Use the cloud when you need a ready-to-run environment without managing hardware. Prices are based on each GPU-hour and billed by the second. Billing starts when the instance runs and stops at shutdown, with no hidden fees.
These are Glows.ai's hourly starting prices in USD. Billing begins when an instance runs and stops at shutdown, with no hidden fees. Actual rates may be lower depending on the region, contract term, and reserved capacity.
| Model | Hourly price from | Typical use |
|---|---|---|
| H100 | $2.96 | A performance baseline for mainstream training and fine-tuning |
| A100 | $1.20 | Training and batch inference with mature frameworks |
| L40S | $0.83 | Inference, vision models, and video processing |
| RTX 6000 Ada | $0.72 | Workstation-class inference, graphics, and simulation workloads |
| RTX 4090 | $0.49 | Entry-level development, education, and research |
KONST's published H100 bare-metal price is $2.00 per GPU per hour. The same GPU costs $2.96 in the cloud. These are not two prices for the same product. They assign idle-capacity risk differently.
With bare metal, you rent the entire server and billing starts by the month. You absorb the cost of idle time in exchange for the lowest unit price, exclusive PCIe and InfiniBand bandwidth, and performance unaffected by neighboring workloads. In the cloud, the platform absorbs idle-capacity risk. You pay by the second only while the instance is running, and the rate includes the platform layer, elastic resource pool, and spare capacity kept ready.
The break-even point usually depends on utilization. For teams with predictable workloads and long-term utilization above sixty percent, bare metal or a long-term contract is almost always less expensive. If demand fluctuates, the architecture is still being validated, or peak traffic needs overflow capacity, the cloud can save more in idle costs than the difference in hourly price. Bare-metal terms of three months or longer receive lower rates and can be combined with cloud resources.
Self-service provisioning with billing only while the instance is running. Suitable for fluctuating demand and validation work.
Reserve predictable baseline demand for a lower unit price. Bare-metal contracts of three months or longer receive lower rates.
When comparing proposals, check how each part is calculated. That gives a more useful comparison than the hourly price alone.
The main charge, based on each GPU-hour and billed by the second.
Runtime storage is included. Long-term storage is billed by capacity used.
First, compare billing granularity. During development with many short jobs, per-second billing can differ by double-digit percentages from rounding each job up to an hour or even a day.
Second, include environment setup time. A lower rate does not help if every start requires reinstalling drivers and frameworks or every epoch requires downloading the dataset again, because the GPU is billed throughout that work. Starting from images and snapshots and avoiding repeated dataset downloads directly reduces the total bill.
Third, check for hidden fees. Uploads, downloads, capacity expansion, and migration often cost extra. KONST does not add charges for these items, so include the other provider's terms in the comparison.
Tell us the model size and expected usage schedule, and we will help estimate a suitable instance and plan.
