OpenWeightsTerminal

Which box pays back

Legacy edition · a research rewrite is awaiting approval. Historical prices and product statements are not current guarantees.

A DGX Spark bought to resell tokens at $1 per million, at 20% duty, does not pay back. A box you already own only owes the electricity.

Street price comes back only if the tokens, at $1 per million, beat the watts. A Spark bought to resell that rate does not.

Update, 29 Sep: price and demand change the answer

The table below assumed $1 per million for every model and 20% of the day busy. Hosts now set their own price, and uncensored builds start at $4 per million because the big providers do not serve them. Served every second of the month at $4, an M4 Pro at 85.5 tok/s grosses about $899, and after the host's 80% share and the power bill it pays back in about 3 months. A DGX Spark at 79.5 tok/s grosses about $836 and pays back in about 8. The limit is no longer the price. It is enough paying callers to keep a box busy.

The clock

The payback table on /payback uses the US street price, residential power at 18.34¢/kWh, $1 per million tokens, and 20% of the day serving. Months are the street price divided by what is left after the power bill. Watts follow the work.

Fast

The top row is Qwen3.5 2B on an RTX 5090: 351 tok/s, $4,639, 575 W, 2.3 years. That print is /pair/qwen35-2b-q4km-5090. Next is an M4 Pro at $2,199 and 45 W. Qwen3.6-35B-A3B at 85.5 tok/s is 4.2 years, /pair/qwen36-35b-4bit-mlx-m4pro. Qwen3.5-4B at 82.8 tok/s is 4.3 years, /pair/qwen35-4b-4bit-mlx-m4pro. The mini is fast because the watts are almost nothing. The 5090 is not a cheap card. It is fast because 351 tok/s still clears 575 W at $1 per million.

Slow

One DGX Spark is $4,999 on the street table. Qwen3.8-Flash-Next at 79.5 tok/s and 240 W is 12 years, /pair/qwen38-flash-exl3-spark. Two Sparks at 94.4 tok/s and 480 W are 23 years, /pair/qwen36-35b-abliterated-nvfp4-spark. Two Sparks at 45 tok/s are 77 years, /pair/qwen38-flash-fp8-spark. An RTX 4090 at $4,495, 40.7 tok/s, 450 W is 40 years, /pair/qwen38-27b-ud-q4k-xl-4090. The power hour eats a weak token hour.

Cheap, or already yours

The cheap US price on this desk is the RTX 5060 Ti 16GB at $614. Its best quoted decode is 23 tok/s, 180 W, 7.0 years, /pair/qwen38-27b-gsq-rco-5060ti. The 10 tok/s uncensored print on the same card is 117 years, /pair/qwen38-27b-turbofc-uncen-5060ti. A cheap card pays back only when decode is high relative to watts. A box you already own has nothing left to pay back but the electricity. Buying a Spark to resell $1 per million tokens at 20% duty is the long case. The full rank is /payback.

Method & limitations

Legacy edition, predating the current research standard. A proposed rewrite is awaiting editorial review. This is not an original experimental report; historical prices and product statements should not be treated as current guarantees.

Sources

See the inline references in this legacy edition. A structured source review has not yet been completed.