Local inference · Cost Structure · Open Models · Apple Silicon · Cost ComparisonTool of September 2026, prices read 12 October 2026
What local AI actually costs
Forty-Three Years Was Last Month’s Price
A one-page calculator asks how long a machine bought to run open models locally takes to pay for itself against renting the same model by the token. On its default settings — a $3,499 Mac Studio, a 27-billion-parameter model, half a million tokens a day — it answers 43.6 years, and the answer travelled. The number is sound: it publishes its data and its formula, and every one of the thirteen figures in the coverage recomputes exactly. What the number depends on is the interesting part. Make the machine’s electricity free and it moves by under four per cent. Change which of twenty live hosts you rent from and the same machine, the same model and the same day give anything from 19 years to 183.
What Happened
The arithmetic checks out; the date on it does not What the tool says, what reproduces from its own files, and what has changed since those files were written.
- 43.6Years to break even, on the tool’s published default settings
- 13 / 13Published figures that recompute exactly from the project’s own data files
- 0 / 98Measured speed rows in its data that cover the machine in that default
- 19 – 183Years to break even across the 20 hosts serving that model today
- 3.8%How much the answer moves if the machine’s electricity were free
| Standing | What | Where it is established |
|---|---|---|
| Confirmed | Every figure in the coverage — the monthly bill, the electricity, the saving, the memory, the speed, the waiting time, even “27 of 39 models fit” — recomputes exactly from the project’s own data and its own published formula. | Four data files and the formula, all published in the open. Recomputed here on 12 October 2026. |
| Confirmed | The prices behind that answer were checked on 3 September 2026. Queried live today under the tool’s own rule — the cheapest active host — the same workload costs about a quarter as much, which pushes break-even to roughly 180 years. | The data file records the rule and the date; the marketplace’s own API was queried on 12 October 2026. |
| Confirmed | Twenty hosts serve that model right now. Input prices span 25-fold and output prices 3.5-fold, so break-even on one machine, one model and one day runs from 19 years to 183. | All twenty endpoints and their prices, from the marketplace’s own API, 12 October 2026. |
| Confirmed | The machine in that default has no measured speed anywhere in the project’s 98 measurements — it had not shipped when the data was compiled — and the project says so itself. | The measurements file, read row by row, and the project’s own list of known gaps. |
| Confirmed, but not verifiable here | The author’s own summary: at a million tokens a day the quickest payback in that capability class is 8.3 years, and it drops under a year only at around 20 million tokens a day. | Written in a public forum thread on 15 September 2026, against prices checked twelve days earlier. |
| Unconfirmed | Whether any of today’s cheapest prices are introductory. A 25-fold spread on one model invites the question, and nothing in the marketplace data answers it. | Twenty endpoints are listed as active; nothing distinguishes a standing price from a promotional one. |
What the number is
A correct calculation, on stated assumptions, from data checked on a stated day, by a project that publishes all three and lists its own gaps.
What it is not
A constant. Two of its inputs move it by a factor of ten or more and both have moved since; the two that look decisive — the machine’s speed and its power — move it by less than five per cent between them.
Timeline
Thirty-nine days from the price check to the price The data was compiled before the machine it describes went on sale, and the figure reached a new audience four weeks later.
- 12 Mar 2025A research group publishes the trend the whole question turns on: the price of a given level of capability has fallen steeply and repeatedly, by as much as 40-fold a year on one benchmark.
- 3 Sep 2026The project records its prices and defaults as checked. The model in the default is priced at $0.32 in and $2.50 out per million tokens, by the rule “cheapest active endpoint”.
- 15 Sep 2026The calculator is posted to a public forum. It draws 47 points and 99 comments, and the first of them refuses the premise: privacy, the commenter writes, makes it pay off instantly.
- 22 Sep 2026The machine in the default setting ships — nineteen days after the data describing it was compiled. Its speed in the calculation is an estimate from memory bandwidth, and is labelled as one.
- 11 Oct 2026The figure is reported to a Chinese-language audience, twenty-six days after it was published and thirty-eight days after the prices behind it were read.
- 12 Oct 2026The same model’s live prices are read across twenty active hosts. By the tool’s own rule the workload now costs about a quarter of the recorded figure, and the twenty prices do not agree with each other either.
The Argument
Four inputs, and only two of them matter Each row below moves one input and leaves the rest at the published defaults.
The formula is published, and reading it is most of the work. The cloud side is charged for input and output tokens. The local side is charged for the electricity burned while generating output — no idle draw, no time spent reading your prompt, no depreciation, no share of the machine’s other uses. Whatever else that is, it is not a calculation rigged against buying the machine: on running costs it is as generous to the local box as arithmetic allows, and the hardware price is left doing all of the work.
