MarsAI · GPU BARE METAL
Dedicated GPU servers
no virtualization layer
Self-host, fine-tune, deploy privately — your compute, in your hands. Bare metal for maximum performance — no virtualization overhead, full InfiniBand, top GPUs on demand.
01·SOLUTIONS
The right GPU for every workload
Start with what you're running, then choose by scenario. Four common workloads with recommended configs — full specs and pricing in the table below.
Self-hosted inference
Deploy open models on dedicated GPUs — high-throughput, large-context inference for your agents.
Fine-tune & train
Train or fine-tune on your own cards — near-linear NVLink multi-GPU scaling, your data never leaves your environment.
Private & compliant
Model and data stay in your environment on a physically isolated, dedicated machine — built for compliance and private deployment.
Batch & rendering
Bulk offline inference, rendering and other cost-sensitive jobs — hourly elastic, cheaper monthly.
02·SPECS & PRICING
Transparent specs. Transparent pricing
One pricing unit — per GPU-hour — with InfiniBand and VRAM right in the table. Monthly reserved is about 30% off.
| Model | GPUs | VRAM / GPU | vCPU | System RAM | NVMe | Network / InfiniBand | Hourly ($/GPU·hr) | Monthly reserved ($/GPU·hr) |
|---|---|---|---|---|---|---|---|---|
| RTX 4090Proven · value | 8 | 24GB GDDR6X | 96 | 512GB | 4TB | PCIe 4.0 | $0.59 | ~$0.41 |
| RTX 5090NEWNext-gen Blackwell | 8 | 32GB GDDR7 | 96 | 512GB | 4TB | PCIe 5.0 | $0.99 | ~$0.69 |
| RTX A6000Inference / render | 8 | 48GB GDDR6 | 128 | 1TB | 8TB | PCIe 4.0 | $1.29 | ~$0.90 |
| A100 SXMClassic training | 8 | 80GB HBM2e | 192 | 2TB | 8TB | NVLink 600 GB/s + RDMA | $1.59 | ~$1.11 |
| H100 SXMBest value | 8 | 80GB HBM3 | 224 | 2TB | 16TB | NVLink 900 GB/s + RDMA | $1.99 | ~$1.39 |
| H200 SXMLarge context | 8 | 141GB HBM3e | 224 | 2TB | 16TB | NVLink 900 GB/s + RDMA | Contact | Contact |
| B200 SXMNEWTop performance | 8 | 192GB HBM3e | 224 | 4TB | 30TB | NVLink 1.8 TB/s + RDMA | Contact | Contact |
| GB200 NVLNEWFlagship cluster | Contact | 192GB HBM3e | Contact | Contact | Contact | NVLink 1.8 TB/s | Contact | Contact |
Prices are indicative starting rates (). Monthly reserved starts from a 1-month commitment — longer commitments cost less per hour; top-end models like the B200 are quote-based, with starting prices on request; contact sales for large clusters and custom SLAs.
03·WHY BARE METAL
Why Bare Metal, not a VM
You get a whole physical machine, not a sliced virtual instance. Four points — all in the specs, all in the bill.
Zero virtualization overhead
Direct physical GPU access, no hypervisor tax — Direct access to the physical GPU for your workload.
Dedicated · single-tenant
The whole server is yours — no noisy neighbors, predictable and reproducible performance.
Full NVLink + InfiniBand
8-GPU NVLink within the node, 3.2Tb/s InfiniBand across nodes — bandwidth never bottlenecks multi-GPU.
Transparent · 0 egress
Priced per GPU-hour, no charge for traffic in/out, no hidden line items.
04·PERFORMANCE & SCALE
Tuned for throughput, built to scale out
Squeeze every FLOP from the silicon — then run the same workload across the whole cluster.
Performance
CUDA acceleration
Native CUDA on raw silicon — nothing between your kernels and the GPU.
TensorRT optimization
Compiled, fused engines for lower latency and higher tokens/sec on NVIDIA cards.
KV cache optimization
Paged KV cache and prefix reuse keep long-context inference fast and memory-efficient.
Scale
Horizontal scaling
Add nodes on demand — throughput grows near-linearly as the fleet scales out.
Multi-node inference
Shard large models across nodes over 3.2 Tb/s InfiniBand, NVLink inside each.
Load balancing
Requests spread across healthy GPU nodes automatically, with failover built in.
05·TRUST & DATA
Production-ready, your data is yours
Stable performance is the floor; isolation, data ownership and compliance are spelled out too.
Single-tenant isolation
Physically dedicated machine — no neighbors, clear security boundary, reproducible results.
Never trained on
Your data stays on your machine, never used for any model training.
A real company
Operated by MarsAI LIMITED (registered in Denver, Colorado, USA).
06·GET STARTED
Up and running, hourly or monthly
No tickets, no day-long setup — get a dedicated machine the moment your account is enabled.
Ready to run
CUDA, drivers and common frameworks pre-installed; training / inference templates out of the box.
Hourly / monthly
Billed per GPU-hour, ~30% off with a monthly reservation, stop anytime.
0 egress
No charge for traffic in/out — move models and data with no hidden bill.
07·SYSTEM ARCHITECTURE
Inside the bare-metal stack
The load balancer fans every request out across parallel GPU nodes — each running the model runtime directly on raw silicon, with no virtualization in the path.
06·FAQ
Frequently asked
Q1How is the GPU billed?
Per GPU-hour, metered to the minute. Monthly reserved is ~30% off; failed or never-started instances aren't billed.
Q2How do commitments work?
Flexible. Hourly has no commitment — stop anytime; monthly reserved starts at 1 month, longer terms cost less per hour ().
Q3Any network / traffic fees?
No. Traffic in and out (egress) is not charged separately.
Q4Will you use my data?
No. Bare Metal is single-tenant; data stays on your machine and is never used for any training.
Q5How do I buy a large cluster?
Contact sales. For clusters beyond self-serve size, custom SLAs, SSO and more, we'll quote you directly.
Dedicated compute. In your own hands
Dedicated machines, transparent pricing, zero egress — launch in minutes.