🟢 All systems operational · Open models on hardware you rent
+ Add Provider
Open weights · your own hardware

LLM Hosting — How Much
GPU You Actually Need

Every question about self-hosting a language model eventually becomes the same question: does it fit, and what does the card cost. This page answers the first half with arithmetic you can check, and the second half with live prices from the catalogue. The model-specific pages are one click away when you have picked a size.

Read the memory table → Which engine, which model
llm — the whole decision
1. pick a size → 8B, 32B, 70B, 120B?
2. pick a precision → FP16, FP8, 4-bit?
3. params × bytes × 1.25 = the floor
4. add the KV cache × concurrent users
✓ now you know which card to rent
# everything else on this page is detail

What you are actually choosing

Rent a card or rent an API
A hosted endpoint bills per token and costs nothing when idle. A rented card costs the same all month and nothing per token. There is a crossover point, it depends entirely on your volume, and nobody can tell you where it sits without your numbers.
Size beats brand
A 32-billion-parameter model needs roughly the same memory whoever trained it. Pick the size your task actually requires first, then choose among the models at that size — not the other way round.
Precision is a price lever
The same weights at 4-bit take a quarter of the memory they take at FP16, which is often the difference between one card and four. Quality does drop; how much depends on the model and the task, and the only honest way to know is to try both.
Weights are the floor, not the bill
Loading the model is the easy part. The KV cache grows with every token of every open conversation, so a card sized exactly to the weights serves exactly one person.
Data locality is the usual reason
Most people who self-host are not chasing a lower bill. They are removing a third party from a path that carries customer data, and that is a decision no price table can make for you.
Read the model’s licence
“Open weights” covers everything from Apache 2.0 to licences that stop being free above a revenue threshold. The engine you serve with and the model you serve are licensed separately, and only one of them is usually permissive.

The memory table, for every size

This arithmetic is ours, not a vendor requirement. No large-model vendor except OpenAI publishes a per-model GPU minimum, so we do the sum in the open: weights take roughly 2 bytes per parameter at FP16, 1 byte at FP8 and 0.5 byte at 4-bit, plus +25% for the KV cache and the serving runtime. It is a floor. This is the reference table the rest of this section is built on: the model pages quote the rows that apply to them, and they quote them from here.
Parameters FP16 FP8 4-bit 4-bit + headroom Card to rent
3B ≈ 6 GB ≈ 3 GB ≈ 1.5 GB ≈ 2 GB Anything with a GPU
7–8B ≈ 16 GB ≈ 8 GB ≈ 4 GB ≈ 5 GB 8 GB
13–14B ≈ 28 GB ≈ 14 GB ≈ 7 GB ≈ 9 GB 12 GB
30–32B ≈ 64 GB ≈ 32 GB ≈ 16 GB ≈ 20 GB 24 GB
70B ≈ 140 GB ≈ 70 GB ≈ 35 GB ≈ 44 GB 48 GB
120B ≈ 240 GB ≈ 120 GB ≈ 60 GB ≈ 75 GB 80 GB
405B ≈ 810 GB ≈ 405 GB ≈ 203 GB ≈ 253 GB A multi-GPU node

The right-hand column is the 4-bit figure rounded up to a card size people actually rent, and it assumes one user at a time. A mixture-of-experts model does not get a discount here: only a fraction of its parameters do work on any given token, but all of them still have to be resident. The exception to all of this is gpt-oss, where OpenAI publishes the figure itself.

Which engine, which model

Two separate choices that get confused with each other. The engine is the software that loads the weights and answers HTTP; the model is the weights. Any engine below serves any model below.

Pick this When Licence More
Ollama One box, one or a few users, and you want it working in two commands MIT Ollama hosting →
vLLM Many people hitting one model at once, and throughput per pound matters Apache-2.0 vLLM hosting →
DeepSeek weights You want a reasoning model, and MIT on the weights themselves matters MIT DeepSeek hosting →
gpt-oss weights You want a published hardware figure rather than an estimate Apache 2.0 GPT-OSS hosting →

Not a ranking, and not sponsored — nobody pays to appear here. If your job is images rather than text the arithmetic on this page does not apply at all; see Stable Diffusion hosting instead. If what you want is an agent that calls a model rather than one that hosts it, that is OpenClaw and Hermes Agent, and neither needs a GPU.

Cheapest GPU at each tier

All GPU plans →

The cheapest live GPU plan in the DohoHub catalogue at each memory tier, out of 142 plans whose provider publishes a GPU memory figure. Prices come from our own feed, normalised so they are comparable and shown in the currency picked in the header. Where one machine is the cheapest way to clear several tiers, it appears once and the note says which tiers it covers.

Provider Plan Specs Price Visit
Servarica
Bee GPU VPS
cheapest at 8–24 GB+ · 24 GB on board
 6 Cores24 GB vRAM30 GB RAM400 GB NVME100 TB traffic10 Gbps
$79.00/mo
 Visit
NovoServe
Supermicro X11 10SFF (GPU)
cheapest at 48 GB+ · 64 GB on board
 64 GB vRAM25 TB traffic
$355.99/mo
renews at $402.52
 Visit
GPU Mart
Enterprise GPU VPS - RTX Pro 6000
cheapest at 80 GB+ · 96 GB on board
 32 Cores96 GB vRAM84 GB RAM400 GB SSDUnlimited traffic1 Gbps
$649.00/mo
 Visit
Contabo
Dedicated NVIDIA H200
cheapest at 141 GB+ · 141 GB on board
 32 Cores141 GB vRAM234 GB RAM1900 GB15 TB traffic
$2,138.25/mo
 Visit

We record the memory a plan advertises, not which card supplies it. That is the single biggest gap in this table: memory decides whether a model loads, and the card’s generation and bandwidth decide how fast it answers, and our catalogue holds the first consistently and the second unevenly. We would rather say so than rank on a number we do not have. None of these providers sells “LLM hosting” as such; they sell GPU servers.

Pick a card that fits the size you need

Filter the catalogue by GPU memory, cores, storage and location — live prices, tracked every six hours.

Find a GPU server →
↑