Which is exactly why the local side of the equation barely moves the answer. At the published defaults the electricity is 3.8% of the cloud bill, so the two numbers that look like the local machine’s whole story — the speed and the power draw — have almost nowhere to go. Halve the speed and you add fourteen months to forty-three years. Quadruple it and you take fifteen months off. Make the electricity free and the answer is still 42 years.
That reframes the project’s most-quoted caveat. Its data holds 98 measured speeds and none for the machine in its own default, which had not shipped when the file was written; the project says so in its list of known gaps. It is a real gap and it decides the time cost completely — the forty seconds a thousand tokens, the quarter-hour of extra waiting a day. It decides the money almost not at all.
The two inputs that do decide it are the API price and the volume. On the volume: the default is half a million tokens a day, and the answer is a fact about that number. Three years’ payback needs 7.3 million tokens a day and one year needs 21.8 million, both inside the machine’s own published ceiling of 34.6 million. The author says the same thing in his own words — under a year “if you’re running agents at around 20M tokens/day”. On the price: see the next section, because one number cannot carry it.
- 3.8%The electricity, as a share of the cloud bill it is meant to beat
- 42.0Years to break even if the machine ran on free electricity
- 7.3MTokens a day needed for a three-year payback, against a default of 0.5M
| Move one input | Break-even | Against 43.6 years |
|---|---|---|
| Speed estimate halved, 25 to 12.5 tok/s | 45.4 | +4.1% |
| Speed estimate quadrupled, 25 to 100 tok/s | 42.4 | −2.8% |
| Power draw doubled, 145 to 290 W | 45.4 | +4.1% |
| Electricity at the official current figure, 18.31¢ | 43.8 | +0.3% |
| Electricity free, 0¢/kWh | 42.0 | −3.8% |
| API price: cheapest live host today | 182.9 | +319% |
| API price: dearest live host today | 19.1 | −56% |
| Usage at 7.3M tokens a day | 3.0 | −93% |
| Usage at 20M tokens a day | 1.1 | −97% |
What Others Add
“The cloud price” is not a number Twenty hosts, one model, one day, and the break-even each of them implies for the same machine.
| Price band, per million tokens | Hosts | This workload | Break-even |
|---|---|---|---|
| Cheapest: $0.040 in, $1.350 out | 2 | $1.85/mo | 182.9 yr |
| Under $4 a month | 9 | $1.85–3.92 | 79.6–182.9 yr |
| $4 to $10 a month | 9 | $5.09–9.46 | 31.7–60.3 yr |
| Over $10 a month | 2 | $14.46–15.54 | 19.1–20.5 yr |
| The figure in the calculator, checked 3 Sep | — | $6.94/mo | 43.6 yr |
The switch nobody mentioned
43.6 years is the kindest answer available
- The tool carries an option to let API prices keep falling. It is switched off by default, so the headline holds today’s price flat for forty-three years.
- Its own note: prices “have not behaved that way”, and switching the option on “pushes break-even out”.
- Past a point, the note says, “the API costs less than your electricity” and break-even “stops being reachable at all”.
What it leaves out
And says so, item by item
- The whole purchase is charged to inference. Nothing is credited for video work, compilation, or anything else the machine does, and nothing is amortised against resale.
- The power figures for the 2026 machines are stand-ins, because the maker has not published any; every affected row is flagged.
- The daily ceiling treats reading your prompt as costing no time. Its own note: the real ceiling is “plausibly 20-30% lower”.
What it cannot price
The first reply refused the premise
- “It pays off instantly, because [the vendors] can no longer see what I’m doing… If I am offloading some of my thought processes to a machine, I want to own that machine.”
- A second reply makes the methodological point in the tool’s favour: the fair comparison is a host running the same open model, which is what it does.
- Offline stability, no change of terms, no change of price and no content rules are in the same category: real, and not expressible in dollars a month.
| Also worth knowing | Detail |
|---|---|
| A dated figure is not a wrong one | The calculation records the rule it used and the day it used it. A price that moved afterwards makes the figure old, not false — and this is a field where a figure can be old in five weeks. |
| Cheapest is a rule, not a host | The data file records the rule and the date, never which provider it landed on, so the September figure and the October one cannot be traced to the same host. |
| The price spread is a question, not a recommendation | Nothing here tested latency, throughput, data retention or reliability at any of the twenty hosts. A 25-fold spread on one model is a reason to look, not a reason to pick the bottom of it. |
| What could not be read | The maker’s own purchase page confirms the prices of the memory tiers either side of the one in the default, but draws the configuration prices with script, so the $3,499 rests on the project’s data file and the source it cites. |
Conclusion
Which way of buying inference suits which buyer The calculation answers one question well. It is not the only question, and it is not the same question for everybody.
Rent it
If the bill is the reason
- Best at: costing nothing up front, never being the wrong machine, and getting cheaper while you sit still.
