Margin and ask price

Sum the cost-floor lines, apply a 15–35% margin, convert $/compute-hour to the ask price field ($/1,000 tokens).

Cost floor (2× A100, mid power, 24/7)

Margin

Public A100 asks sit around $1.10–$1.99 / GPU-hour. This box is two GPUs, so a $1.80 / box-hour ask is $0.90 / GPU-hour — below those lists if you sell the pair as one instrument. If you price per GPU, double the box floor first.

Convert to price per 1,000 tokens

Sustained 160 tokens/s = 576,000 tokens / hour.

$/1,000 tokens  =  ($/compute-hour) ÷ (tokens_per_hour / 1000)
                =  1.805 ÷ 576
                =  $0.00313

Submit that as a decimal string on POST /v1/sellers/{id}/asks: "0.00313". The field is per 1,000 tokens, strictly positive.

If decode is 80 tokens/s, tokens/hour halves and price doubles. Measure your own tokens/s; do not copy 160.

Spare-capacity variant

Drop hardware from the floor (spare capacity): $0.115 + $0.226 = $0.341 / h. At 25% margin: $0.426 / h → $0.00074 per 1,000 tokens at 160 tok/s. That is the incremental ask. Going lower than $0.341 / h loses money on every fill.

Worked end-to-end numbers for three operators: scenarios.