Self-hosting an LLM or paying per token: where the break-even really is
Running an open-weight model on a GPU you rent is cheaper than paying per token only past a volume threshold, and dearer below it. The GPU bills every hour, busy or idle; an API bills only the tokens you use. With OVHcloud's list price for one H100 (€2.80 an hour before VAT, read on 6 October 2026) and Evanesca's price for Llama 3.3 70B, the threshold sits around 1.6 billion tokens a month for one card — and a 70B model realistically needs two.
Published 6 Oct 2026 · checked on 20 Oct 2026
The whole question is one division
Renting a GPU and calling an API are billed in different units. The machine bills time; the API bills consumption. Comparing them takes one line:
tokens per month = (hourly price of the machine × 730) ÷ price per token730 is the number of hours in an average month. Put both sides on the same tax basis: Evanesca's prices include VAT, so the machine is counted with VAT too. The result is the same with VAT left out of both sides — the rate cancels — but mixing a price before tax with one after tax would make self-hosting look about a fifth cheaper than it is.
The machine side: what an H100 costs
OVHcloud's public catalogue for France, read on 6 October 2026, lists the h100-380 instance — one NVIDIA H100 with 80 GB of memory, 30 vCores and 380 GB of RAM — at €2.80 an hour before VAT, or €3.36 with it (catalogue, price page). Run all month, that is €2,452.80 with VAT; the monthly plan for the same machine costs €1,940 before VAT.
| Instance | GPUs | Per hour, excl. VAT | Per hour, VAT incl. | A month round the clock, VAT incl. |
|---|---|---|---|---|
| h100-380 | 1 × H100 80 GB | €2.80 | €3.36 | €2,452.80 |
| h100-760 | 2 × H100 80 GB | €5.60 | €6.72 | €4,905.60 |
| h100-1520 | 4 × H100 80 GB | €11.20 | €13.44 | €9,811.20 |
These are list prices. They vary by region and commitment, and storage and outbound traffic can be billed on top. Treat them as a starting figure for your own quote, not as the quote.
The API side, and the break-even
Evanesca serves Llama 3.3 70B on Scaleway in Paris, with OVHcloud as a fallback, for €1.512 per million tokens, input and output alike, VAT included — the amount actually debited, not an infrastructure cost (all prices). Dividing the machine by that price gives the break-even:
| Machine | Per month, VAT incl. | Break-even per month | Tokens per second, day and night |
|---|---|---|---|
| 1 × H100, hourly, round the clock | €2,452.80 | 1.6 billion | 620 |
| 1 × H100, monthly plan | €2,328.00 | 1.5 billion | 590 |
| 2 × H100, hourly, round the clock | €4,905.60 | 3.2 billion | 1,230 |
| 4 × H100, hourly, round the clock | €9,811.20 | 6.5 billion | 2,470 |
For scale: Evanesca's smallest credit purchase, €2, buys about 1.3 million tokens on Llama 3.3 70B, and €5 of credit about 3.3 million — or about 30 million input tokens on Gemma 4 26B, the cheapest input price in Evanesca's catalogue. Set those figures against last month's real volume, from your logs, not your forecast.
Memory decides how many cards you need
Estimates usually go wrong here: they price one card and then deploy on two. A 70-billion-parameter model stored at two bytes per parameter needs about 140 GB for its weights alone, against 80 GB on an H100. Quantised to one byte, the weights fall to about 70 GB — they fit, but leave almost nothing for the KV cache, the memory that grows with context length and with the number of requests served at once. A 70B model with usable context and batching wants two cards, which doubles the threshold to about 3.2 billion tokens a month.
Architecture changes the picture. OpenAI's card says GPT-OSS 120B fits “into a single 80GB GPU” thanks to quantised weights (model card, read on 6 October 2026). But it is also cheap per token on Evanesca: for three input tokens per output token, its blended price is €0.4410 per million, and one H100 would have to serve about 5.6 billion tokens a month to beat it.
