The breakeven, in one line
Self-hosting trades a per-token price for a fixed monthly cost, so it wins only when your volume is high enough that the equivalent API bill would exceed that fixed cost. The crossover is simply fixed monthly host cost ÷ API price per token. Below it, pay per token; above it, a dedicated host can be cheaper — with one big caveat below.
Why idle time is the real decider
An API charges you for exactly the tokens you use. A GPU charges you for every hour it exists, busy or not. If your traffic is spiky — heavy during the day, quiet at night — a host you use 30% of the time costs more than three times its "effective" rate, which can erase the entire saving. Steady, high-utilization workloads are where self-hosting genuinely pays; bursty ones usually favour the API well past the raw breakeven.
A worked example
At 500M tokens a month and a $3-per-million blended API rate, the API costs about $1,500/month. A dedicated host at $1,800/month all-in is moreexpensive here — you don't cross breakeven until ~600M tokens. Push volume to 1.2B tokens and the API would be ~$3,600 while the host stays $1,800: now self-hosting saves ~$1,800/month — provided the GPU is busy enough to deliver that throughput. The decision is volume andutilization, never volume alone.
How to read your breakeven number
The calculator returns the monthly token volume where the two paths cost the same. Place your real usage against it:
- Below ~70% of breakeven— stay on the API. You'd be paying a fixed GPU bill to serve a load the per-token price handles for less, with zero ops. Not close.
- Within ~70–130% of breakeven— the API almost always still wins. At the margin, the host's ops time, idle hours, and reliability work tip it back; the paper saving is inside the error bars.
- Above ~130% with steady, predictable traffic — self-host becomes worth modelling seriously. Confirm you can keep the GPU busy (see utilization below) before you commit.
- Above breakeven but bursty— usually still the API, or a hybrid: API for the spiky tail, a reserved host only for the steady base load. Don't buy 24/7 capacity for a daytime peak.
Common mistakes
- Counting only the GPU rental. The sticker rate is a fraction of the real number. Deployment and maintenance engineering, monitoring, and on-call all load onto the fixed side — and they recur.
- Assuming 100% utilization. The single biggest error. You pay for 720 hours a month whether you use 700 or 200. Divide the headline rate by your honest utilization to get the cost per useful hour before comparing anything.
- Comparing a spot price to an API SLA.A $1.65/hr marketplace H100 and a managed API are not the same product — one can be reclaimed mid-request. Compare on-demand to API, or price the reliability work you'll do to make spot safe.
- Ignoring model-version lag. A managed API ships the new frontier model on day one; a self-host stays on whatever you deployed until you do the upgrade work. That gap has a real, if invisible, cost.
- Forgetting egress and storage. Moving weights and data in, and serving responses out, bills separately and can rival the compute line on data-heavy workloads.
How to use it
Estimate your real monthly token volume, your blended API rate (price your exact model mix in the LLM API Cost Calculator), and an honest all-in host cost including ops. If you're below breakeven, the answer is easy. If you're above it, pressure-test utilization before committing — and remember the API also buys you instant access to new models and zero maintenance.
What this assumes — and what it deliberately leaves out
The breakeven is intentionally a clean ratio so you can reason about it. That means it models the things that dominate the decision and omits the ones that vary too much to assume on your behalf:
- Assumes: a fixed all-in monthly host cost you provide (GPU + the ops you load into it), a blended API price per token, and that you compare like-for-like throughput. Breakeven = fixed host cost ÷ API price per token.
- Leaves out (model it yourself):your actual utilization curve, autoscaling headroom, egress/storage, batching and quantization gains, committed-use or spot discounts, and the engineering cost of new model upgrades. These don't cancel out — most push the real breakeven higher than the raw number, favouring the API longer than the napkin math suggests.
Where the inputs come from
| Input | Source / how to get it honestly |
|---|---|
| GPU $/hr (the host cost floor) | Live, source-linked on-demand rates in our GPU Price Index — dated source checks, not a vendor estimate. |
| Blended API price per token | Price your exact model mix in the LLM API Cost Calculator rather than guessing a single rate. |
| Utilization | Your own observability — requests/hour over a representative week. If you don't have it, you are not ready to self-host. |
| Ops / engineering load | Your estimate of deploy + maintain + on-call hours × loaded cost. Conservative is correct here. |
Frequently asked questions
When is self-hosting an LLM cheaper than using an API?
Only above a breakeven volume. Self-hosting swaps a per-token price for a fixed monthly GPU + operations cost, so it wins only once your usage is high enough that the equivalent API bill would exceed that fixed cost. Below the breakeven, the managed API is cheaper and far less work; above it, a dedicated host can save money — if you keep it busy.
What is the breakeven volume for self-hosting?
It's the fixed monthly self-host cost divided by the API price per token. If a host costs $1,800/month all-in and the API charges $3 per million tokens, breakeven is 600 million tokens per month. Process less than that and the API is cheaper; process more and self-hosting starts to pay — assuming high utilization.
What costs does self-hosting add beyond the GPU rental?
The GPU is the visible cost; the hidden ones decide the outcome. Engineering time to deploy and maintain the stack, idle-GPU waste (you pay 24/7 but may use a fraction of the day), autoscaling and reliability work, and the lag in getting new model versions a managed API ships day one. A host used 30% of the time erases most of the headline saving.
Should I self-host to save money?
Treat the breakeven as a floor, not a green light. If you're well below it, don't — the API is cheaper and simpler. If you're well above it and your traffic is steady enough to keep GPUs busy, model the full cost (including ops and idle time) before switching. Many teams that 'should' self-host on paper still come out ahead on a managed API once their real utilization is counted.
Independent analysis, not infrastructure advice. It compares cost only — not latency, data residency, or control, which can each justify self-hosting on their own. Validate against your real utilization and ops overhead before switching. GPU rates show their last verification date; see methodology.