- At this usage the bill is a few dollars a month, and the hardware would need decades of them.
- The trade-off: your work goes to somebody else’s machine, and the price, the terms and the model can change without you.
Buy it
If the volume is already there
- Best at: turning a fixed cost into unlimited use. The machine does not charge by the token, so the heavier the day the better it looks.
- Three years’ payback at 7.3 million tokens a day, about one year at 20 million — both within the machine’s own ceiling, and both a long way above the default.
- The trade-off: it is slower per token than a data centre, and the hosted price you are beating is falling underneath you.
Buy it anyway
If the reason was never the bill
- Best at: the things no calculator has a column for. The work never leaves the room, the terms cannot change, and it runs with the network down.
- Where confidentiality is a legal requirement rather than a preference, the break-even question does not arise at all.
- The trade-off: be honest that this is what you bought. The arithmetic above is not an argument against it, and it is not an argument for it either.
Settled
The arithmetic is right and reproducible from published data. At half a million tokens a day the machine does not pay for itself on the bill, by a wide margin, and the margin is not sensitive to anything about the machine.
Open
What the hosted price will be. It fell by roughly four times in thirty-nine days on one reading, twenty hosts disagree with each other by 25-fold today, and the trend the tool cites is from an analysis its own note says predates this year.
Hold this loosely
Any single break-even figure, including every one on this page. They are all snapshots of a price that moves weekly, computed at a usage level somebody chose. Re-run it with your own two numbers; the rest barely matters.
- Work out your own two numbers before anything else: tokens a day, and the price of the host you would actually use. Together they move the answer by a factor of ten or more. The machine’s speed and its electricity move it by under four per cent between them.
- Check the date on any break-even figure, including this one. The prices behind the headline were five weeks old when it reached a new audience, and the workload they describe had changed by about four times in the meantime.
- Price the host you would use, not “the cloud”. Twenty of them serve this one model today and their input prices differ by 25-fold. Which one you pick changes the answer more than anything you could buy.
- If you bought the machine for privacy, say so and stop doing the sum. The calculation charges the whole purchase to inference and credits nothing for anything else the machine does. It answers a question about money; it was never the question you were asking.
Quick notes
- the report of the calculator in TechBang, 11 Oct 2026
- the calculator itself at sunkcost.ai, and its repository — the README carrying the formula and the list of known gaps, and the four data files for hardware, models, defaults and measured throughput
- the Show HN thread of 15 Sep 2026, for the author’s own figures and the first replies
- OpenRouter’s models and endpoints APIs, queried 12 Oct 2026 for all twenty live prices
- Apple’s Mac Studio specifications and purchase pages
- Epoch AI, ‘LLM inference prices have fallen rapidly but unequally across tasks’, 12 Mar 2025
- and the US Energy Information Administration’s Electric Power Monthly, via its API. All read 12 Oct 2026. The slides also draw on OpenRouter’s prompt-caching guide and the repository’s commit history, both read 15 Oct 2026.
Reliable
Worth believing
Sharing a subject
- Local LLMs and Buying a Mac — A Technical ReportApple Silicon · Local inference
- Inkling: Thinking Machines' Open Model, and the Rival It Learned FromOpen Models
- Nvidia's Confirmed $12.9bn Acquisition of Hugging FaceOpen Models
- Nine Years, Same MarginCost Structure
Sharing a source domain
- The Claude Code Quality Postmortem — Three Unrelated Changesgithub.com · ycombinator.com
- Local LLMs and Buying a Mac — A Technical Reportapple.com · github.com
- Avalanche's RWA Boom, and Why AVAX Still Fallsgithub.com · githubusercontent.com
- Swap Across Five Platforms — Compress First, Spill Laterapple.com · githubusercontent.com
- Enough, or Less? Context Engineering Said Both in One Weekgithub.com · ycombinator.com
- 9 Billion Requests a Day, Moved Without Rebuilding a Filegithub.com
- Grok Build Shipped Whole Repos to xAI's Cloudgithub.com
- Five Names for One Stack, and No Agreed Ordergithub.com
- POCO X8 Pro vs X8 Pro Max — A Taiwan Buyer's Guidetechbang.com
- It Began as Tuning an Agent. By March It Meant Everything but the Model.github.com
- One Term, Coined Twice, Two Weeks Apart. The Joke Won.github.com
- Dorado v4 to v6 — What Changed, How Much, and Whether It Was Goodgithub.com
- Nvidia's Confirmed $12.9bn Acquisition of Hugging Faceopenrouter.ai
- Z.ai's Chinese-Chip Claim for GLM-5.3-Flashopenrouter.ai
- Shunting Bulk Reads to a Cheaper Model — What the 90% Actually Cutsgithub.com
- The Front Stopped Moving. The War Did Not.eia.gov
- Decades Rather Than Yearsapple.com