At the other end, Kimi K3 counts 2.8T parameters (104B active): even at one byte per parameter, its weights need about 2,800 GB, far beyond the 320 GB of the four-GPU instance. Models of that size are a multi-machine project.
Utilisation is what actually decides
Turn the threshold into throughput and the picture sharpens: 1.6 billion tokens a month is about 620 tokens per second, held day and night. We give no throughput figure for your model on your hardware: it depends on quantisation, batch size, context length, the mix of prompt and output, and the serving stack. Measure it under your own load. What the arithmetic shows is that the break-even sits close to the machine's saturation point.
So the real question is whether you can keep the GPU busy every hour you pay for. An internal tool used in office hours — eight hours a day, five days a week — uses about 173 of the 730 hours in a month, roughly 24%: each token it serves costs about 4.2 times the saturated figure. Peaks make it worse: the machine is sized for your busiest hour and paid for in every quiet one.
When self-hosting does win
- Sustained, saturating volume on a single model — the one case where the division clearly favours the machine.
- Custom or fine-tuned weights that no provider serves.
- A model that no managed operator offers yet.
- A hard requirement no provider meets: network isolation, a latency target, a specific location.
Self-hosting does not settle jurisdiction by itself. A GPU rented from a US-controlled cloud, even in an EU region, stays within reach of the CLOUD Act; one rented from a provider with no non-EU control is outside its direct reach. The operator decides, not the fact of running the model yourself — the GDPR and CLOUD Act guide explains why.
Decide in six steps
- Take last month's real token volume from your logs, input and output separately.
- Multiply by the published price of the model you would use: that is your API bill.
- Size the machine you would need — memory first, then throughput — and multiply its hourly price by 730.
- If the API bill is below the machine, stop: the API wins.
- If it is above, divide the break-even by 2,628,000 to get the tokens per second you would have to hold, and ask whether your traffic holds it at 3 a.m. on a Sunday.
- Add the hidden costs to the machine side before you conclude.
If the API side wins, Evanesca has no subscription: credit from €2 that never expires, and the exact cost under every answer (pricing). The guide to open-weight models helps pick the model.
Frequently asked questions
Is self-hosting an LLM cheaper than an API?
Only above a volume threshold. A rented GPU bills every hour, busy or idle, so it wins only when kept close to saturation. Against Evanesca's price for Llama 3.3 70B, one H100 at OVHcloud's October 2026 list price breaks even around 1.6 billion tokens a month, and the two cards a 70B model realistically needs around 3.2 billion.
How many tokens a month before a rented H100 pays off?
Divide the machine's monthly cost by your price per token. At €3.36 an hour with VAT, one H100 run all month costs €2,452.80; against €1.512 per million tokens, that is about 1.6 billion tokens, or about 620 tokens per second held around the clock.
Can a 70B model run on a single H100?
Barely. At two bytes per parameter its weights need about 140 GB against 80 GB on the card; quantised to one byte they fit in about 70 GB but leave little room for the KV cache that serving several users needs. Size the machine by memory first, then by throughput.
Does self-hosting make me GDPR-compliant?
Not by itself. It removes one processor from the chain, but you still run on someone's infrastructure, and the company that controls it decides whether the CLOUD Act can reach the data. The rest of the GDPR — legal basis, minimisation, retention, rights — applies whoever runs the model.
Read next
- AI API aggregators: one key for many models, and the question of who runs themAn AI API aggregator puts many models behind one OpenAI-compatible key. How OpenRouter, Together AI, Groq, Fireworks AI and Evanesca differ.
- How to choose an AI API aggregator for your companyFive questions to ask an AI API aggregator before you sign — jurisdiction, data kept, billing, catalogue, compatibility — and how five providers answer them.
- Open-source AI gateways and inference servers you can host yourselfLiteLLM, vLLM, SGLang, Ollama, TGI, Xinference: what each self-hosted tool does, what running it costs, and when a managed EU API is the simpler route.