The figure attached to this story is not in the record. Neither the CFTC's release nor the reporting around it puts a number on GPU residual value, or on the share of a chip's book value that a public rental price would support [18]. What the material does support is narrower and worth stating plainly: a settlement price for on-demand rental of two Nvidia chip generations, assembled by Silicon Data, with each CME contract standing for one month of rental [5][6].
That unit is why the collateral conclusion does not follow automatically. Unused GPU time cannot be held and sold later [7], and a working paper using data supplied by Silicon Data and Ornn found that longer-term rental agreements give only an upper bound on value rather than a fair price [c8b]. A curve built this way tells a lender what an hour of capacity is expected to earn. Turning that into a percentage of hardware value requires assumptions about utilisation, useful life and discount rate, and none of them sit inside the index.
The calendar is tighter than the number. The comment window closes no earlier than October 18, roughly two weeks after CME's target launch date [16], which means the public argument over how the benchmark is built runs alongside a listed contract rather than ahead of it. CME holds the only published launch date among the three US efforts [14]. ICE's version, which normalises capacity by energy use through the NATIVX COIL Index, is expected later in 2026 [11], and Architect Financial Markets is building a venue on the CFTC-designated contract market it acquired with IMX Health, with exchange-for-physical links to Compute Desk [12].