Reference

Appendix C — Derived Quantities#

Three of this report's sourcing tags describe where a number came from: primary, reported, unverified. None of them describes a number that came from nowhere — that the report computed from figures it cites elsewhere. Those are a fourth class, and until this appendix they carried no marking at all. This part lists them, shows the arithmetic, and states the assumptions each one rests on.

The reason to do this is the argument in Section 3. A cost-shift claim that cannot be tested on the record is not a claim a commission can adjudicate; that is the report's finding about other people's numbers, and it applies to its own. A derived figure printed without its inputs asks the reader for trust that the report elsewhere declines to extend. Every entry below is written so that a reader who disagrees with an assumption can substitute their own and see what changes.

C.1 Two hour figures for the same curtailment rate (Section 6)#

As printed: Section 6 and the References give roughly 22, 44 and 88 hours a year in one place and 85, 177 and 366 hours in another, for the same three curtailment rates.
Why it needs an entry: both are correct and they measure different things, but nothing on the page says so, and a reader comparing them has no way to reconcile a fourfold difference.

The 22, 44 and 88 figures are full-curtailment-equivalent hours: the rate is a share of energy, so 0.5% of the 8,760 hours in a year is 43.8, the time it would take to curtail that energy if the load were shed completely. The 85, 177 and 366 figures are hours in which any curtailment occurs, which is the operationally meaningful number because curtailment in the study is partial. The two reconcile at an average depth of 43.8 ÷ 177 = 24.7%, and that is consistent with the retention statistics Section 6 already prints: at least half the new load retained in 88% of curtailment hours, three-quarters in 60%, nine-tenths in 29%.

The distinction matters to the thing Section 6 is arguing about. A contract written against 44 hours of full interruption is a different product from one written against 177 hours of shallow, mostly-partial reduction, and the second is both easier to accept and harder to price. Section 6 states the energy-versus-time distinction explicitly; what it does not do is connect the two hour counts, which is why they read as a discrepancy.

C.2 Rack-level energy storage against the emission band (Annex T.1)#

As printed: roughly 4.7 kJ per rack, on the order of 115 milliseconds of coverage against the swing, and less in service.
Inputs: 65 joules per GPU (vendor, primary); 72 GPUs per NVL72 rack; rack rating near 135 kW; a swing of about 30% of rack power, from the vendor's own peak-reduction figure.

Stored energy is 72 × 65 = 4,680 J. The denominator is the swing, not the rack: the buffer is not asked to run the machine, only to supply the difference between what the rack draws and what the grid delivers. Against a 40 kW swing that is 4,680 ÷ 40,000 = 117 ms. Two corrections both cut it. A capacitor bank discharged only to 70% of rated voltage yields 1 − 0.7² = 51% of stored energy, or 2,390 J. And a periodic swing is not rectangular: the energy in a half-cycle of a sinusoid is the swing multiplied by the period and divided by π, which is 64% of the swing multiplied by half the period that a rectangular estimate gives. Working it that way, the usable buffer covers 56% of a half-cycle at 3 Hz, 19% at 1 Hz and 4% at 0.2 Hz.

What the entry rests on. The 70% discharge floor is an engineering assumption, not a published specification, and the conclusion is sensitive to it: at a 50% floor the usable energy is 3,510 J and coverage at 3 Hz rises to 83%. The sinusoidal treatment is also an idealisation — a real training waveform is closer to a square wave, which moves the requirement back toward the rectangular figure. What survives both sensitivities is the ordering: the buffer is sized for the transient, the ramp controls handle the cycle, and neither alone spans the band. That ordering, not any single millisecond figure, is what Section 4 relies on.

C.3 Ratios and products stated in the text#

Four figures are simple arithmetic on cited inputs and are recorded here so that a reader checking them finds the basis rather than a discrepancy. ERCOT's queue against its all-time peak: 438 ÷ 85.5 = 5.12×, printed as roughly five times. Its preliminary 2032 forecast against the same peak: 367,790 ÷ 85,500 = 4.30×, printed as more than four times. Data-center-attributable capacity cost: 21.3 ÷ 47.2 = 45.1%, printed as about 45%. And the Texas financial-security posting for a 500 MW campus: 500 × $50,000 = $25 million, which is the gross posting — the roughly 20% refundable at commercial operation is not netted out of the printed figure.

Closed in v1.39. The forecast fan formerly in T.7 carried no stated base, and its 110 GW spread implied a base near 219 GW at a ten-year horizon — a figure matching nothing the report uses, against a US summer peak of about 760 GW. Rather than infer an assumption, the bullet was restated on the NERC and ERCOT revisions the report already sources, so there is no derived quantity left to record.