Should you buy a GPU and self-host an open LLM, or just keep paying per token for an API? It comes down to volume. Enter your monthly usage and your hardware, and this calculator shows the break-even point — the moment owning the GPU becomes cheaper than the API bill.
Your usage
Your self-host rig
| API cost (your volume) | — |
| Self-host cost (GPU amortized) | — |
| Self-host cost (electricity) | — |
| Self-host total / month | — |
Self-hosting runs open-weight models (free weights), so this compares the per-token API bill against owning hardware. It assumes your GPU can keep up with the volume (a single GPU has a tokens/sec ceiling) and ignores your setup/maintenance time. Check what a GPU can actually run in our VRAM calculator, and current API prices in the cost calculator.
Remember: self-hosting runs open-weight models, so factor in the quality difference versus a frontier API — and use our VRAM calculator to confirm your GPU can actually run the model you want.
Quick answer: Is it cheaper to self-host an LLM or use an API?
It depends almost entirely on your sustained monthly token volume. For most users a hosted API is cheaper: below roughly 50 million tokens a month, per-token pricing beats the cost of buying and running a GPU. Self-hosting only pays off at high, steady volume — typically tens to hundreds of millions of tokens a month feeding agents, batch jobs or a whole team — where an owned GPU kept busy amortises its upfront cost against effectively unlimited inference. The break-even is a range, not a fixed line: it moves with model size, GPU tier, electricity price and, above all, how heavily you keep the card utilised.
- Low volume (under ~50M tokens/month): the API wins almost every time.
- The decision zone (~50M–500M tokens/month): a genuine toss-up, but only worth self-hosting if you have dedicated engineering time to run it.
- High sustained volume (agents, batch or a whole team, 500M+ tokens/month): a well-utilised owned GPU can cut inference cost by up to ~5x.
- A GPU only pays for itself if it is kept busy — an idle card still costs its full capex and power.
- Budget “flash” or small-model APIs are so cheap that self-hosting is rarely justified at any volume.
Frequently asked questions
When does a local GPU pay for itself?
A GPU pays for itself once your steady API spend on an equivalent open model consistently exceeds the all-in cost of owning one. A consumer GPU of roughly $2,000–$2,500 spread over three years works out to about $55–$70 a month, plus $30–$60 in power — so owning one costs roughly $85–$130 a month all-in. If your comparable API spend stays under that, the API is cheaper; once it runs well above it — on the order of a few hundred dollars a month — the card recovers its $2,000–$2,500 upfront cost within about a year and saves money after that.
How much VRAM do I need to run an LLM locally?
At 4-bit quantisation, budget roughly 0.5–0.6 GB of VRAM per billion parameters, plus extra for the KV cache that grows with context length. In practice a 7B model needs about 8 GB, a 13B model 12–16 GB, and a 4-bit 70B model roughly 40–48 GB. That means a 7–13B model fits on a single 8–16 GB card, but a 4-bit 70B is too big for any single consumer GPU — even a 32 GB card falls short, so it takes two 24 GB cards, a 48 GB workstation card, or a smaller, more aggressively quantised model.
What are the hidden costs of self-hosting an LLM?
The GPU price is only part of it — once you add electricity, cooling and especially engineering time, self-hosting typically costs 3–5x the raw GPU rental figure. Plan for 10–20 hours a month of setup, monitoring and maintenance; that labour, plus the cost of a card sitting idle between jobs, is what sinks most naive break-even estimates.
Does the model size change the break-even point?
Yes. Bigger models need more or pricier GPUs, which raises the capex you have to amortise, but they also carry the highest per-token API prices, which pushes the break-even lower in volume terms. A small 7–13B model is cheap to self-host but also cheap via API, so APIs usually win; a 70B-plus open model is where a busy owned GPU tends to deliver the largest savings.
Does electricity make self-hosting too expensive?
Usually not — power is a minor line item next to the hardware. A single desktop GPU run flat out draws roughly 400–600 W, which at typical electricity rates works out to only around $30–$60 a month if it runs continuously. The dominant cost of self-hosting is the upfront hardware and the engineering time to operate it, not the electricity.
Should a small team or startup self-host or use an API?
Start with an API. Until you have high, predictable volume and someone to run the infrastructure, per-token pricing is cheaper, faster to ship and carries no upfront risk. Self-hosting is worth revisiting only when your monthly bill is large and stable — typically once you are consistently past tens of millions of tokens a month on a comparable open model.
