A $3,499 Mac Studio saves 22 cents a day against hosted inference in Sunk Cost's model

Two inputs produce the 44 years. Sunk Cost prices the featured workload at about $6.94 a month through an OpenRouter endpoint and 26 cents a month in electricity ^[6]. The gross saving is $6.68 a month, roughly $80 a year, and $3,499 divided by $80 is 43.6 years ^[1][2].

Most of that hosted bill is prompt. At 468,750 input tokens a day and 32 cents per million, input costs about $4.56 a month; the 31,250 daily output tokens cost about $2.38, so two thirds of the bill is input ^[5][7][4]. Sunk Cost does not price prompt caching, and prompt caching works on the input side ^[16].

The two assumptions most likely to be wrong move the answer very little. Generation speed is an estimate of 24.7 tokens per second, derived from memory bandwidth, an assumed bytes-per-token figure and a 75% efficiency factor, on a configuration nobody has measured in production ^[10][11]. Power is 145 watts under load, taken from Apple's published maximum for the previous M4 Max Mac Studio ^[12]. Delete the electricity line entirely and break-even goes from 44 years to 42 ^[5].

Working backwards from the published inputs, 614 GB/s at 75% efficiency divided by 24.7 tokens per second implies about 18.6 GB read per token ^[9][6]. In a dense model the weights stream from memory once per token, so the rate is just a bandwidth quotient. Sunk Cost labels it an estimate ^[10].

The speed estimate matters more on the waiting side. Sunk Cost puts a 1,000-token response at 40 seconds on the Mac and 13 seconds through an API running at 80 tokens per second, and totals the extra delay at about 15 minutes a day ^[15]. That checks out: 31,250 output tokens is about 31 responses of 1,000 tokens, 27 seconds slower each, or 14 minutes ^[7]. Over a year the operator waits about 91 hours to save $80, which is 88 cents an hour ^[8].

For payback inside three years at these prices, the workload has to carry about $97 a month of hosted spend. That is roughly 15 times the modeled one: 7.3 million tokens a day, and about five hours a day of local generation at 25 tokens per second ^[9]. It also assumes the September 3rd rates hold ^[7]. Sunk Cost includes an annual API decline input, and RuntimeWire's reason for it is that hosted vendors spread new hardware and utilization gains across many customers while the Mac owner captures none of those improvements unless faster software or a better local model arrives ^[14][18].

Sunk Cost excludes the value of keeping code and documents on-device, working offline, riding out provider outages, using the Mac for unrelated work, and resale ^[16]. Those are the arguments a buyer has left. The calculator's own numbers give the API 80 tokens per second against the Mac's 25, so latency is not one of them for this configuration ^[15][10].