Buying AI Servers in the Cloud

What this article covers

• When a cloud-based AI server makes sense.
• Which providers offer GPU instances.
• What to watch for when renting.
• Cost calculations and alternatives.
• Key differences from local hardware.

Introduction: Buying AI servers in the cloud

Owning your own AI PC or server isn’t always the right choice. If you experiment occasionally, train models that need computing power for a limited time, or simply don’t want to operate expensive hardware yourself, GPU instances in the cloud are a practical option. Cloud providers offer high-performance GPUs, plenty of RAM, and fast networking for shorter or longer periods. That said, cloud costs add up quickly, and data protection matters.

This article explains what to look for when renting a cloud-based AI server.

Key terms

• GPU instance: A virtual server with a GPU attached.
• VPS: Virtual private server, usually without a GPU.
• Bare metal: A physical rented server.
• Spot pricing: Discounted but cancellable cloud resources.
• On-demand: Flexible resources at standard rates.
• Hourly rate: Billing per hour.
• Egress: Outbound data transfer costs.
• Reserved instances: Long-term bookings at reduced rates.

When does a cloud server make sense?

• Quick experiments without your own hardware.
• Training runs lasting more than a couple of hours.
• Load spikes that your local hardware can’t handle.
• No space or desire to maintain your own equipment.
• Access needed from multiple locations.
• Testing costs before investing in your own setup.

If you use AI consistently and regularly, you should evaluate a local or rented dedicated server instead.

Popular providers

Hetzner

• GPU servers with Nvidia RTX 3090, RTX 4090, and H100 in select locations.
• Transparent pricing and GDPR-compliant.
• Hourly and monthly rates available.
• Excellent choice for European users.

AWS EC2

• g4dn, g5, p4d, and p5 instances.
• Flexible scaling.
• Higher prices, though spot instances are cheaper.
• Extensive tooling ecosystem.

Google Cloud

• Compute Engine with GPUs.
• A2/A3 instances for AI workloads.
• Strong Kubernetes integration.
• Watch out for high data transfer costs.

Microsoft Azure

• NC, ND, and NV series.
• Strong Windows and enterprise integration.
• Custom fine-tuning models available.

Other providers

• RunPod, Vast.ai, Lambda Labs, and Paperspace specialize in GPU clouds.
• Often cheaper, but with less support.

What to check with a provider

• GPU model and VRAM.
• RAM and CPU core count.
• Storage capacity and type.
• Bandwidth and data transfer costs.
• Location and GDPR compliance.
• Cancellation and availability terms.
• Pricing structure and hidden fees.

Cost examples

Provider	GPU	RAM	Rough hourly rate
Hetzner	RTX 4090	128 GB	approx. 2-4 EUR
AWS g5.xlarge	A10G	16 GB VRAM	approx. 1.30 USD
GCP a2-highgpu-1g	A100	85 GB	approx. 3-4 USD
Azure NV48ads	A10	24 GB	approx. 3 USD

Prices change constantly. You’ll need to calculate your own costs.

Calculating your own costs

Cost per hour × hours per day × days per month

Example: 2 EUR per hour × 4 hours per day × 30 days = 240 EUR per month.

For the same budget, owning your own server can be cheaper long term.

Data protection and security

• Data shouldn’t sit in the cloud without a contract.
• Check GDPR and data processing agreements.
• Use encryption for storage and transfer.
• Access via SSH keys, not passwords.
• Set up firewalls and reverse proxies.
• Don’t store API keys or secrets in images.

Differences from local hardware

Aspect	Cloud	Local
Initial investment	Low	High
Ongoing costs	Variable	Power and maintenance
Scaling	Simple	Limited
Data protection	Depends on provider	Under your control
Latency	Network-dependent	Low
Flexibility	High	Hardware-bound

Further reading and resources

• BotServ.de Building an AI PC
• BotServ.de Buying a GPU for local AI
• BotServ.de Energy efficiency
• BotServ.de Notebooks for AI
• BotServ.de Cloud cost calculator

FAQ: Cloud AI servers

Should I rent or buy? With regular use, owning hardware often pays off. For experiments, cloud is cheaper.

What GPU do I need in the cloud? At least 12 GB VRAM for 7B models, 24 GB or more for larger ones.

Is my data safe in the cloud? Only if the provider is GDPR-compliant and uses encryption.

How much does a cloud GPU instance cost? Depending on the provider, anywhere from a few cents to several euros per hour.

Can I keep models on a cloud server? Yes, but watch billing carefully and back up regularly.

Sources and further reading

• Hetzner Cloud: https://www.hetzner.com/cloud/
• AWS EC2 GPUs: https://aws.amazon.com/ec2/instance-types/
• Google Cloud GPUs: https://cloud.google.com/compute/docs/gpus/
• Azure GPU VMs: https://azure.microsoft.com/products/virtual-machines/

Summary: Buying AI servers in the cloud

A cloud-based AI server works well for flexible experimentation, short-term training, and comparing costs with local hardware. European providers like Hetzner are attractive for data protection, while hyperscalers offer scaling and many features. If you use the cloud, scrutinize hourly rates, data transfer costs, storage, and contract terms carefully. For ongoing, regular AI workloads, your own server or a dedicated rental solution is usually the more economical choice.