Hedging inference costs

If you rent GPUs by the hour, your bill rises when the market rate rises. A LONG plot on the same crop pays when that happens, so it offsets part of the extra cost.

A long 7-day plot pays out when the rent you are exposed to rises

Your exposure

Count the GPU-hours you will buy at market rates in the coming week, per model. A team running 8 H100s around the clock buys 8 × 168 = 1,344 H100-hours a week. A 10% rise in the rate costs them about 10% of that bill.

Sizing a hedge

A parimutuel payout depends on how the pot is split, not on the size of the move, so a plot is a bet on direction. The payout on a winning LONG plot is plot × (1 + shortPot / longPot). Size the plot so that payout covers what a rise would cost you, and check the split on the crop page before you plant.

Worked example

H100 SXM index at Monday 00:00 UTC: 2.61 $/h. Weekly bill at that rate: 1,344 × 2.61 = 3,508 USD.
The pot shows 600 USDG LONG and 900 USDG SHORT. You plant 100 USDG LONG.
The index ends the week at 2.74 $/h. Your plot pays 100 × (1 + 900 / 700) × 0.99 = 226.29 USDG, against a rise of 175 USD on the bill.
Had the index fallen, you would lose the 100 USDG and pay less for the GPUs.

Automate it

An agent can run the same logic every Monday with the SDK: read the index, read the pot, plant over x402, and let the keeper harvest.