Last updated: 2026-06-22

The self-host vs API break-even explained

The self-host vs API break-even is the monthly token volume at which the fixed cost of hosting a language model on rented or owned GPU capacity equals the variable cost of calling a hosted API that charges per token. Self-hosting substitutes a large fixed monthly cost and a low marginal cost per token for the API's pure per-token pricing. Below the break-even volume the API is cheaper because the fixed cost is underused; above it the fixed cost amortizes across enough tokens that self-hosting wins.

The formula

Let $F$ be the fixed monthly GPU cost, $p_a$ the API price per million tokens, and $p_v$ the self-host variable cost per million tokens. The two options cost the same when the fixed cost equals the per-token savings on the volume processed. Solving for the break-even token volume $T_b$:

A break-even exists only when $p_a > p_v$; if the API's per-token price does not exceed the self-host marginal cost, self-hosting never overtakes it. At an actual monthly volume $T$, the two monthly costs are

and the self-host saving is $S = C_a - C_s$, positive above break-even and negative below it.

Worked example

Take a fixed GPU cost of F = \1{,}500$per month, an API price of$p_a = $10$per million tokens, and a negligible self-host variable cost$p_v = $0$ on already-provisioned capacity. The break-even volume is

Below 150M tokens the API is cheaper; above it self-hosting is. At an actual volume of T = 500 million tokens, the API costs C_a = 500 \times \10 = $5{,}000$while self-hosting costs the flat$C_s = $1{,}500$, a saving of$S = $3{,}500$ per month. The gap widens as volume grows, because the API cost scales linearly while the self-host cost stays flat once capacity is provisioned.

Variations and caveats

The single API price is a blend of input and output rates weighted by their share of traffic; the variable self-host cost captures electricity and bandwidth, which the Inference Energy Cost Calculator calculator can quantify. The model assumes provisioned capacity is sufficient for the volume. If utilization is low — idle GPUs waiting on sporadic traffic — the effective fixed cost per token rises and break-even moves higher; if a single instance cannot serve peak load, additional instances make part of the "fixed" cost behave as variable. Self-hosting also adds responsibility for scaling, uptime, model updates, and security that a managed API absorbs.

Application

The break-even volume is the cost threshold for an infrastructure decision. Comparing it against projected monthly usage — which LLM Monthly Cost Calculator helps estimate — indicates which option is cheaper today, and how much headroom exists before the answer flips. Because published API prices and GPU rates change over time, and the model omits one-time setup and idle time, the figure is a baseline to validate against current quotes and measured utilization.