Field note № 1 measures this at the wall.

3. Cloud cost

Cloud/mo = (in_Mtok * in_price + out_Mtok * out_price)

Millions of tokens in and out per month, times the provider's posted per-million prices. Trivial arithmetic — the entire game is in which baseline you pick. Compare against a premium frontier model and local looks great; compare against an ultra-cheap open-weight API and it often never pays off. That choice moves the verdict more than any hardware decision you will make, which is why Field note № 2 exists.

4. Break-even

Break-even_months = build_cost / (Cloud/mo - Electricity/mo)

If Cloud/mo <= Electricity/mo → no break-even, ever.

Each month of ownership "earns" the cloud bill you skipped, minus the power you burned. Divide the build cost by that and you get the payoff month. Note what is not in the denominator: amortization. The build cost is the debt being paid down — it lives in the numerator. Counting it again as a monthly cost would double-charge ownership and quietly flatter the cloud. And if the cloud bill is at or below your electricity bill, there is no break-even at any horizon, and the engine says so in plain words instead of showing you a 400-month fantasy.

How numbers become verdicts

The thresholds, from engine.mjs: OWN at 18 months or less, HYBRID from 18 to 36, RENT beyond 36 — and RENT regardless whenever break-even lands past your own amortization horizon, or never arrives at all.

Why those numbers. Eighteen months is half the default 3-year life of the rig: pay off inside that and you spend most of the hardware's useful life in the black, with enough margin to survive a cloud price cut or two. Thirty-six months is the whole default life: a rig that pays for itself on its last day of relevance didn't really pay for itself. And the horizon rule catches the edge case — if you chose a 2-year horizon and break-even lands at month 30, that's a RENT, because the rig is due for replacement before the math closes. The engine also attaches a confidence level (break-even under 12 months is high; near a threshold is low) so a 33-month HYBRID doesn't masquerade as certainty.

The sovereignty score

Four axes, each 0–10: privacy (prompts never leave the machine — 10 for any local rig), offline (works on a plane, in an outage — 10), rate-limit freedom (no 429s, no deprecations — 10), and setup ease, the honest variable: a Mac scores 9, a dual-3090 build scores 3 and costs you a weekend and some forum posts.

The score is never mixed into the dollar math, and that is a design decision, not an oversight. The moment you assign privacy a dollar value, you can tune that value until the verdict says whatever sells hardware. Keeping the axes separate keeps the money math falsifiable: you can check every dollar on this page against a bill. Buying a rig for sovereignty is a completely legitimate decision. Calling it savings is not, and we won't.

Where the data comes from

Four JSON files, served openly under /data/, refreshed quarterly. Each carries a lastVerified stamp that the interface shows you — so when our data is stale, we're stale in public.

Prices are indicative, not quotes. Every preset in the app links out to a live price so a stale table can never quietly cost you money.

A worked example, run through the real engine

One rig, three baselines, three different honest answers. The setup: a used RTX 3090 build ($1,200 all-in, 450 W active, 50 W idle), the coding-assistant preset (7.5M tokens in + 7.5M out per month, 3 hours active, 5 idle), the national electricity rate, 3-year amortization. Local cost is flat regardless of baseline: $33.33 amortization + $9.04 electricity = $42.37/mo. These rows are the engine's actual output, not estimates of its output.

Same rig, same watts, same rate — the baseline does all the deciding. Field note № 3 takes this rig apart in detail, and the break-even index has it pre-computed for every state: for instance vs Sonnet in California or vs DeepSeek in Texas.

What the model leaves out

Plainly, so nobody has to discover it in the comments:

• No resale value. A used 3090 will fetch something in three years. We count it at zero, which understates the case for owning. We would rather err in that direction.
• No repairs. Fans die, PSUs pop, used cards were sometimes mined on. There is no failure line item.
• No time-cost of setup. Your assembly weekend costs $0 in the math. It shows up in the sovereignty score's setup axis instead — scored, never priced.
• Token counts are estimates. The usage presets are tuned to real patterns, but your month is not our preset. If you know your numbers from a real bill, enter them.
• Prices drift. Everything here was verified 2026-07-05 and refreshes quarterly. Cloud pricing in particular tends to fall, which usually makes our OWN verdicts more conservative over time, not less.

Audit the engine yourself

The entire model is one readable JavaScript file — pure functions, no dependencies, served unminified at /js/engine.mjs. The exact same file runs in your browser, in our tests, and behind every pre-computed page on this site. Read it, run it, argue with it. If you find an error, tell us; correcting it in public is the business model.

Run your own numbers →

Every dollar figure on this page is the output of an engine run against the published data tables — nothing was hand-rounded into a better story.