AI team: ten experiments/week, two GPU hours each. €2/h rental seems obvious. Three months later same nightly training: bill exceeds €350/mo dedicated RTX never ordered.
Cloud GPU = flexibility. GPU server = fixed cost. Experimentation wants first; repeated production shifts second.
Trap: dedicated for POC; on-demand cloud for 24/7 prod without inference autoscaling.
On-demand cloud GPU
Many cards without capex, scale up/down, spot discounts. Weaknesses: availability, expensive dataset egress, opaque billing.
| Phase | Cloud | Dedicated |
|---|---|---|
| 2-wk POC | Yes | No |
| Recurrent fine-tune | Spot | If >40% util |
Dedicated GPU server
Fixed price, local disk, data on machine. Paid idle, manual upgrade. Profile: nightly training, stable inference.
Calculate €/GPU-hour used over 90 days — not catalog €/h.
Hidden cloud costs
Egress to bucket, snapshots, image license, attached disks. Spot preempt without checkpoint = lost run.
The peak: experiment cheap, produce predictably
Log before deciding
Without GPU-hour log per experiment, cloud vs dedicated stays intuitive. For thirty days note model, duration, spot interrupt, real cost. Multiply by twelve for annual projection vs monthly RTX or L40S server.
Consider confidentiality: training corpus with personal data requires EU region and DPA, cloud or bare metal. Hourly card price does not replace that diligence.
Go further with the inquiry
Apply the article to your concrete case: what volume, what service level, what contractual constraints from clients or regulators? Write answers before opening the pricing page. Request written documents — full grid, DPA, certification scope, renewal terms — not homepage screenshots.
Compare at least two providers via the directory and compare tool on identical thirty-six-month assumptions. Share the short list with whoever signs the contract and whoever operates infrastructure daily. Archive terms dated at subscription so due diligence is defensible later.
Go further with the inquiry
Apply the article to your concrete case: what volume, what service level, what contractual constraints from clients or regulators? Write answers before opening the pricing page. Request written documents — full grid, DPA, certification scope, renewal terms — not homepage screenshots.
Compare at least two providers via the directory and compare tool on identical thirty-six-month assumptions. Share the short list with whoever signs the contract and whoever operates infrastructure daily. Archive terms dated at subscription so due diligence is defensible later.
Log before deciding
Without GPU-hour log per experiment, cloud vs dedicated stays intuitive. For thirty days note model, duration, spot interrupt, real cost. Multiply by twelve for annual projection vs monthly RTX or L40S server.
Consider confidentiality: training corpus with personal data requires EU region and DPA, cloud or bare metal. Hourly card price does not replace that diligence.
Log before deciding
Without GPU-hour log per experiment, cloud vs dedicated stays intuitive. For thirty days note model, duration, spot interrupt, real cost. Multiply by twelve for annual projection vs monthly RTX or L40S server.
Consider confidentiality: training corpus with personal data requires EU region and DPA, cloud or bare metal. Hourly card price does not replace that diligence.
Decide and move forward without blind spots
Start by framing your real scenario on one page: volumes, client constraints, required support level, and a thirty-six-month budget including domain, backup, and essential add-ons. Compare at least two hosts via the directory and compare tool with identical assumptions, archive pricing grids dated on subscription day, then have both the contract signer and daily infrastructure operator validate the short list.
Frequently asked questions
Cloud or dedicated fine-tuning?
Cloud trials; dedicated if recurrent runs >40% util/mo.
Spot for AI?
Yes batch + checkpoint; no client demo without save.
Minimum VRAM?
7B inference ~8–12 GB; fine-tune often 24 GB+.
EU hosting?
OVH, Scaleway, Outscale — region and dataset egress.
Experiment pays à la carte; industrialize stops paying idle GPU.
