Dedicated bare metal for the whole batch,
full performance with no virtualization loss
Usage rights to a whole batch of GPU servers go to a single lessee. The equipment sits in a KONST data center, and the lessee gets physical machines rather than virtualized slices, so all performance is available. The lessee logs in remotely over a dedicated VPN and manages everything above the operating system, while KONST keeps ownership of the equipment.
Environment pre-built,
managed remotely by the lessee
Delivery starts from the initial operating system installation. KONST completes machine setup, drivers, the base environment, the separate network segment, and the dedicated VPN. Once connected, the lessee decides on frameworks, scheduling, and data storage.
- 01KONST completes installation and environment setup
Servers are racked, drivers and the base environment are built, and a separate network segment and dedicated VPN tunnel are configured.
- 02The lessee connects remotely and manages independently
Log in remotely with an account. The team decides on framework versions, scheduling tools, and data storage.
- 03The lessee completes environment acceptance
Measure cross-node bandwidth and latency with NCCL, then move to production use once confirmed.
Annual terms, with capacity reserved exclusively during the term
Terms run by the year, priced per GPU per hour and settled monthly. Most contracts include take-or-pay, paid on the agreed capacity. That batch of equipment is not released to others during the term, and the lessee controls scheduling.
Annual term,
priced per GPU per hour
The term is measured in years, priced per GPU per hour, and settled monthly.
Billed on agreed capacity,
capacity not released to others
Take-or-pay terms apply: payment is on the agreed capacity and does not vary with the month's actual usage hours. That batch of equipment is not released to others during the term.
Fees are prepaid,
and a longer term means a lower unit price
A longer term means a lower unit price, and fees are prepaid in one payment at the start. Term and project schedule are aligned before signing, and how payments are handled on early termination is stated in the contract.
Isolated network segment, external access through a dedicated VPN
Leased equipment sits in a separate network segment, with traffic separated from other tenants at the network layer. All external connections go through a dedicated VPN, and data does not cross the public internet. Nodes inside the cluster communicate over a high-speed network, and a dedicated line can be added for exchanging data with an existing data center.
Leased equipment sits in a separate network segment, with traffic separated from other tenants at the network layer.
External connections go through a dedicated VPN, and data does not cross the public internet. Nodes inside the cluster communicate over a high-speed network.
When exchanging large volumes of data with the owner's existing data center, a dedicated line can be added to avoid transfer over the public internet.
Full specs of the H200 node
Accelerators currently available to rent include H100, H200, B200, and B300. The table below lists the full specs of the H200 node. Other models differ in accelerator generation and in power and cooling, and are configured by scale and site.
| Item | Specifications |
|---|---|
| Accelerator | HGX H200 |
| Processor | Intel Xeon 8558P, dual socket, 48 cores |
| RAM | DDR5-5600 ECC |
| System drive | M.2 SSD |
| Network interface | 400G dual-port, ConnectX-7 with BlueField |
| Power supply | 3200W Titanium class |
| Security module | TPM 2.0 |
Note: The table shows a reference configuration for the H200 node. Actual delivered specs are as stated in the project quotation.
Multiple bare-metal projects delivered in Taiwan and Malaysia
An H100 engagement in Taiwan at the scale of a hundred servers on a one-year term, a long-term contract at the self-built Taichung data center running alongside a platform, and a B300 site in Malaysia at the scale of a hundred servers have all been delivered. KONST keeps equipment ownership, and lessees serve downstream customers with the capacity.
H100 at the scale of a hundred servers,
one-year engagement delivered
One-year term. Delivery covered the initial operating system installation, a separate network segment, and a dedicated VPN. KONST keeps equipment ownership, and the lessee serves its downstream customers.
Self-built Taichung data center,
long-term contract and platform in parallel
The self-built Taichung data center hosts both the long-term contract and the platform. The long-term contract covers base capacity, and the remaining capacity connects to the cloud platform for external supply.
B300 at the scale of a hundred servers, delivered whole for lease
Once the equipment is in place, the whole capacity is supplied externally as a single whole-batch lease.
FAQ
How does bare metal differ from GPU cloud?
Besides overseas GPU clouds, what should be considered when renting locally?
Can the GPU count be increased or reduced mid-term?
Who owns the equipment, and what does the rental fee cover?
From installation to ready for use, where does KONST's delivery scope end?
Services often evaluated together
Need dedicated GPU bare metal for a whole batch?
Tell us the GPU model, quantity, and term, and we will reply with the configurations available to rent and the activation timeline.