Published evidence
A manufacturer specification, a dated rate card, or a reported deal.
Example: 72 GPUs in an NVL72 rack, from the manufacturer’s specification.
EVIDENCEMETHODS & SOURCES
How to interpret the numbers
The desk combines published facts, author estimates, and scenario calculations. Their sources, dates, and boundaries determine what each number can support.
Author estimates
A range inferred from adjacent evidence, with more uncertainty than a directly quoted value.
Example: an estimated price range on the atlas. Keep the estimate label attached.
INFERENCEModel outputs
A calculation under stated assumptions. Its result changes when those assumptions change.
Example: $2 reserved ÷ $6 on demand = 33.3% breakeven utilization.
MODELA SIMPLE NORMALIZATION
What does one GPU-hour cost?
Divide the instance price by the number of GPUs included. This makes the billing units comparable; different hardware, networks, and terms can still deliver different value.
HYPOTHETICAL 8-GPU INSTANCE
$24per instance-hour÷8GPUs=$3per GPU-hour
Running continuously for a 730-hour month would cost $17,520 before additional charges. This is an arithmetic example, not a provider quote.
MODEL BOUNDARIES
The question each tool can answer
| Tool | Useful for | Keep in mind |
|---|---|---|
| Capacity planning | Comparing a reserved layer with on-demand capacity at given rates. | Demand is assumed to repeat. Sorting imported hours removes chronology; growth, outages, minimum bills, and counterparty risk are omitted. |
| Network performance | Estimating one ring all-reduce communication component alongside your compute and overlap assumptions. | Application goodput, congestion, sharding, storage, and other NCCL optimizations are outside the model. Use measured bandwidth when available. |
| Network layouts | Counting switches and links under explicit radix, endpoint, rail, oversubscription, and redundancy assumptions. | A vendor’s allocation of ports and redundant planes may differ. |
| Contract scenarios | Tracing how a defined incident and illustrative terms produce exposure or service credits. | The amounts are not probability-adjusted or additive savings. The fictional clauses are not legal advice; review real agreements with counsel. |
| Rack & pod models | Understanding physical relationships among compute, networking, power, and cooling. | These are explanatory models, not construction drawings. Geometry, placement, and some cable counts are simplified. |
How to read the evidence labels
- LIST
- A public provider rate card or official specification. Configuration, region, term, and verification date belong with the value.
- MKT
- A dated marketplace observation. It does not guarantee available capacity.
- DEAL
- A value reported in a filing, earnings call, customer announcement, or attributable press report. It may not compare directly with list prices.
- EST
- An author calculation or range derived from adjacent public evidence. The default atlas view shows on-demand observations with source links. Select “All observations” to include estimates; compare the evidence grade before using a value.
- MODEL
- An illustrative scenario output. Its assumptions are part of the result.
What a “GPU-hour” price includes
Prices are shown in US dollars per GPU-hour. An instance price is divided by the number of GPUs in that instance. Unless a source note says otherwise, values exclude tax, storage, support, egress, and data-transfer charges.
The chip’s form factor, host CPU and memory, network, region, availability, and commitment terms can make two GPU-hours very different products. Use the atlas to ask better questions about a quote; its points are not offers you can execute.
How dates and updates are handled
Archived observations keep their original verification dates and historical-data labels. A new quote requires a new source and verification date; changing a label does not refresh the evidence.
Interactive calculations can be used with your own assumptions. Published hardware and price facts should be checked against the linked source before a purchasing or engineering decision.
Made by Alex Yen · source code & corrections
Designed, written, and implemented by Alex Yen. The source code is on GitHub; corrections are welcome through the issue tracker.
Original code is MIT licensed. The pinned Three.js runtime retains its upstream MIT notice. Linked specifications, rate cards, filings, and reporting remain subject to their original terms.