How the price is made
Every number on the pricer comes from a live index anchor per GPU, volatility estimated from that GPU's realized series, a basis adjustment for the provider you'd actually rent from, a Bachelier option value, and a disclosed spread. Every quote here is indicative, subject to the provider you are matched with and their basis to the index.
1. The instrument
A capacity call is a physically-settled booking option. You pay a non-refundable fee at trade for the right, without any obligation, to N GPUs at a locked hourly rate (the strike) for a defined future window. The writer is a named provider that hard-holds the capacity and keeps the fee in every outcome. Exercise notice is due seven days before the window opens. If you exercise, 80% of the fee is credited against the GPUs you consume.
Note that the strike is set on the writer's own rate card rather than the index. Physical delivery at a named venue means settlement basis is zero by construction. The index drives pricing, where basis belongs.
2. Live, transaction-based indices per SKU
Pricing anchors to the Ornn OCPI daily indices, the public, transaction-based rental benchmarks published per GPU. Each SKU carries its own live anchor and its own realized series; volatility is estimated per SKU from that series, so a B200 quote is priced off B200 price rather than an H100. The forward curve off each anchor carries a −3% drift across the horizon, consistent with the persistent contango in rental rates as new supply lands.
3. Volatility is measured from the data
Daily returns of each SKU's realized series, annualized over 252 trading days. The live regime uses an exponentially-weighted moving average (λ = 0.94, the RiskMetrics standard):
σ²t = λ·σ²t−1 + (1 − λ)·r²t
GPU rental rates regime-switch, with weeks of drift punctuated by supply squeezes, and a booking priced today should reflect today's regime. The calm and squeeze bounds on the vol slider are the best and worst rolling 30 days of the same series, so the slider's range is measured from that history.
4. Basis
Everyone rents at some venue's card, and cards track the index imperfectly. The model decomposes a venue's realized rate into a systematic part and an idiosyncratic part:
venue = anchor × (1 + β · index return) + idiosyncratic
σ²venue = β²·σ²index + σ²idio
A venue that reprices with the market (β ≈ 0.9, low idio) is nearly the index; a deep discounter with a flat card tracks it loosely (β ≈ 0.7, higher idio). Looser tracking means more uncertainty about the card you'll actually face, which prices straight into the fee.
One caveat on the quote itself: two clusters of the same GPU are not the same product. Interconnect (NVLink versus PCIe), memory per card, storage and networking, and region all move the real number. The price on the page is a basic estimate off the index. A binding quote reflects the actual spec, the provider's own rate card, and their basis.
5. Bachelier
With window forward F, strike K, and dollar standard deviation s = F·σ·√T (T = lead + half the window, in years), the fair value of the call per GPU-hour is the normal-model price:
fair = (F − K)·Φ(d) + s·φ(d), d = (F − K) / s
6. From fair value to your fee
fee = max( fair × 1.10, 0.75% × F )
The floor (0.75% of the window forward) keeps small, far-from-strike quotes from rounding to zero. Higher vol, longer lead, tighter strike, or a loosely-tracking card all move the quote, so the fee is a priced percentage that shifts with the inputs.
7. Efficacy
Why not just sign a reservation? If your dates are certain and your horizon is a year, a reservation is cheaper per hour, so sign one. The call is for the other case: a real date that might move, a 4-week need that doesn't justify a 12-month commitment. You're paying a few percent for the right to be wrong.
Why would a provider write this? Fee income on capacity that would otherwise sit idle, plus a pre-qualified buyer at the end of it. Providers already sell term discounts, prepaid credits, and take-or-pay. This is the same machinery at smaller ticket sizes, priced by a model rather than a negotiation.
Is the quote on this page binding? No. It is the model's indicative price at live inputs. A quote comes back within 48 hours, struck on a named provider's actual rate card.
For inquiries, contact@plainviewinstruments.com.
