Enter your own GPU price and throughput. See where self-hosting beats the API, and where it never does.
Runs entirely in your browser. The GPU price and throughput you type are not sent anywhere, and no model API is called.
API rates come from the LLM pricing card, checked 2026-09-22.
Always-on: every GPU is billed 24 × 7, 730 hours a month. Scale-to-zero is not modelled yet.
Enter the GPU price per hour to see results. The page does not fill in a price, because a quote you actually got is the only honest input.
A self-hosted model costs the same whether it serves ten requests or ten million, as long as the GPUs stay on. An API bill starts at zero and grows with every request. This page puts both monthly figures side by side for the traffic you describe, and finds the daily request count where the two lines cross.
The break-even is the lowest request count at which the API costs at least as much as your GPUs. When traffic outgrows the cards you entered, the page adds GPUs and the self-host line steps up, so there can be more than one step on the way. If the API stays cheaper all the way to 100 million requests a day, the page says there is no crossing rather than inventing one.
A GPU delivers its benchmark throughput only while it is busy, and queueing delay climbs steeply as it nears full load. The page plans for a utilization ceiling, 70% by default, and bills the GPU-hours that ceiling implies. A lower ceiling needs more hours for the same traffic, which pushes the break-even up or leaves it where it is. The sensitivity table shows the same question at 20, 40, 60 and 80 percent.
The VRAM estimate counts only the weights: parameters times bits, divided by eight, plus 20% for runtime overhead. It does not include the KV cache. That memory grows with context length and with the number of requests served at once, so a model that fits on paper can still need a bigger card or a shorter context.
The answer depends entirely on the numbers you enter. Engineering time, monitoring and storage are excluded unless you enter them as fixed monthly cost. Latency, model quality and data-handling rules are not compared either. An open-weight model is not the same model as the hosted API one, so a lower bill is a fair comparison only when the quality is good enough for your job.
No, and it does not try. GPU prices change quickly and availability often matters more than price, so the price field starts empty. Enter the quote you actually got. The API side uses the same rate card as the LLM pricing page, with the date it was checked.
From your own serving stack. Run the model with the framework and batch size you plan to use, and measure output tokens per second per GPU and prefill tokens per second. Published figures are usually best-case numbers for short prompts, so they tend to flatter self-hosting.
Because a GPU is billed for every hour it runs. At low traffic most of those hours sit idle, which is why self-hosting looks expensive for small workloads. Only the always-on mode is modelled here; scale-to-zero, with cold-start delays, is not included yet.
Under the assumptions you entered, the API stays cheaper than your GPUs at every request count up to 100 million a day. That usually means the GPU price is high or the model's throughput is low. It is not a claim that self-hosting never pays.
When the traffic needs more GPU-hours than the cards you entered can deliver under the utilization ceiling, the break-even calculation adds GPUs. The monthly cost above keeps the count you entered, and the capacity warning tells you how many you would need.
Here GB always means 10⁹ bytes, for both the VRAM you enter and the weights estimate. If a vendor's spec sheet uses GiB, convert it before you enter the number.
At 16-bit precision the weights of a 70-billion-parameter model take about 140 GB, at 8-bit about 70 GB and at 4-bit about 35 GB. With the 20% runtime overhead the page adds, 8-bit (about 84 GB) already exceeds one 80 GB card, while 4-bit (about 42 GB) fits. Enter 70 as the model size and pick 16, 8 or 4 bits to see the estimate.
Per-million token rates for each model, with the day each card was checked.