Embedding Generation: Cost at 1B Tokens/Month
What does it cost to run embedding generation at 1B tokens/month? Self-hosted dedicated GPU vs API provider pricing.
Monthly Cost Comparison at 1B tokens/month
| Provider | Monthly Cost | Pricing Model | vs GigaGPU |
|---|---|---|---|
| GigaGPU (RTX 3090) | £89/mo | Fixed | — |
| OpenAI text-embedding-3-small | £20/mo | Per-tokens | API is cheaper at this volume |
| Cohere Embed v3 | £100/mo | Per-tokens | 11% cheaper with GigaGPU |
| Voyage AI | £120/mo | Per-tokens | 26% cheaper with GigaGPU |
One Billion Tokens: The Crossover Point for Premium Embedding Providers
At one billion tokens per month, a clear split emerges in the embedding market. OpenAI text-embedding-3-small remains remarkably cheap at £20/month — well below a dedicated GPU’s £89 fixed cost. But higher-quality embedding providers tell a different story: Cohere Embed v3 hits £100/month and Voyage AI reaches £120/month.
A dedicated RTX 3090 at £89/month now undercuts Cohere by 11% and Voyage AI by 26%. And unlike APIs, the GPU cost stays flat whether you process 1B or 2B tokens — your marginal cost is zero. For teams that need higher-quality embeddings than OpenAI’s budget model but want to avoid Cohere or Voyage pricing, self-hosted models like BGE-large or E5-large deliver comparable quality.
Annual savings potential: Up to £372 per year compared to the most expensive API option, assuming consistent 1B tokens/month usage.
The Case for Self-Hosted Embeddings at Scale
- Growth protection: At 1B tokens, you are one product launch away from 2B or 5B tokens. Self-hosted cost stays at £89 while API costs scale linearly with every additional token.
- Model customisation: Fine-tune embedding models on your domain data. Retrieval accuracy on specialised content often improves 15-30% with domain-adapted embeddings.
- Data governance at scale: One billion tokens represents a massive corpus of text being processed externally via APIs. Self-hosting eliminates that data exposure.
- Pipeline efficiency: Run embeddings as part of a larger GPU workload — the same RTX 3090 can handle embedding generation, LLM inference, and reranking.
When APIs Still Make Financial Sense
- OpenAI-tier pricing: At £20/month for 1B tokens, OpenAI text-embedding-3-small is genuinely hard to beat on pure cost — if the embedding quality meets your retrieval needs.
- Operational simplicity: An API call is one line of code. Self-hosted embeddings require server management, model updates, and capacity monitoring.
- Infrequent re-embedding: If you embed a document corpus once and rarely update it, API costs are a one-time expense rather than recurring.
Recommended Hardware
The RTX 3090 at £89/month provides 24 GB VRAM for running high-quality embedding models like BGE-large or E5-large, with throughput for 1B tokens/month and 20-30% headroom. Pre-configured with CUDA, Docker, and inference frameworks.
Self-Host Premium Embeddings for £89/Month
Get Cohere-quality embeddings without Cohere pricing. Run any open-source embedding model on dedicated GPU hardware with no per-token fees.
