Hetzner suspended my "unmetered" VPS at 8TB, fair or fine print?

I want to correct myself because I did the math wrong in my head. 8TB is actually a lot for LLM inference if you're just doing API calls. I was thinking training, my bad. For serving a 7B model with 4-bit quant, you're moving maybe 4GB per load, not pulling constantly. 8TB would be millions of requests.

Still stands that providers oversell though. The 10Gbps shared thing was a guess, not a fact.

Brazilian here, I feel the pain on cheap euro pricing that turns expensive with IOF and spread. Hetzner bills in EUR so that 4 becomes 6 real quick with my card.

More relevant: I had a CX11 in Falkenstein, moved to Vultr São Paulo because latency to Rio users was 220ms vs 35ms. The "unmetered" was never the issue, the routing was. If your cousin's audience is in Argentina, why is the server in Germany anyway?

Istanbul to Hetzner Falkenstein is 45ms, not bad. But I learned the hard way: Turkish regulators sometimes block German IPs in waves, had to move a client to Vultr London. Unrelated to bandwidth but another reason "cheap and unmetered" isn't the whole picture.

For video processing specifically, have you looked at Wasabi or Backblaze for egress? Pair a small compute VPS with object storage, might be cheaper than pretending one box does everything.

My nephew corrected me. Water bill is metered, you pay per gallon. I meant "unlimited" like the buffet where they kick you out for staying four hours. Bad analogy, sorry.

The Dell guy was talking about my PowerEdge T140, not relevant. I should stop posting before coffee.

Late to this but I wanted to check something. From memory Hetzner's cloud terms have something about reasonable use and their FAQ says "most customers use under 1TB." I asked support once what happens at 5TB, they said "it depends." Depends on what? Phase of moon?

I moved my personal stuff to OVH in Gravelines, they have explicit 10TB on the 5 EUR plan. Not great but at least it's a number. Paris to Gravelines is 180km, latency under 5ms. Lyon users should care about this more than benchmark numbers.