Size your capacity.
Compare it to the API.
How many GPUs your team needs to self-host a coding model, what that costs per seat, and how it compares to paying an API per token for the same work. Hardware only, not the time to run it yourself.
Qwen3-32B-class coder in FP8, near-lossless vs full weights. Needs FP8 hardware (Ada or Hopper).
Prices from Podcode's RunPod templates (secure, on-demand). Set 0 for hardware you already own.
Everything sent and generated, including context an agent re-sends each step.
Each active session's context lives in GPU memory, so longer contexts mean fewer sessions per pod. The longest supported context depends on the model.
Billed hours; pause the pod the rest of the time. ~21.7 working days a month, applied to both sides.
Each pod fits about 13 sessions in memory at 32k context and about 80 by compute. Seat 38 adds the next pod (+$553/mo). Cost grows in steps, so between steps each new teammate costs nothing extra.
| Seats | Pods | Hardware/mo | Per seat/mo | API per seat | Difference |
|---|---|---|---|---|---|
| 1 | 1 | $553 | $553 | $89 | +$464/seat |
| 2 | 1 | $553 | $277 | $89 | +$187/seat |
| 5 | 1 | $553 | $111 | $89 | +$21/seat |
| 7break-even | 1 | $553 | $79 | $89 | −$10/seat |
| 8you | 1 | $553 | $69 | $89 | −$20/seat |
| 10 | 1 | $553 | $55 | $89 | −$34/seat |
| 15 | 1 | $553 | $37 | $89 | −$53/seat |
| 20 | 1 | $553 | $28 | $89 | −$62/seat |
| 30 | 1 | $553 | $18 | $89 | −$71/seat |
| 37 | 1 | $553 | $15 | $89 | −$75/seat |
| 38 | 2 | $1,106 | $29 | $89 | −$60/seat |
| 50 | 2 | $1,106 | $22 | $89 | −$67/seat |
Throughput is normalized from public vLLM/SGLang batch benchmarks and scaled by GPU memory bandwidth. Memory assumes 90% of VRAM for serving, ~33GB of weights and an FP8 KV cache. API prices are list rates; cache writes and long-context surcharges aren't modeled. Treat this as a capacity-planning estimate: engine, context length and tuning all move the real numbers.