Slide deck
BackForty-Three Years Was Last Month’s Price
A calculator says a $3,499 Mac bought to run open models locally pays for itself in 43.6 years. Every figure reproduces exactly from its own published data. But only two inputs move it — the API price and how much you use it — and the API price is now 39 days out of date.
Explainer · prices read 12 October 2026
Does local AI pay for itself?
How the Sunk Cost calculator works, the ideas behind it, and what its number really depends on.
For beginners and practitioners1 / 16
The question
Buy a machine, or rent the AI by the word?
Buy: run it at home
Pay once for the hardware, then only for electricity. Like buying a car.
Rent: use an online API
Pay a small fee for every piece of text the model reads and writes. Like taking a taxi.
The one question
How long until the purchase earns back its price?
2 / 16
The tool
Sunk Cost is a free, open-source pay-back calculator
Built by one developer to test a popular claim: buy a Mac, run models locally, and it pays for itself. Its README says what it is for: “It is willing to say ‘this doesn’t pay off’. That is the product.”
- 39current open models, of 55 in its data
- 56machines: Macs, graphics cards and more
- 15 Sepposted to Hacker News, 2026
The data, the formula and the gaps are all published3 / 16
The headline result
43.6 years to pay back a $3,499 Mac
A Mac Studio with the M5 Max and 64 GB, running Qwen3.8 27B, at 500,000 tokens a day. The saving is only about $0.22 a day.
- 43.6years to break even, on the published defaults
- 13 / 13published figures that recompute exactly
- 3 Septhe day its API prices were checked
Correct arithmetic, on prices from a stated day4 / 16
Concept 1 · Tokens
AI reads and writes in tokens, and you pay per token
What is a token?
A small piece of text, often part of a word. Models count, and are priced, in them.
Input and output
Input is what you send: the question, files, chat history. Output is what the model writes back.
Price per million
As the tool recorded it on 3 September: Qwen3.8 27B at $0.32 per million input and $2.50 per million output. Writing costs more than reading.
5 / 16
The method
The whole calculation is three steps
- 1Daily API billinput tokens × input price + output tokens × output price
- 2Daily electricity at homehours spent writing output × machine watts × electricity price. Nothing else on the local side.
- 3Break-even daysmachine price ÷ (API bill − electricity). If the saving is zero or less: never.
The formula is published in the project’s README6 / 16
Worked example
Walking the default to 43.6 years
Step Calculation Per day API input 468,750 tokens × $0.32 per million $0.150 API output 31,250 tokens × $2.50 per million $0.078 Home electricity 0.35 h × 145 W × $0.17/kWh $0.009 Saving 0.228 − 0.009 $0.22 Break-even $3,499 ÷ 0.2196 = 15,936 days 43.6 yr Electricity barely counts
It is 3.8% of the API bill. The 145 W is the project’s own stand-in, because Apple publishes no power figure for the 2026 machines; make the electricity free and the answer is still 42 years.
7 / 16
Concept 2 · Memory bandwidth
Home speed is limited by how fast memory can be read
To write each word, the machine re-reads the whole book. A wider pipe, or a thinner book, means faster writing.
- 614 GB/smemory bandwidth of the M5 Max
- 18.6 GBread per token: weights plus KV cache
- 24.8 tok/s614 ÷ 18.6 × 0.75 efficiency: an estimate
- 0 / 98measured speed rows that cover any M5 machine
8 / 16
Concept 3 · Context window and KV cache
Long chats need more memory, and run slower
Context window
How much text the model keeps in view at once, such as 32k tokens. Its short-term memory.
KV cache
Notes the model keeps on every token it has read, so it need not work them out again. They grow with every token: 2.15 GB at 32k for this model.
The fit test
Weights plus KV cache must fit in usable memory. A 64 GB Mac offers 48 GB to the GPU, and 27 of the 39 current models fit at 32k.
9 / 16
Concept 4 · Prefill and decode
Reading a prompt is fast; writing is slow
Prefill: reading
The whole prompt is processed at once, in parallel, limited by raw compute.
Decode: writing
One token at a time, re-reading memory for each, limited by bandwidth. Tens of tokens a second at home.
The simplification
The tool times decode only and treats input as free. Its README says the real daily ceiling is plausibly 20–30% lower on input-heavy work.
10 / 16
Concept 5 · Prompt caching
Re-sent text is often billed at a steep discount
When the start of a prompt matches one the provider saw recently, it reuses its stored work — a cache hit — and charges less for those tokens.
Provider Cached input, against the normal price OpenAI 0.25–0.50× Gemini, Moonshot AI 0.25× DeepSeek; Alibaba Qwen, when asked for 0.1× Not modelled
Sunk Cost charges all input at the full price. For agents that mostly re-read their context, renting is cheaper than shown, so pay-back is longer still.
Cache prices as OpenRouter lists them, read 15 October 202611 / 16
What changes the answer
Usage moves the date, not the finish line
- 436 yr50,000 tokens a day: light personal chat
- 43.6 yr500,000 tokens a day: the default
- 13 months20 million tokens a day: agents around the clock
The finish line
The machine is worth about 8 billion tokens of saved API spend. Use it faster and you reach that line sooner. Falling API prices push the line further away.
Default Mac Studio with Qwen3.8 27B, prices of 3 September12 / 16
The price under the number
“The cloud price” is not one number
Twenty hosts serve this model. Same machine, same model, same day, priced on 12 October:
API price This workload Break-even The tool’s recorded figure, 3 Sep $6.94/mo 43.6 yr Cheapest live host, 12 Oct $1.85/mo 182.9 yr Median of 20 live hosts $5.24/mo 58.5 yr Dearest live host, 12 Oct $15.54/mo 19.1 yr Old, not false
By the tool’s own rule — the cheapest active host — the answer 39 days later is closer to 180 years than 44. The switch that lets API prices keep falling is off by default, so 43.6 is the kindest answer the defaults give.
13 / 16
Reception
The press repeated the number; the community challenged it
Media
The 43.6 years travelled, reaching Taiwan’s TechBang on 11 October, nearly four weeks after launch.
Hacker News
The speed estimate is “at least 2x” too low; it pays back fast at 24/7 use; privacy matters more than money.
The author’s changes
A field for your own measured speed and a best-buys page, both on launch day; on 18 September the usage slider was raised to 100 million tokens a day.
14 / 16
Limitations
What the calculator leaves out, and which way it pushes
Left out Included, home hardware would look… Resale value, other uses, privacy better Speculative decoding, MTP, 8-bit KV cache faster, but the saving barely moves Many agents batched on one machine better, at very high volume Prompt-caching discounts on the API worse Prefill time, idle power, depreciation worse 15 / 16
Takeaways
Buy for privacy or heavy agents, not for savings
- Use your own two numbers. Real tokens a day from your provider’s dashboard, and the price of the host you would actually use. Together they move the answer tenfold or more.
- Enter a measured speed, from llama.cpp or vLLM, rather than trusting the estimate — it decides your waiting time, if not your bill.
- Discount for caching. Agentic work with many cache hits makes renting cheaper still.
- Check model quality. The tool compares open models; a frontier subscription is a different trade-off.
The calculator: sunkcost.ai16 / 16