Annual term, priced per GPU-hour
Rental terms are measured in years. Pricing is per GPU-hour and settled monthly.
Bare Metal GPU Rental
Dedicated bare metal with full performance and no virtualization overhead
A single client receives exclusive use of the full group of GPU servers. The physical machines, not virtualized slices, run in a KONST data center with their full performance available. The client connects through a dedicated VPN and manages everything above the operating system. KONST retains ownership of the equipment.
SOLUTION｜ The KONST approach
A prepared environment that clients manage remotely
KONST handles server installation, the initial operating system, drivers, the base environment, an isolated network segment, and a dedicated VPN. After connecting, the client chooses the frameworks, scheduler, and data storage.
- 01
KONST installs the systems and base environment
Rack the servers, install drivers and the base environment, and configure an isolated network segment and dedicated VPN.
- 02
The client manages the environment remotely
Connect through a remote account and choose framework versions, scheduling tools, and data storage.
- 03
The client accepts the environment
Use NCCL to measure cross-node bandwidth and latency, then move into production after acceptance.
FEATURE｜ Service details and specifications
Annual contracts with capacity reserved exclusively
Terms are annual, priced per GPU-hour and settled monthly. Most contracts include take-or-pay: the client pays for contracted capacity, the equipment is not offered to anyone else during the term, and the client controls scheduling.
Pay for contracted capacity, which is not released to others
Under take-or-pay, charges follow contracted capacity rather than actual hours used that month. The equipment remains exclusive throughout the term.
Prepaid fees, with lower unit prices for longer terms
Longer terms have lower unit prices, and fees are prepaid at the start. The term and project schedule are aligned before signing, and the contract states how payments are handled after early termination.
FEATURE｜ Service details and specifications
Network isolation with all external access through a dedicated VPN
Rented equipment sits on an isolated network segment, separated from other tenants at the network layer. All external access uses a dedicated VPN, so data does not cross the public internet. Cluster nodes communicate over a high-speed network, and a separate circuit can connect an existing data center.
Isolated at the network layer
Rented equipment occupies a dedicated network segment, separated from other tenants' traffic at the network layer.
External access through a dedicated VPN
All external connections use a dedicated VPN, keeping data off the public internet. Cluster nodes communicate over a high-speed network.
Optional dedicated circuit to an existing data center
A dedicated circuit can carry large data transfers to the owner's existing data center without using the public internet.
FEATURE｜ Service details and specifications
Complete H200 node specification
Available accelerators include H100, H200, B200, and B300. The table shows a complete H200 node configuration. Other models differ in accelerator generation, power, and cooling and are configured for the project scale and site.
| Item | Specification |
|---|---|
| Accelerator | HGX H200 |
| Processor | Dual Intel Xeon 8558P, 48 cores |
| Memory | DDR5-5600 ECC |
| System drive | M.2 SSD |
| Network interface | 400G dual-port, ConnectX-7 with BlueField |
| Power supply | 3200W Titanium |
| Security module | TPM 2.0 |
Note: This is a reference H200 node configuration. The project quotation defines the actual delivery specification.
EVIDENCE｜ Results and data
Multiple bare metal deployments delivered in Taiwan and Malaysia
Completed projects include a one-year H100 contract at roughly 100-server scale in Taiwan, long-term contracts running alongside the platform in KONST's Taichung data center, and a B300 site at roughly 100-server scale in Malaysia. KONST retains equipment ownership, while clients use the capacity to serve their downstream customers.
Hundreds of H100 servers delivered under a one-year contract
The one-year delivery includes initial operating system installation, an isolated network segment, and a dedicated VPN. KONST retains the equipment, and the client serves downstream customers.
Taichung data center combines long-term contracts and platform capacity
The self-built Taichung facility supports long-term contracts and the cloud platform at the same time. Contracts cover base capacity, while remaining capacity is offered through the cloud platform.
A full B300 deployment at roughly 100-server scale
After installation, the full block of capacity is offered under a single rental agreement.
Frequently asked questions
How does bare metal differ from GPU cloud?
Bare metal provides physical machines with no virtualization layer, so all performance is available, but the client manages the operating system and cluster. GPU cloud is preconfigured and quick to provision, but resources are shared with other tenants. Long-running training at full load generally uses bare metal.
What should we consider when renting locally instead of using an overseas GPU cloud?
Consider data residency, network latency, and support hours. If data cannot leave the country, equipment location is mandatory. Local connections have lower latency than transoceanic links, and the support team works in the same time zone as the client.
Can we add or remove GPUs during the term?
Additional GPUs are usually added as new nodes and priced under the existing contract, subject to available capacity. Take-or-pay limits reductions, so clients usually adjust downward at renewal.
Who owns the equipment, and what does the rental fee cover?
KONST retains ownership. The client pays for usage rights, including data center space, power, network, and operations. KONST bears procurement funding, residual-value, and generational-refresh risk.
What does KONST deliver between installation and availability?
KONST installs the operating system, drivers, and base environment, then configures an isolated network segment and dedicated VPN. The client manages frameworks, scheduling, and data storage after handover.
Need a dedicated group of bare metal GPUs?
Tell us the model, quantity, and term. We will respond with an available configuration and provisioning schedule.