Finding the Optimal Combination of Pricing Levers

Gross margin on a GPU cloud is not set by a rate card. It is the output of four interacting commercial levers applied to different slices of the same physical fleet: how much capacity is sold as certainty, how much is multiplexed into tokens, how the perishable residual is priced and hedged, and how tightly the discount waterfall is governed. Optimizing any one lever in isolation moves revenue between slices rather than raising fleet margin.

This paper sets the unit cost floors from the cost stack, then walks the four levers in the order they must be decided, with the arithmetic for each threshold: the 60–75% take-or-pay anchor, the 21.3% token crossover density, the $2.10 hedge strike against a $1.70 spot clear, and the 237,442-token outcome-versus-metered crossover.

No lever can be evaluated without the floors it must clear. From the cost stack in the primer, per GPU-hour:

• Lever 1 — contract mix — The share of installed capacity sold as multi-year take-or-pay certainty versus held merchant.
• Lever 2 — packaging — Whether a given slice is sold as GPU-hours or multiplexed into tokens, endpoints and outcomes.
• Lever 3 — dynamic price — The sequential price path on residual capacity, plus the financial hedge on hours that would otherwise perish.
• Lever 4 — discount governance — The list-to-realized waterfall: what the desk is allowed to concede, and under what control.

The levers are decided in order, because each one sizes the slice the next one operates on.

┌────────────────────────────────────────────────────────────────────────┐
│ 1. Set Take-or-Pay Floor  ──► 2. Token Multiplexing ──► 3. RL Residual │
│    (60–75% @ $2.05/hr)         (Density > 21.3%)        & Futures Hedge│
└───────────────────────────────────┬────────────────────────────────────┘
                                    ▼
                         4. Discount Waterfall
                         (Cap leakage < 0.5%)

Target rate: list $2.35, realized $2.05, hard floor $1.95 per GPU-hour. The role of this lever is not margin maximization — it is transferring utilization risk to anchor customers so that fixed depreciation is covered before any merchant hour is sold.

1,024-GPU cluster at 100% contract coverage ($45.3M capex)

Rule. Never let take-or-pay drop below 60% of installed capacity. Below that, fixed depreciation of $1.01/GPU-hr is spread across too few contracted hours and dilutes fleet gross margin regardless of how well the merchant book prices.

Token serving is a multiplexing arbitrage: the same node is sold many times over to concurrent requests, so revenue per GPU-hour scales with traffic density rather than with the rate card.

Crossover density

crossover = $16.80 base cost ÷ $78.84 peak revenue ≈ 21.3%

At 60% traffic density the node earns $47.30/hr ($5.91/GPU-hr) — a 75.6% gross margin against roughly 33% on the same silicon sold as reserved capacity.

Rule. Route uncommitted capacity into token serving only when sustained traffic density is expected to exceed 21.3%. Below that threshold the endpoint destroys value relative to simply renting the node.

The remaining 15–25% of the fleet is spot and on-demand. Pricing here is sequential, not static: a customer priced out today defers and may return, so today's discount destroys tomorrow's price. That is the structure a contextual bandit exploits in the RL simulation tab.

• Dynamic policy — Hold price at $2.40–$2.60 while the deferred-demand backlog is high; discount to $1.60–$1.80 only into soft demand weeks, to clear perishable hours above the $0.485 cash opex floor.
• Hedge the tail — Short compute index contracts on 30–35% of uncommitted volume at a $2.10 strike. If spot clears at $1.70 the hedge returns +$0.40/hr, holding cash EBITDA margin above 70%.
• Never hedge the core — Take-or-pay hours are already fixed; hedging them converts a stable book into a speculative one.

Outcome versus metered pricing

Pricing an agent workflow at $0.52 per resolved task, against a token cost of $0.63/M and a metered list of $2.19/M:

crossover = ($0.52 ÷ $2.19) × 1,000,000 = 237,442 tokens per task

Sell outcome-based pricing for agent tasks consuming under 237k tokens, where it carries the higher margin; revert to metered token rates for long-context tasks above that line.

Waterfall discipline

• Dynamic policies & cohort elasticity — The RL simulation tab on the cockpit.
• Cluster hedge & payback sensitivity — The compute futures page.
• List-to-realized discount waterfalls — The pricing & packaging tab on the cockpit.
• Full unit-cost economics and worked P&Ls — The P&L primer, sections 4 and 5.

All figures are illustrative — synthetic data built from public benchmarks, consistent with each other but not any specific company's economics.