Table of Contents
DeepSeek's models — particularly the V2 16B Lite and V3 671B MoE — have an unusual cost profile. The DeepSeek-hosted API is genuinely cheap (often the cheapest tier of any frontier-class model API), which makes the self-hosting math less obvious than for, say, Llama 3 70B vs Together AI. This page is the proper comparison.
For DeepSeek-V2 16B Lite: self-host on an RTX 5090 at £0.13/1M tokens beats DeepSeek's own API (£0.18/1M output) at any meaningful volume. For DeepSeek-V3 671B: the official DeepSeek API at £0.85/1M output is essentially impossible to beat — V3 self-hosting requires 8× H100 hardware at >£15K/mo, only worth it for data-residency reasons.
Which DeepSeek? V2 16B vs V3 671B
- DeepSeek-V2 16B Lite — 16B params, 2.4B active per token. Single-GPU friendly. The most-deployed DeepSeek variant.
- DeepSeek-V2 236B — 236B total, 21B active. Multi-GPU only. Strong on reasoning.
- DeepSeek-V3 671B — frontier-class MoE. 37B active per token. Multi-node only.
- DeepSeek-Coder-V2 — code-specialised V2 Lite. Same hardware profile.
For self-hosting purposes the practical target is V2 16B Lite. V3 is too big for any single server we rent.
Self-hosting cost per million tokens
Methodology matches our other cost pages: 60% utilisation, vLLM 0.6.3, 50-thread driver.
DeepSeek-V2 16B Lite:
| GPU | Monthly | tok/s | Precision | Cost per 1M |
|---|---|---|---|---|
| RTX 3090 | £159 | 410 | AWQ-INT4 | £0.28 |
| RTX 4090 | £289 | 590 | AWQ-INT4 | £0.30 |
| RTX 5090 | £399 | 780 | AWQ-INT4 | £0.30 |
| RTX 5090 | £399 | 950 | FP8 | £0.24 |
| RTX 6000 Pro | £899 | 1,420 | FP16 | £0.50 |
DeepSeek-V3 671B: needs 8× H100 / H200 minimum. We can spec this as a custom build, lead time 4–6 weeks, monthly cost >£15K. Cost per 1M tokens at 60% utilisation works out to ~£3.50 — meaningfully more than the official API.
DeepSeek API pricing
DeepSeek's official API pricing (as of late 2026):
| Model | Input per 1M | Output per 1M | Notes |
|---|---|---|---|
| deepseek-chat (V2) | £0.10 | £0.18 | Quietly raised from $0.14/1M earlier |
| deepseek-coder | £0.10 | £0.18 | Same pricing as chat |
| deepseek-reasoner (V3 / R1-class) | £0.45 | £0.85 | Frontier reasoning |
DeepSeek API pricing is among the cheapest of any frontier-class model API.
Break-even token volume
For DeepSeek-V2 16B Lite vs the official API:
- RTX 5090 self-hosted: £399/mo. API at £0.20/1M = break-even at ~2.0B output tokens/month.
- Below 2B tokens, API wins.
- Above 2B tokens, self-hosting wins.
For V3 671B vs the official API:
- 8× H100 self-hosted: £15,000+/mo. API at £0.85/1M = break-even at ~17.6B output tokens/month.
- That is an enormous workload — only the largest enterprise deployments cross it.
- Below 17B tokens (the vast majority of deployments), API wins on cost.
Non-cost reasons to self-host DeepSeek
For DeepSeek specifically, cost is rarely the actual driver. The real reasons we see customers self-hosting DeepSeek:
- Data residency — the official DeepSeek API is China-hosted. UK and EU customers in regulated sectors cannot use it. See private AI hosting.
- No prompt logging — DeepSeek's privacy policy is more permissive than most Western APIs. Self-hosting eliminates that question.
- Latency — Chinese egress + your application location means typical TTFT of 800–1500 ms. Self-hosted in your region: 200–400 ms.
- API stability — DeepSeek's official API has had several multi-hour outages in 2026. Self-hosted is your uptime.
- Custom fine-tunes — DeepSeek does not offer fine-tuning on their hosted API. Self-hosting is the only path.
Verdict
- If cost is the only factor and your volume is <2B tokens/month: use DeepSeek's official API.
- If you need data residency / GDPR / privacy: self-host on a 5090 (V2) regardless of volume.
- If you need V3 and you are not Anthropic-scale: use the official API. V3 self-hosting is rarely the right choice.
- If you are DeepSeek-Coder-specifically: self-hosting is the default for any team large enough to justify a dedicated server (~10+ devs).
Bottom line
DeepSeek is the rare model where the official API is genuinely competitive on cost. Self-hosting wins on privacy and latency, not pounds-per-token. Pick self-hosted DeepSeek-V2 16B if you need data residency or a custom fine-tune; otherwise the API is fine.
For deployment specifics on V2 see can RTX 5090 run DeepSeek?; for the broader picture our best GPU for DeepSeek guide.
