Table of Contents
1M docs at avg 500 tokens via BGE-large: 5060 Ti ~25 minutes, 4090 ~12 minutes. Pro-rated GPU time is pennies; the real comparison is against hosted APIs at ~£12 (OpenAI) to ~£40 (Cohere) for the same job. Self-hosted ingest is ~80x cheaper at scale.
Compare
| Method | £ for 1M docs |
|---|---|
| Self-hosted 5060 Ti | ~£0.15 |
| Self-hosted 4090 | ~£0.13 |
| OpenAI text-embedding-3 | ~£12 |
| Cohere Embed v3 | ~£40 |
How derived
Self-hosted = pro-rata of monthly tier × minutes used. 5060 Ti at £119/mo and ~600 embeddings/sec finishes 1M docs in ~25 minutes — roughly £0.10 of compute, plus a small share of fixed costs. Reasoned estimates from observed BGE-large batch-32 throughput.
Ongoing vs one-off
One-off ingest costs are tiny either way. The real cost shows up if you re-embed often: weekly schema change, model swap, or large continuous backfill. At ~10M docs/month sustained, OpenAI bills £120/month; the same on a £119 5060 Ti tier is essentially free of marginal cost.
Verdict
Self-hosted embedding ingest is ~80x cheaper than hosted at scale. Even for a single backfill the payback is immediate; for ongoing pipelines it is overwhelming.
Bottom line
~80× cheaper. See embedding migration.
