Carmen Li(@carmenli):1/ GPU reserve contracts and GPU futures can price the exact same future months differently. That disagreement isn’t just academic. It can reveal an arbitrage opportunity—or simply tell a compute buyer which way to hedge. 2/ Start with the physical market. If a 6-month H100 reserve costs $2.52/hour and a 12-month reserve costs $2.30/hour, what are months 7–12 actually costing? The answer is not $2.30. 3/ The math: 12 × $2.30 = $27.60 6 × $2.52 = $15.12 The remaining six months cost $12.48. Divide by six, and the reserve-implied forward price for months 7–12 is $2.08/hour. 4/ Now compare that with GPU futures covering the same six months. If the futures strip averages $2.25/hour, the financial market is pricing the same capacity $0.17 above the physical reserve-implied rate. That gap matters. 5/ A $0.17/hour spread sounds small. But over six months: • One GPU: approximately $745 • 100 GPUs: approximately $74,460 • 1,000 GPUs: approximately $744,600 Small pricing disagreements become meaningful at infrastructure scale. 6/ In theory, arbitrage should close the gap. Buy cheaper reserved capacity. Sell the more expensive futures. When futures settle against the spot index, the index exposure cancels, leaving the spread. But compute is not a frictionless commodity. 7/ A GPU contract represents specific hardware, in a specific location, from a specific provider. That creates basis risk, operational risk, utilization risk and—in the harder direction—fulfillment risk. This is why a persistent spread may not be “free money.” 8/ The participants best positioned to close these gaps are compute providers and large buyers. They already control physical capacity, understand utilization and can manage both sides of the trade. Financial traders alone may not be able to enforce convergence. 9/ For most AI companies, the lesson is simpler: You don’t need to execute the arbitrage. You need to read it. Compare: Reserve-implied forward rate vs. Futures strip + your expected provider basis Then choose the cheaper—or operationally safer—route. 10/ One more important point: a forward price is not simply a prediction of where spot prices will be. It also reflects scarcity, financing, access, flexibility, balance-sheet costs and market frictions. The curve is a price—not a prophecy. 11/ As GPU futures begin trading, we’ll finally be able to observe how the physical and financial compute markets interact. Where do the gaps emerge? Who closes them? How much is guaranteed access worth? Those answers will help define compute as an asset class. Full analysis:

2026.09.15 12:29

1/ GPU reserve contracts and GPU futures can price the exact same future months differently. That disagreement isn’t just academic. It can reveal an arbitrage opportunity—or simply tell a compute buyer which way to hedge. 2/ Start with the physical market. If a 6-month H100 reserve costs $2.52/hour and a 12-month reserve costs $2.30/hour, what are months 7–12 actually costing? The answer is not $2.30. 3/ The math: 12 × $2.30 = $27.60 6 × $2.52 = $15.12 The remaining six months cost $12.48. Divide by six, and the reserve-implied forward price for months 7–12 is $2.08/hour. 4/ Now compare that with GPU futures covering the same six months. If the futures strip averages $2.25/hour, the financial market is pricing the same capacity $0.17 above the physical reserve-implied rate. That gap matters. 5/ A $0.17/hour spread sounds small. But over six months: • One GPU: approximately $745 • 100 GPUs: approximately $74,460 • 1,000 GPUs: approximately $744,600 Small pricing disagreements become meaningful at infrastructure scale. 6/ In theory, arbitrage should close the gap. Buy cheaper reserved capacity. Sell the more expensive futures. When futures settle against the spot index, the index exposure cancels, leaving the spread. But compute is not a frictionless commodity. 7/ A GPU contract represents specific hardware, in a specific location, from a specific provider. That creates basis risk, operational risk, utilization risk and—in the harder direction—fulfillment risk. This is why a persistent spread may not be “free money.” 8/ The participants best positioned to close these gaps are compute providers and large buyers. They already control physical capacity, understand utilization and can manage both sides of the trade. Financial traders alone may not be able to enforce convergence. 9/ For most AI companies, the lesson is simpler: You don’t need to execute the arbitrage. You need to read it. Compare: Reserve-implied forward rate vs. Futures strip + your expected provider basis Then choose the cheaper—or operationally safer—route. 10/ One more important point: a forward price is not simply a prediction of where spot prices will be. It also reflects scarcity, financing, access, flexibility, balance-sheet costs and market frictions. The curve is a price—not a prophecy. 11/ As GPU futures begin trading, we’ll finally be able to observe how the physical and financial compute markets interact. Where do the gaps emerge? Who closes them? How much is guaranteed access worth? Those answers will help define compute as an asset class. Full analysis: