VPS Capital, LLC owns and operates dense NVIDIA A100 compute in a colocated datacenter facility. We rent our servers to teams training, fine-tuning and serving large AI models — by the hour, with no commitment.
Our live fleet. Every machine below is powered on, listed and available to rent today — this is not a catalogue of what we could build.
| Server | GPUs | CPU / RAM | Storage | Network | Status |
|---|---|---|---|---|---|
|
VPS-SERVER01
ID #144579
|
8× A100 SXM4
80GB each · 640GB total · NVLink 300 GB/s
|
EPYC 7763
128 cores · 1TB RAM
|
1.6TB NVMe
3,346 MB/s
|
810 / 866 Mb/s
Static IP · 256 ports
|
Online
✓ Verified
|
Servers are rentable whole or partitioned — take 1, 2, 4 or all 8 GPUs on a node. Additional capacity is being commissioned; this table is updated as each server comes online.
Specifications for our A100 SXM4 server class — the configuration deployed in VPS-SERVER01.
Most A100 rentals are single cards or pairs. Eight on one NVLink fabric puts frontier-scale open models inside a single server.
Where a dense 8× A100 server earns its keep over newer, thinner instances
Full-parameter and adapter fine-tuning across 8 GPUs with NVLink-speed gradient exchange. 1TB of host RAM covers optimiser-state offload.
Throughput-oriented offline jobs — dataset labelling, embeddings, synthetic data, evaluation runs — where cost per token matters more than latency.
Image and video generation at scale. 80GB per card handles high-resolution latents and long sequences without aggressive tiling.
Tensor-parallel serving of 400B–750B parameter models on a single server via vLLM, SGLang or TensorRT-LLM.
Simulation, molecular dynamics and general CUDA workloads. A100 FP64 throughput remains strong for scientific compute.
Colocated, monitored hardware with 99.8%+ measured reliability — multi-day training runs are not a gamble.
Our capacity is listed on the Vast.ai marketplace, which handles account setup, billing, SSH keys and container templates on our behalf. You can also contact us directly for sustained-use or reserved-capacity arrangements.
Sign up and add credit at cloud.vast.ai. No commitment and no minimum term.
Filter for GPU type A100 SXM4 and set the count to 8, or search machine ID 144579 directly. Our host name shows as VPS Capital.
Use a stock template — PyTorch, vLLM, CUDA — or supply your own Docker image. CUDA drivers are already in place.
On-demand for guaranteed runtime; interruptible bidding for materially cheaper throughput work that tolerates preemption.
Instances come up in minutes with SSH and Jupyter access, plus direct port mappings for anything you need to expose.
Who you are transacting with
VPS Capital, LLC is an infrastructure company that owns GPU server hardware and rents compute capacity on it. We purchase the servers outright, place them in a colocated datacenter facility with redundant power and cooling, and operate them ourselves — we are the hardware owner and operator, not a reseller of someone else's cloud.
What we sell: metered access to GPU compute — dedicated and shared NVIDIA A100 server time, billed by the hour, along with the local NVMe storage and network bandwidth attached to each rental. Customers are AI and machine-learning teams, research groups and independent developers who need large-VRAM hardware without buying it.
Questions about the hardware, reserved capacity, sustained-use pricing, or a workload you are not sure fits? We operate the metal and answer directly.
support@vpscapital.netVPS Capital, LLC · GPU & dedicated server rentals