06 / CAPACITY ECONOMICS

Sizing a GPU commitment

Reserved GPUs cost less per hour, but you pay while they sit idle. Compare that fixed bill with renting capacity only when work arrives.

THE ECONOMIC QUESTION

Will you use the discount?

At $2.00 per reserved hour and $6.00 on demand, a GPU must be needed more than 33.3% of the time for a commitment to save money. At the threshold, the two bills are equal.

The same rule applies to the next GPU you reserve. Covering the busiest hour of the year can leave it idle through the rest.

ONE GPU · A 100-HOUR PERIOD

Reserve it for all 100 hours$200

Rent it for 50 busy hours$300

Bars show paid hours. Prices are illustrative and follow the inputs below.

CAPACITY MODEL

Choose how much to reserve

100 GPUs of average demand
400 GPUs at the peak

Reserved rate · $/GPU-hour

On-demand rate · $/GPU-hour

The reserved rate is paid every hour. The on-demand rate applies only to work above your reservation.

Reserved & working Reserved & idle Rented as needed

How many GPUs will you reserve? 100 GPUs

Cost per GPU-hour served

—

Change from all on demand

—

Reserved utilization

—

Work served on demand

—

Reserved bill—

On-demand bill—

Total per modeled hour—

Hours are sorted from highest to lowest demand, so this is a distribution, not a calendar. The model assumes on-demand capacity is always available.

Demand assumptions & importWorkload shape, rates & CSV data

Price ratio 3.0× · $2.00/hr

Peak-to-average demand 4.0×

Average demand · GPUs

Observed hourly demand · CSV or text

Synthetic demand · 180 modeled intervals

Assumptions: synthetic stationary demand; 100-GPU average; $2.00 committed; $6.00 elastic; hourly billing.

How the commitment model worksThe model & its limits

The committed layer is paid for every hour whether used or not. Demand above it is served at the elastic rate; unused committed capacity is paid idle. Import a single demand column or place demand in the final column after a timestamp. One optional header row and blank lines are allowed. Missing, invalid, or negative demand values stop the import; earlier columns are never substituted. Commitments use whole GPUs, and imported scenarios are optimized against the observed distribution.

commit a GPU ⇔ its expected utilization ≥ committed ÷ elastic price

The cost curve often has a broad minimum. Sweep K from zero to the peak and watch modeled cost fall, bottom out, and rise. Sizing error may be less important than duration, delivery, concentration, and counterparty exposure; this tool does not price those risks.

Volatility lowers the committed layer. Holding rates constant, a higher peak-to-average ratio places more demand in low-frequency intervals and moves the modeled optimum down.

A larger committed-rate discount raises the committed layer. Breakeven utilization equals committed price divided by elastic price, so lower committed pricing makes less frequently used capacity economical in the model.

Limitations. The tool assumes the observed distribution repeats, elastic capacity is always available, rates are flat, and one GPU-hour is interchangeable with another. It omits growth, minimum billing increments, placement, performance, outages, and contract optionality.

Decision supported: How many GPUs should we commit?
Primary output: Cost-minimizing commitment and utilization
Model boundary: Duration, not chronology; no outages or counterparty risk

Model output, not a demand forecast. Imported observations preserve the demand distribution but discard their time order; apply your own growth, seasonality, and contract-risk judgment. Imported files stay in the browser and are not embedded in share links.

model · newsvendor critical ratio · load-duration framing after Modal, “How to price serverless” (2026) · methods & limitations