Capabilities
AI Servers & GPUs
We help you source and configure GPU servers and host platforms for training, fine-tuning, and inference — quoted to the workload you intend to run.
What we help with
Keepseed supports inquiries for AI-class GPU servers and the complementary CPU, memory, storage, and host networking those nodes need. We work from your target specification, or we help you shape one: GPU class, GPU count per node, interconnect, and the surrounding cluster assumptions.
Training, fine-tuning, and inference often want different balances of GPU class, memory, and host I/O. If you already have a bill of materials, we quote to it. If you have a workload and a set of constraints, we will propose a configuration we can support and then quote that.
We treat the server as a configuration to be quoted, not as a public inventory list. Availability depends on the platform, quantity, and destination, and is confirmed in the quotation.
Networking and liquid cooling for the same deployment can be included in the same inquiry when you want a coherent package rather than a node quoted in isolation.
Configurations we can quote
These are examples of configurations we can quote, not a catalogue of stock on hand. Availability is confirmed in the quotation. This site does not publish list prices.
-
H100-class 8-GPU training nodes
Eight-GPU nodes in the H100 class, quoted as a training or fine-tuning building block, including the host options needed to complete the node.
-
H200-class 8-GPU nodes
Eight-GPU nodes in the H200 class, quoted to the memory and training or inference assumptions you share.
-
B300-class platforms
B300-class systems quoted as specified for a given deployment, rather than as a single standard bill of materials.
-
Inference-oriented GPU servers
GPU servers shaped around a model and latency target, where the GPU class, GPU count, and host I/O follow the inference job rather than a training-node template.
-
Complementary host options
CPU, memory, local NVMe, and high-speed NICs quoted to match the node — either as part of a server configuration or as complementary options to one you already have.
Practical considerations
-
Interconnect
GPU-to-GPU topology and the cluster fabric often matter as much as the accelerator class. Name them early if they are already decided.
-
Thermal design
Air or liquid cooling should be stated when the hall or the platform already imposes a constraint. We can quote cooling as a related package.
-
Host balance
CPU, memory, and NVMe should be sized to the workload, not copied from a generic bill of materials that may not fit the job.
-
A coherent package
NICs, optics, and switches are easier to quote with the node than after a server configuration has been frozen in isolation.
What to include in an inquiry
- The workload — training, fine-tuning, or inference — and any model or latency target that should shape the quote
- GPU class and GPUs per node, plus how many nodes you are asking us to quote
- Interconnect preferences, including intra-node GPU fabric and the planned east-west network
- CPU, memory, and local NVMe assumptions, or a request that we propose them
- Target NIC and fabric speed, if already known
- Destination country and any site or import constraints we should account for
- The timeline you are working toward, and whether networking or cooling should be quoted alongside
Ready to specify a configuration?
Share the use case, the target build, quantity, destination country, and timeline. We will take it from the inquiry.