The six tiers, defined
A price only means something next to prices of the same kind. Every row on TideRack carries one of these tiers and one GPU form.
Secure on-demand
Provider-run capacity you rent by the hour or minute, no contract, no interruption by design. This is the like-for-like baseline.
69 price points in this snapshot, 18 providers. Lowest H100: $1.68 · Latitude.sh (unspecified, Secure)
Community cloud
Capacity from third-party hosts run through the provider platform (for example RunPod Community Cloud). Cheaper, with less control over where and on what hardware it runs.
9 price points in this snapshot, 1 providers. Lowest H100: $1.99 · RunPod (PCIe, Community)
Marketplace (host-set)
An open market where independent hosts set prices (Vast.ai). We show the lowest listed price and, where shown, the median. Quality and reliability vary by host.
7 price points in this snapshot, 1 providers. Lowest H100: $2.13 · Vast.ai (SXM, Market)
Spot / preemptible
Interruptible capacity that can be reclaimed. Cheap per hour, but a job may need to restart: the estimator adds your own rerun rate.
18 price points in this snapshot, 6 providers. Lowest H100: $0.79 · Nebius (SXM, Spot)
Reserved / committed
A price for a committed term. The term is not always on the pricing page we read, so these are shown as "starting from" and never mixed into on-demand comparisons.
6 price points in this snapshot, 3 providers. Lowest H100: $2.72 · Hyperstack (SXM, Reserved)
Serverless (per second)
Billed per second only while your code runs. We convert the per-second rate to an hourly figure for comparison.
9 price points in this snapshot, 2 providers. Lowest H100: $3.949 · Modal (SXM, Serverless)
Forms
| Form | Meaning |
|---|---|
| SXM | GPU on an HGX board with the fastest GPU-to-GPU fabric. Normally the multi-GPU training form. |
| PCIe | Standard add-in card. Usually cheaper, lower GPU-to-GPU bandwidth. |
| NVL | NVLink-bridged PCIe pairs or the vendor's "NVLink" naming; shown separately and never ranked against SXM. |
| form not stated | The vendor page names the GPU without a form. These rows are listed but excluded from like-for-like estimator runs. |