Cheapest Cloud GPU to Host DeepSeek V2 / V2.5
DeepSeek V2 uses a Mixture-of-Experts architecture with 236B total parameters but only 21B active per token. This makes it extremely efficient for inference — the active parameter count is comparable to a 21B dense model, but with much higher quality. FP8 quantization cuts the 472 GB FP16 footprint in half to 236 GB, fitting on 2x H100 SXM5 GPUs with tensor parallelism.
VRAM Sizing Matrix
Memory consumption at different quantization levels for DeepSeek V2 / V2.5. Includes model weights only — KV-cache adds 2-32 GB depending on context length and batch size.
| Precision | VRAM Required | Fits On | Quality Loss |
|---|---|---|---|
| FP16 / BF16 | 472 GB | Multi-GPU cluster | None (full precision) |
| FP8 | 236 GB | Multi-GPU or single large GPU | <1% perplexity delta |
| INT4 / AWQ | 128 GB | Single consumer GPU | 1-3% quality degradation |
Hardware Bottleneck Advice
KV-cache expansion, tensor parallelism constraints, and production considerations for DeepSeek V2 / V2.5.
KV-Cache & Context Scaling
21B active parameters per token. At 32k context, KV-cache adds ~8 GB. With TP=2, each GPU holds 118 GB weights + 4 GB KV-cache = 122 GB per GPU.
Tensor Parallelism
TP=2 mandatory. Each GPU holds 118 GB of 236B FP8 weights. NVLink required between the 2 GPUs for MoE expert routing.
Live Cloud Pricing for H100 SXM5
Filtered to H100 SXM5-class hardware — the recommended GPU for hosting DeepSeek V2 / V2.5.
| Provider | GPU & VRAM | Interconnect | Spot Rate | On-Demand | Monthly | Status | Action | |
|---|---|---|---|---|---|---|---|---|
Community | NVLink 4.0 (900 GB/s) | $1.89 / hr | $4.72 / hr | $1,157 / mo | Instant | |||
Bare Metal | NVLink 4.0 (900 GB/s) | $2.29 / hr | $5.73 / hr | $1,401 / mo | Instant | |||
Dedicated | NVLink 4.0 (900 GB/s) | $2.99 / hr | $7.48 / hr | $1,830 / mo | Instant | |||
Cloud | NVLink 4.0 (900 GB/s) | $3.49 / hr | $8.73 / hr | $2,136 / mo | Instant |
Production Runbook
Ready-to-deploy vLLM / SGLang launch command for DeepSeek V2 / V2.5.
vllm serve deepseek-ai/DeepSeek-V2.5 --tensor-parallel-size 2 --max-model-len 32768 --dtype autoRelated hosting guides