Bare metal GPU
High-performance bare-metal GPUs with full machine isolation and low-cost operation — built for AI, ML, and deep-learning workloads. Submit a reservation and our team will craft a plan with you.
Register demand · no charge · no inventory lock · reply within 24 hours
BARE METAL GPU
Dedicated GPU infrastructure
Model training
Large context
Next generation
Visual compute
Dedicated performance · on-demand delivery
Use cases
Four core scenarios with purpose-built bare-metal configs. Prices are indicative; exact quotes follow reservation confirmation.
Compute for large language model training with multi-GPU and distributed setups
8x NVIDIA H100 SXM per node
Memory
80 GB HBM3 per GPU · 640 GB total
Interconnect
NVLink 900 GB/s + RDMA
Features
¥36.00/GPU/hour
8x NVIDIA B200 SXM per node
Memory
192 GB HBM3e per GPU · 1,536 GB total
Interconnect
5th-gen NVLink 1.8 TB/s + RDMA
Features
LLM serving, real-time chat, multimodal generation, and agent inference — low latency, high throughput
8x NVIDIA H200 SXM per node
Memory
141 GB HBM3e per GPU · 1,128 GB total
Interconnect
NVLink 900 GB/s + RDMA
Features
8x NVIDIA RTX 4090 per node
Memory
24 GB GDDR6X per GPU · 192 GB total
Interconnect
PCIe 4.0 · Ada Lovelace
Features
¥6.80/GPU/hour
Agent batch jobs, AIGC generation, and cost-efficient inference
8x NVIDIA RTX 5090 per node
Memory
32 GB GDDR7 per GPU · 256 GB total
Interconnect
PCIe 5.0
Features
8x NVIDIA A800 per node
Memory
80 GB HBM2e per GPU · 640 GB total
Interconnect
NVLink + RDMA
Features
¥18.00/GPU/hour
GPU-accelerated dynamics, remote modeling, and molecular science
8x NVIDIA H100 SXM per node
Memory
80 GB HBM3 per GPU · 640 GB total
Interconnect
NVLink 900 GB/s + RDMA
Features
¥36.00/GPU/hour
8x NVIDIA L40S per node
Memory
48 GB GDDR6 per GPU · 384 GB total
Interconnect
PCIe 4.0
Features
Core advantages
Built for high-performance computing — peak performance, full control, and enterprise reliability
Direct hardware access without hypervisor loss, fully unlocking GPU compute
CUDA, cuDNN, PyTorch, and TensorFlow can be preinstalled so you start training and inference quickly
Enterprise availability commitments, dedicated support, and 24×7 response for continuity
Dedicated physical servers and isolated networking with custom security policies for compliance
Model gateway, on-demand GPUs, and secure runtimes under OpenRoute. Reserve a card type first, then scale as you grow.
Coming soon
A RunPod / Novita style marketplace for idle compute: pool idle IDC GPUs with per-second billing and off-peak discounts. Gateway users can both buy cheap compute and contribute their own idle GPUs for revenue.
Short-term or monthly whole-machine isolation, 1-on-1 sales planning, delivery as fast as same day.
Reserved access for waitlist users: per-second billing, start/stop on demand, early-bird pricing.
Buy compute to cut costs, or contribute idle GPUs to the scheduling pool for revenue.
Delivery, billing and compliance — what you may want to know