AI server: buy, rent in the cloud or host in France?
There are three ways to get computing power for AI: rent it by the hour in the cloud, buy a machine and install it on your own premises, or buy a machine and have it hosted. The right choice comes down to three questions: how many compute hours per month, how sensitive your data is, and how large your model is. Here is how we help our customers decide.
Data center GPU prices are covered in detail in our guide NVIDIA H100, H200, B200 and B300 prices in 2026.
The three options compared
| Criterion | GPU rented by the hour in the cloud | Machine purchased, on your premises | Machine purchased, hosted with us in France |
|---|---|---|---|
| Setup time | A few minutes | Delivered ready to use in 7 days | 7 days, installed and connected by us |
| Cost | Billed by usage, high with continuous use | One-off purchase, plus electricity | One-off purchase, plus hosting |
| Confidentiality | Depends on the provider and the law that applies to it | Maximum, the data never leaves your premises | High: your machine, in France, with a French company |
| Noise, heat, electricity | Nothing to manage | Must be planned for, especially for a rack server | Managed by us, with monitoring |
| Ideal for | Occasional peaks, testing | Daily use, sensitive data | Intensive use, 8-GPU servers |
When does buying pay off?
A rented GPU costs little for a few hours, but a lot if it runs every day. To compare, divide the price of the machine by the monthly saving: the monthly cost of the equivalent rental, minus electricity and hosting. The result is the number of months after which the purchase has paid for itself.
The more regular your usage, the shorter that period. A machine that sits idle at night can even generate income: our GPU profitability calculator estimates what renting out its computing power can earn.
DGX Spark-type workstation, multi-GPU workstation or 8-GPU server?
| Machine | Memory | Who it is for | Price |
|---|---|---|---|
| NVIDIA GB10 Grace Blackwell 128 GB workstation (DGX Spark type) | 128 GB of unified memory | Developers, testing large models on a desk | On quotation |
| Threadripper workstation, 1 to 3 RTX PRO 6000 Blackwell | 96 to 288 GB | Teams, models with 70 billion parameters and more, fine-tuning | From €21,990 excl. VAT |
| 8-GPU NVIDIA H200 SXM server | 8 × 141 GB | Inference for dozens of users, training | On quotation |
| 8-GPU NVIDIA B200 server or B300 Blackwell Ultra | Models with several hundred billion parameters | Training and large models | On quotation |
| 8-GPU AMD Instinct MI355X server | Alternative to NVIDIA | Large models, training | On quotation |
The GB10 workstation loads large models in a compact case, but its unified memory is slower than that of a dedicated graphics card. It is well suited to testing and development; to serve a team, a multi-GPU workstation responds faster. To put a model such as Mistral Small 4 into production at full precision, Mistral recommends at least four H100 GPUs, two H200s or one B200 system.
AI graphics card: which one to choose?
The first criterion is video memory: it sets the size of the model you can load. The second criterion is memory bandwidth: it sets how fast the answers come.
- 16 GB: the RTX 5060 Ti and RTX 5070 Ti in our Essential workstations, for models of up to 14 billion parameters.
- 32 GB: two 16 GB cards, cheaper than an RTX 5090 32 GB, which is now hard to find.
- 72 to 96 GB: RTX PRO 5000 72 GB and RTX PRO 6000 96 GB, for models with 70 billion parameters.
- 141 GB: H200 NVL in PCIe format, for enterprise servers.
Hosting with us: how does it work?
Just as with our miners, your machine can be installed, powered and monitored on our premises, with remote access. You remain its owner. It runs on French electricity that was 95.2% low-carbon in 2025, according to RTE. It is the ideal solution for an 8-GPU server, which is too loud and too power-hungry for an office.
Have a project? Tell us about your models, your software and your budget: we will propose a tested configuration, delivered or hosted. Call us on +33 6 32 49 26 41, write to contact@happymining.eu, or request a quote. You can reserve with a deposit, refunded in full if you decline the quote.
Our server and hardware guides
- NVIDIA H100, H200, B200 and B300 prices in 2026
- For an individual workstation: AI PC: which computer for local AI?
- For the legal framework: Local AI in business: GDPR and AI Act
Frequently asked questions
What is an AI server? A server fitted with several GPUs and designed around them: power supply, cooling and a fast link between the cards, so it can run continuously.
Workstation or GPU server? A workstation is quiet and fits under a desk. A rack server offers more power, but it is loud and needs a dedicated power supply: it is made for a server room or for hosting.
Can I start small and scale up? Yes. The Essential 16 GB workstation accepts a second graphics card, and we will rework the configuration with you if your needs grow.
Sources