Servers, bandwidth, and inference on one invoice.
Every resource on the Carpathian platform carries a published rate. Your invoice is the sum of the ones you are running, prorated to the day you add or release each one, and it arrives as a single monthly bill rather than a set of separate charges to reconcile.
Base Spark instance
Price a footprint before you provision it
Set the resources, add as many instances as you plan to run, and the total updates against the same catalog rates the dashboard charges. The figure below is not an estimate or a starting-from number, it is what the configuration costs.
Configure Spark instances
Every instance in this configuration gets the same resources. Mixed sizes are priced the same way, one instance at a time.
Bandwidth is billed once per network and shared by every server on it, so the plan above covers all 1 instance rather than being charged against each one.
Provision partway through a month and the first invoice covers only the days remaining, then $10.00 a month after that. There is no setup fee and no minimum term.
What each component costs
Every resource is priced on its own, so memory scales without dragging cores along with it and disk grows without pushing you to the next size up. These are the live catalog rates, read from the same source the dashboard bills against.
Two ways to run
A Spark instance is a virtual machine on a compute node we own and keep well under capacity. A Quasar instance is a physical machine of your own, for workloads that would rather not share anything.
Spark instance
Virtual machines with full root access, built from the components above and resized whenever the workload changes.
- vCPU, memory, and disk scale independently
- Ubuntu, Debian, CentOS, or Windows Server
- SSH, RDP, or screen sharing from the first boot
- Standard disk images you can export at any time
Every instance includes the base network tier, monitoring, and DDoS mitigation.
Quasar instance
In developmentDedicated bare metal with an optional GPU, managed by Carpathian and configured in the builder the same way a Spark instance is.
- A physical machine rather than a share of one
- Hardware maintained and replaced by Carpathian
- Configured component by component, as Spark is
Dedicated hardware is expensive to overprovision, so start with what the workload needs and scale up as it grows. Quasar is not yet available to provision.
Network plans
Every instance arrives with a base speed tier and data allowance. Beyond that, a network plan belongs to your organization rather than to a machine, so every server on the network shares one plan instead of each carrying its own bandwidth charge.
100 Mbps / 50 GB
100 Mbps with 50 GB of data
500 Mbps / 500 GB
500 Mbps with 500 GB of data
1 Gbps / 3 TB
1 Gbps with 3 TB of data
Unlimited data
An add-on to any speed tier. The data allowance comes off and the speed you chose stays.
When you reach the data allowance
You decide in advance what happens, and the setting belongs to the organization rather than to an individual server. Neither option interrupts anything or takes a machine offline.
- Throttle modeThe default. Speed drops to the base tier until the next billing cycle and nothing further is charged, so the invoice cannot move on you.
- Overage modeOpt in and full speed continues past the allowance, with the additional usage billed and capped at $120 a month.
AI inference rates
Models are metered per million input and output tokens at the rate published for each one. Usage lands on the same monthly invoice as your infrastructure, with no separate AI subscription and no minimum spend.
Medium grade
Enhanced reasoning and deeper analysis. Grades group models by capability and cost, and every one of them is reachable through the same API key and the same endpoint.
What the price includes
Why the rates look the way they do
The hardware is ours. It sits in our own racks in a Midwest datacenter, and much of it is equipment retired on a depreciation schedule rather than because anything failed, rebuilt to spec and burned in before it carries a workload.
Because we are not servicing the cost of new equipment or reselling somebody else's capacity, the difference shows up in the rate rather than in a margin. That is also why the rate can stay put: there is no promotional period to expire and no second-year price to discover.
- Compute nodes we own, in a facility our own staff can walk into
- Node density kept low, so a busy neighbour is not your problem
- One team accountable from the metal up to the application
Questions about billing
How is my monthly bill calculated?
Am I billed for CPU time, requests, or traffic?
What happens if I add or remove a resource partway through the month?
Is bandwidth billed per gigabyte?
Do I need a contract or a minimum term?
How is AI usage priced?
Can I price a configuration before I commit to it?
What is a Quasar instance?
Work that is priced by the project
Consulting, infrastructure assessments, migration planning, and custom application work are quoted against the scope rather than sold from a rate card. Tell us what you are trying to build and you will get a figure and a timeline from the engineers who would do the work.
Configure it, price it, then deploy it.
Create an account, build the footprint you priced above, and have it running with an IP address and root access in minutes. Nothing on the invoice will be a component you did not choose.