Server Plans and Pricing: VPS, GPU and Dedicated

Check certification requirements for the specific data center and service you intend to order. If your project requires security certifications, audit reports or a data processing agreement, request the applicable documents and confirm their scope, validity and covered legal entity before purchase. A facility certification does not automatically certify your application or satisfy its legal obligations. You remain responsible for application access controls, data handling and any sector-specific requirements. Include your hosting location and data residency constraints when discussing the deployment with support.

Start with the API documentation available for your account and the particular product. Confirm which operations are supported—such as instance creation, power control or configuration changes—and check authentication, permissions and request limits before building automation. Do not assume a Terraform provider, SDK or Kubernetes integration is included unless it is documented for that service. Test scripts against a non-production instance, keep credentials out of source code, and handle timeouts and retries carefully so a failed response does not create duplicate orders. Ask support to verify any required integration before committing to a deployment.

Choose a location close to your users and application dependencies, then test the actual plan's route from the networks that matter to you. CN2 GIA can be relevant for traffic to mainland China, but routing varies by carrier, direction and time of day; the label alone does not guarantee a particular latency. Use a plan-specific test IP and compare peak-hour Ping, MTR and transfer results. Bandwidth, measured in Mbps or Gbps, is a rate; monthly transfer, measured in GB or TB, is a data allowance. A 10Mbps connection has a theoretical ceiling of about 1.25MB/s before overhead. Check metered directions, reset dates, overage rules and fair-use restrictions separately. A 1Gbps port does not imply unlimited transfer or sustained 1Gbps throughput.

Check the backup features included with your selected plan: whether backups are automatic or manual, how often they run, where they are stored, how long they are retained and whether restores incur charges. A snapshot on the same storage system is not a substitute for an independent backup. Keep an off-server copy of important files and database backups, and test restoration before relying on it. Recovery point objective (RPO) describes how much recent data you can afford to lose; recovery time objective (RTO) describes how quickly service must be restored. Set these targets for your application and verify them through a restore exercise rather than assuming a universal recovery guarantee. Before reinstalling or cancelling a server, export any data you need to retain.

The SLA attached to your service or written quotation determines the availability target, exclusions and any service credits. Review what is measured, which incidents qualify, the claim deadline and the evidence required; a credit schedule for one product may not apply to another. Support response time is also different from resolution time. When reporting an incident, include the instance or order ID, timestamps with a time zone, affected services and relevant diagnostic results. If you need a contractual response target, dedicated contact or managed application support, agree those requirements before ordering rather than assuming they are part of every plan.

RTX 4090: 24GB GDDR6X, 16384 CUDA cores, 82 TFLOPs (FP32), ideal for cost-effective training and inference ($250-350/month). Tesla A100: 40/80GB HBM2e, 312 TFLOPs (TF32), ECC memory, enterprise-grade reliability ($1100-2000/month). H100: 80GB HBM3, 1000 TFLOPs (FP8), 4th-gen Tensor Cores, for frontier AI research ($3500+/month). Multi-GPU configurations: 2-8 cards with NVLink (600GB/s interconnect) or NVSwitch. PCIe Gen4/5 support. Real-world benchmarks: ResNet-50 training 40x faster than 8-core CPU, GPT-3 fine-tuning viable on A100.

Deep Learning Stack images: CUDA 12.2, cuDNN 8.9, PyTorch 2.1, TensorFlow 2.14, JAX, Jupyter Lab, VS Code Server, NGC containers. Python scientific stack: NumPy, Pandas, scikit-learn, Matplotlib. Distributed training frameworks: Horovod, DeepSpeed, Megatron-LM, Ray. Container runtime: Docker, Kubernetes with GPU operator. Bare-metal Ubuntu/CentOS also available for custom setups. One-click deployment of popular models: Stable Diffusion, LLaMA, Whisper. Free environment setup assistance and optimization consulting included.

Scaling options: 2-8 GPUs per node via NVLink (600GB/s bandwidth), multi-node clusters for large-scale training. Supported parallelism: data parallel (PyTorch DDP/FSDP), model parallel (Megatron, DeepSpeed), pipeline parallel. Performance: 4x A100 achieves 3.5-3.8x linear speedup. Network: 10-25Gbps InfiniBand or Ethernet for inter-node communication. Cluster management: Kubernetes, Slurm, or custom orchestration. Example: Train 175B parameter model across 8-node (64-GPU) cluster. Consulting provided for distributed architecture design and optimization. Ideal for enterprise ML teams and research institutions.

