Hetzner GPU Server
Hetzner’s new GEX44 GPU server, built around an RTX 4000 SFF Ada with 20 GB VRAM, is drawing interest as a relatively cheap way to run AI workloads compared to US hyperscalers and other GPU rental platforms. Commenters weigh its strengths—very competitive price/performance, solid long‑term reliability for many users, and Hetzner’s cost-optimized hardware stack—against notable drawbacks such as limited VRAM for serious training, stricter account verification that can block some signups, and weaker support and tooling compared with full‑featured cloud ecosystems. Overall it is seen as attractive for sustained, raw compute needs if you can handle more operational responsibility and potential support friction.
Overall sentiment on Hetzner (non-GPU)
- Many long-term users report excellent price/performance, especially for dedicated servers and auctions.
- Common praise: low latency, generous bandwidth, strong uptime (multi‑year, even decade‑long runs), and responsive support when needed.
- Hetzner’s bare‑bones style and lack of “managed” features are seen as acceptable or even desirable for technically capable users.
Reliability and business suitability
- Some claim Hetzner is unsuitable for “crucial business” workloads, citing bad experiences, sudden shutdowns, or poor support.
- Others strongly disagree, reporting years of stable production use with few incidents and prompt hardware replacements.
- Consensus: great for cost‑sensitive workloads if you have good backup/replication and can tolerate handling more ops yourself.
Account verification, geography, and onboarding friction
- Multiple users report being asked for passport scans and still getting rejected (notably from Asia, but also EU/US), leaving Hetzner with their documents and no account.
- Others say that after failed automatic checks, manual verification via support quickly resolved issues.
- Verification and limited instant trials are explained as anti‑fraud and abuse mitigation measures, especially for dedicated servers.
- Some perceive the process as hostile or opaque; a Hetzner representative frames it as necessary for safety and ToS enforcement.
New GPU server (GEX44, RTX 4000 SFF Ada)
- Priced around €184/month; roughly comparable to ~$0.30/hr in other GPU clouds when normalized, and cheaper than some competitors.
- GPU has 20GB VRAM: adequate for 13B‑class models and quantized larger LLMs but seen as limited for serious training; better suited to inference or small‑scale tuning.
- Benchmarks shared: Mixtral variants run acceptably (tens of tokens per second) but are VRAM‑constrained, sometimes forcing CPU offload.
- Several commenters argue that for real training, cards like A5000/A6000/A40/L40(Ada) or short bursts on A100‑class instances make more sense.
Cost: renting vs buying hardware / big cloud
- Hetzner is often 3–10× cheaper than major clouds for raw compute if workloads are continuous.
- Reasons cited: bulk hardware buying, cheaper datacenter power, custom servers/racks, and aggressive cost engineering.
- Tradeoffs: renting sacrifices asset ownership and resale value, but avoids capex, co‑location fees, and hardware management.
- For spiky or short‑lived GPU jobs, pay‑per‑second clouds (e.g., GCP, others) may be more economical than a monthly Hetzner GPU.