{
 "title": "Revenue per GPU-hour for an inference provider: DeepSeek-V4.1-Flash on 4x H200",
 "measured": "2026-10-27",
 "platform": "RunPod Secure Cloud, 4x H200 141 GB, $5.29 per GPU-hour",
 "engine": "vLLM 0.32.0, TP4, --kv-cache-dtype fp8 --tokenizer-mode deepseek_v41 --engram-config {\"cpu_offload\":true} --gpu-memory-utilization 0.95 (vLLM recipe for H200)",
 "model": "deepseek-ai/DeepSeek-V4.1-Flash (FP8 weights, FP4 experts)",
 "workload": "chat-mix-v1, random-token prompts (no cache hits), median 1,024 input / 256 output tokens",
 "results": {
  "tpot50": {
   "slo": "p99 TTFT <= 2 s, p99 TPOT <= 50 ms",
   "rate_rps": 5.091,
   "input_tok_s": 8177,
   "output_tok_s": 2047,
   "node_power_kw": 2.38,
   "revenue_usd_per_gpu_hour": {
    "low ($0.14/$0.42)": 1.8,
    "median endpoint ($0.21/$0.87)": 3.15,
    "common ($0.30/$1.20, 12 of 38 endpoints)": 4.42,
    "highest ($0.60/$2.40)": 8.84
   }
  },
  "tpot10": {
   "slo": "p99 TTFT <= 2 s, p99 TPOT <= 10 ms (~100 tok/s per user)",
   "rate_rps": 0.328,
   "input_tok_s": 461,
   "output_tok_s": 108,
   "node_power_kw": 0.92,
   "revenue_usd_per_gpu_hour": {
    "low ($0.14/$0.42)": 0.1,
    "median endpoint ($0.21/$0.87)": 0.17,
    "common ($0.30/$1.20, 12 of 38 endpoints)": 0.24,
    "highest ($0.60/$2.40)": 0.48
   }
  }
 },
 "openrouter_prices": {
  "snapshot": "2026-10-27T04:38Z",
  "model": "deepseek/deepseek-v4.1-flash",
  "endpoints": 38,
  "providers": 33,
  "output_usd_per_m": {
   "min": 0.18,
   "p25": 0.6,
   "median": 0.87,
   "p75": 1.2,
   "max": 2.4
  },
  "input_median_usd_per_m": 0.21,
  "endpoints_at_0.30_1.20": 12,
  "median_output_fp4_endpoints": 0.25,
  "median_output_fp8_endpoints": 0.75
 },
 "formula": "revenue per GPU-hour = (input tok/s x input price + output tok/s x output price) / 1e6 x 3600 / 4 GPUs; full load; no cache-hit discounts"
}