Example chat provider

Your workload as the model sees it

Page 2 of 7

Your yield and how it moves

2 distinct runs between 2026-03-27 and 2026-05-17; framework tags vllm:v0.18.0, vllm:v0.21.0. The latest published run is older than sixty days, so yields may have moved since.

Expected progress is the mean monthly log change of each yield across the calibration window (8 transitions); shocks are the residuals, resampled jointly for prefill and decode at the snapshot level. Any reservation or hedge sized on today's yield is sized for a quantity that expected progress alone will change by the figures above every month.

Page 3 of 7

Where your cost risk comes from

• GPU price, 37 percent of six-month cost variance: futures or reservations address it.
• Provider basis, 6 percent: nothing standard addresses it; only a contract that settles on your provider's own rate would.
• Your demand, 44 percent: reservations address part of it; forecasting is the lever.
• Serving yield, 15 percent: nothing addresses it; progress-aware sizing avoids compounding it.
• Utilization, 0 percent: your autoscaling policy is the lever.

Diagnostic: the five Shapley shares at six months sum to 103 percent (the exact value is one hundred; the difference is Monte Carlo estimation error of the closed Sobol indices). Shares are simulated under the base scenario with 20,000 paths.

Page 4 of 7

Procurement policies compared

All policy figures are simulated under the base scenario. P(overrun) is the share of simulated paths whose cumulative six-month cost exceeds 1.25 times plan; ES95 is the mean overrun of plan in the worst five percent of paths. Reserved rows marked assumed use a reserved rate based on an assumed 15 percent discount from your on-demand rate; the reserve-for-peak rows still overflow on demand when simulated demand exceeds the reserved capacity, so neither is free of on-demand exposure.

This comparison assumes a reserved rate of 85 percent of spot (assumed, 2.61 USD per GPU-hour), a demand shape whose always-on trough is 49 percent of the mean, expected efficiency progress of -2 percent per month in the prefill yield that dominates this mix, and a flat index. Change any of these and the ranking can change; page 7 shows how far.

Under these inputs, reserve baseline plus progress-aware strip has the lowest expected shortfall, at +41 percent of plan in the worst five percent of simulated six-month paths.

Page 5 of 7

Proxy-index hedge simulation

HE is one minus the variance of hedged cumulative cost divided by the variance of unhedged on-demand cost; a negative value means the strip added variance. At twelve months the minimum-variance strip is 118 percent of the fixed-conversion strip and the progress-aware strip is 106 percent of it. Each contract count already reflects your basis level of 1.000 index units per GPU-hour. Settlement correlation sensitivity, minimum-variance HE of six-month cost: correlation 1.0 gives 38 percent; correlation 0.9 gives 31 percent; correlation 0.7 gives 18 percent.

One-index theoretical bound

Under the assumptions below the largest share of cost variance any linear hedge on the index can remove is 37 percent at six months (37 at three months, 35 at twelve), from a coefficient of variation of 25 percent for the index and 31 percent for your quantity exposure. simulated

• positive index level S and positive quantity exposure X
• S and X independent
• a hedge linear in S
• settlement period matching the exposure month
• no correlation between basis and the index
• frictionless sizing (no margin, fees or liquidity limits)
• procurement and settlement indices coincide

Applies when the procurement and settlement indices coincide. Proxy-index mismatch can only reduce achievable hedge effectiveness relative to this idealised case.

Range at six months

Minimum-variance HE of six-month cost across the sensitivity grid: 18 / 38 / 64 percent (minimum, median, maximum). The grid is one-at-a-time changes to index volatility (25, 35, 60 percent), demand forecast error (half, base, one and a half times), basis volatility (0, 10, 20 percent), efficiency shock scale (0.5, 1, 1.5), adoption lag (0, 1, 3 months) and settlement correlation (1.0, 0.9, 0.7), each at N = 12000 paths.

Page 6 of 7

Sensitivity, provenance, scope

Ornn's Compute Price Index is a volume-weighted, winsorized mean of executed on-demand rental transactions over a rolling one-hour window, from verified providers. It is not the CME settlement index, which is published by Silicon Data.

Data files used

Scenario: index volatility 35 percent a year, index drift 0 percent a year, basis volatility 10 percent, risk premium 0 percent, settlement correlation 1.0, demand forecast error 10 percent a month, growth 3 percent a month, 20,000 paths, seed 20260908. Intake hash e84bdffdedd5.

Page 7 of 7