Ngl this started as a joke and became a trauma log. I needed GPU inference nodes, not even training, just LLM hosting for a side project. 24GB VRAM minimum, 48GB preferred.
Host 1: Contabo. "GPU instances coming Q2." Q2 passed. Q3 passed. Ticket response: "the team is excited." I left.
Host 2: OVHcloud. Actually had A40s. Price fine. Then they migrated my instance to "new hardware" and the CUDA version was incompatible with my compiled wheels. 6 hours rebuilding. Left.
Host 3: Vultr. Advertised RTX A6000. Delivered... a6000 via passthrough from some guy's gaming rig? Performance was 30% under spec. Lol
Host 4: Hetzner. DDoS on neighbor IP, null-routed my /24 for 18 hours. No communication.
Host 5: InterServer. Billing system double-charged, took 11 tickets to fix. Support used emojis unironically.
Host 6: Hostinger. "Unmetered" except 100Mbps after 2TB. Buried in AUP.
So I made this. Full comparison, pain rated 1-10.
| Host | GPU | Price/mo | Downtime | Pain |
|------|-----|----------|----------|------|
| Contabo | promised | €0 | ∞ | 4 |
| OVHcloud | A40 | $420 | 6h | 6 |
| Vultr | "A6000" | $380 | 0h | 8 |
| Hetzner | A40 | €450 | 18h | 9 |
| InterServer | A100 40GB | $680 | 0h | 5 |
| Hostinger | A40 | $390 | 0h | 7 |
Now? I need host 7. Guess who has actual A40s in stock, verified benchmarks, and I never finished the trial with.