# Self-hosted vs hosted face recognition: the real total cost

> A clear-eyed TCO comparison of running open-source face recognition yourself versus a hosted API — GPUs, redundancy, model updates, staffing, and when each option actually wins.

Source: https://sightradar.com/blog/self-hosted-vs-hosted-face-recognition · By SightRadar Engineering · Updated 2026-07-17 · ~8 min read

> **TL;DR** Open-source face models can be cheap to license (check each project's actual license and model weights — terms vary) but they are not free to run: you pay for GPUs or CPU inference, redundancy, a vector database, monitoring, model updates, and the engineering time to operate all of it. A hosted API like SightRadar has a clean per-photo price (batch index $0.00062/photo) and shifts the infra ops to the vendor — you still own integration, monitoring, and governance. Self-hosting tends to win at very high, steady volume with in-house ML ops or a hard data-residency requirement; a hosted API tends to win for early, spiky, or small-team products.

"The model is open source, so it's free" is the most expensive sentence in a face-recognition roadmap. The model weights are free; running them reliably is not. Let's price the whole thing honestly.

## What self-hosting actually costs

*Hosted still needs integration, monitoring, and governance — the vendor covers the inference infra, not your whole stack.*

| Cost line | Self-hosted | Hosted API |
| --- | --- | --- |
| Model license | Varies — check each project's license | In the per-photo price |
| Inference compute | GPU/CPU instances, 24/7, sized for peak | In the per-photo price |
| Redundancy / failover | You run multi-AZ | Vendor-managed |
| Vector database | You operate + scale it | Vendor-managed |
| Model updates | You re-benchmark and re-index | Vendor-managed |
| Infra monitoring & on-call | Your team | Vendor's (you still monitor your integration) |
| Specialist staffing | ML/infra engineers | Not required for the engine |
| Scaling to a burst | Provision ahead or throttle | Elastic, pay per photo |

> **Note:** The trap is sizing compute for peak but paying for it at idle. Event and gallery workloads are bursty — a wedding weekend then nothing — so a 24/7 GPU fleet sits mostly idle while still billing.

## When self-hosting genuinely wins

- **Very high, steady volume** where per-photo API cost would dwarf a well-utilized fleet.
- **Hard data-residency / air-gap** requirements that forbid sending images to any third party.
- **In-house ML ops** already exists — you have the team and the vector-DB expertise.
- **Deep model customization** — you need to fine-tune on your own domain.

## When a hosted API wins

- **Early or uncertain volume** — you don't want to provision a fleet for demand you can't predict.
- **Small teams** — no one to carry a GPU on-call rotation.
- **Bursty workloads** — pay per photo during the burst, nothing at idle.
- **Time to market** — a working index/search in an afternoon vs weeks of infra.
- **Rekognition compatibility** — if you're already on AWS's API, a compatible hosted API is a smaller move than standing up your own stack.

> The honest test: multiply your realistic monthly photo volume by the per-photo API rate, then compare it to a fully-loaded self-hosted cost including idle compute, redundancy, and a fraction of an engineer's salary. Self-hosting rarely wins until volume is both high and steady.

Compare a per-photo API price against your fleet estimate. [See SightRadar pricing](/pricing)

## Frequently asked questions

### Is self-hosted face recognition cheaper than a hosted API?

Only at very high, steady volume with in-house ML ops. Open-source models are free to license but you pay for GPUs or CPU inference running 24/7, redundancy, a vector database, monitoring, model updates, and engineering time. For early, spiky, or small-team products, a hosted API with a per-photo price and no ops burden is usually cheaper in total cost of ownership.

### What are the hidden costs of running open-source face recognition?

Compute sized for peak but idle most of the time, multi-AZ redundancy, operating and scaling a vector database, re-benchmarking and re-indexing on model updates, monitoring and on-call, and the ML/infra staffing to run it all. These dwarf the (zero) model-license cost.

### When should I self-host instead of using an API?

Choose self-hosting when volume is very high and steady, when a hard data-residency or air-gap rule forbids sending images to a third party, when you already have in-house ML ops, or when you need deep model customization. Otherwise a hosted API is faster and cheaper overall.

---

SightRadar is an AWS Rekognition–compatible face recognition API — index, search, compare, and detect faces at per-photo rates below AWS Rekognition (batch and search about half its Group-1 reference), self-serve with no quota tickets. Docs: https://sightradar.com/docs · Pricing: https://sightradar.com/pricing · Start: https://sightradar.com/login