Enterprise management: sub-accounts with role-based access control (RBAC), resource quotas per team/project, cost allocation tags, usage dashboards with GPU utilization metrics. Pricing flexibility: hourly on-demand ($0.8-$1/hr for RTX 4090), monthly reserved instances (20-30% discount), annual commitments (40-50% discount), spot instances (up to 70% off for interruptible workloads). Cost optimization: auto-shutdown during idle periods, GPU scheduling, resource recommendations. API integration for automation and CI/CD pipelines. Dedicated TAM and quarterly business reviews for annual contracts. Volume discounts available for 10+ GPUs.

Storage options: 1-4TB NVMe local SSD per GPU server (3000-7000MB/s throughput), expandable to 20TB+. Object storage (S3-compatible): $0.02/GB/month, 10Gbps internal network transfer to GPU servers. Dataset pre-loading service: ship hard drives to datacenter, we deploy to your servers. Transfer methods: direct upload (Hong Kong 100Mbps-1Gbps, US up to 10Gbps), rsync incremental sync, parallel multi-threaded uploads. Access public datasets: fast downloads from Hugging Face, Kaggle, ImageNet (no egress charges within region). Example: 100GB dataset upload in 15 minutes via S3 pre-load + internal 10Gbps download.

Dedicated servers give you the resources of a physical machine, while cloud servers provide a virtual machine with the allocation specified by the plan. There is no universal performance multiplier: CPU generation, storage layout, resource sharing and the application itself all matter. Dedicated hardware can suit sustained database, virtualization or compute workloads; cloud servers can be easier to start with when capacity needs are smaller or change frequently. Compare the total cost, management responsibilities, backup arrangements and upgrade process, then benchmark your own application before choosing.

All dedicated servers include full IPMI (Intelligent Platform Management Interface) out-of-band management at no extra cost. Features: remote KVM console access (control server as if physically present), power management (reboot, power cycle even if OS crashed), virtual media mounting (install any OS via ISO remotely), hardware monitoring (CPU temperature, fan speeds, power supply status, disk health), event logging and alerts. IPMI operates independently of the server OS, providing access during system failures. Delivery time: standard configurations deploy within 4-6 hours of order; custom configurations (specific CPU/RAM/storage) require 1-3 business days. 24/7 IPMI access via secure web interface.

Hardware RAID controller included with all dedicated servers. RAID options: RAID 0 (striping, maximum performance, no redundancy), RAID 1 (mirroring, 2 disks, 100% redundancy), RAID 5 (3+ disks, single-disk fault tolerance, balanced performance/capacity), RAID 10 (4+ disks, mirrored stripes, best performance + redundancy). Disk types: SATA SSD (economical, 500MB/s), enterprise SATA SSD (higher endurance), NVMe SSD (3000MB/s+, lowest latency), SAS HDD (large capacity storage). Recommendations: RAID 10 + NVMe for databases requiring high IOPS; RAID 5 + SATA SSD for web servers balancing cost and reliability; RAID 1 + NVMe for small high-performance setups. Hot-swap support and spare disk options available for mission-critical deployments.

Use the payment methods, currency and billing terms shown at checkout for the selected service. Review the total, any gateway fees and renewal terms before paying. For cryptocurrency payments, use the exact network, address and amount supplied with the order. If your business needs a purchase order, bank transfer, company receipt or special billing arrangement, confirm the available process and documents with support before placing the order. Do not assume deferred payment or a particular local tax document is included. After payment, check the order status before attempting a second payment.

Standard support: 24/7/365 coverage via ticket system, live chat, email, phone. Response times: critical issues (server down, network outage) within 15 minutes, high priority within 1 hour, normal within 4 hours. Enterprise support: dedicated Technical Account Manager (TAM), direct phone hotline, priority queue, 15-minute guaranteed response for P1 incidents, quarterly business reviews (QBR), architecture design reviews, proactive monitoring. Support scope: infrastructure troubleshooting, OS installation, network diagnostics, security hardening, performance optimization, migration assistance. Excludes: application-level code debugging (available as professional services). Multi-language: English, Simplified Chinese, Traditional Chinese, Japanese. On-site support available (chargeable).

Eligible first-time cloud server purchases have a 24-hour refund request window, subject to the applicable conditions, including less than 1GB of used transfer. Submit the request within 24 hours of purchase; approval and the time for funds to arrive are separate matters governed by the service terms and payment method. GPU, dedicated and custom resources may be excluded, so check the selected product before paying. To limit evaluation costs, test the network first and start with a short billing term. Where hourly GPU billing is available, confirm the minimum charge and whether stopping an instance leaves storage or other charges running.