{
 "about": "Base case: GB200 NVL72 capacity, per MW of IT load, serving DeepSeek R1 (FP4) at an interactive SLO of 125 tokens/s per user, 1k input / 1k output. A SCENARIO built from cited inputs, not observed operator data. Each param: value, range, unit, evidence (spec | filing | independent | vendor | estimate | assumption), source.",
 "params": {
  "it_kw_per_gpu": {
   "value": 1.667,
   "range": [
    1.4,
    1.95
   ],
   "unit": "kW IT per GPU (system level)",
   "evidence": "spec (via search)",
   "source": "NVIDIA GB200 NVL72 rack ~120 kW / 72 GPUs, https://www.nvidia.com/en-us/data-center/gb200-nvl72/ ; InferenceX per-MW figure implies ~1.87 kW all-in"
  },
  "pue": {
   "value": 1.3,
   "range": [
    1.1,
    1.54
   ],
   "unit": "ratio",
   "evidence": "assumption bounded by survey",
   "source": "Uptime Institute Global Data Center Survey 2025: global average 1.54 (n=681); new liquid-cooled AI halls run lower (assumption)"
  },
  "power_load_factor": {
   "value": 0.8,
   "range": [
    0.6,
    1.0
   ],
   "unit": "average draw / nameplate",
   "evidence": "assumption",
   "source": "assumption; GPUs at partial load draw below TDP"
  },
  "electricity_usd_per_kwh": {
   "value": 0.0813,
   "range": [
    0.04,
    0.15
   ],
   "unit": "$/kWh",
   "evidence": "statistic (via search)",
   "source": "EIA US industrial average 2024: 8.13 c/kWh (EIA Electric Power Monthly Table 5.3, via search); Jan-Mar 2026 8.6-9.3 c"
  },
  "colo_usd_per_kw_month": {
   "value": 195.94,
   "range": [
    150,
    275
   ],
   "unit": "$/kW-month, excl. power",
   "evidence": "market survey (via search)",
   "source": "CBRE North America Data Center Trends H2 2025: 250-500 kW average $195.94/kW-month (The Register, 4 Mar 2026); Q1 2026 N. Virginia $190-235"
  },
  "capex_per_gpu": {
   "value": 54167,
   "range": [
    40000,
    70000
   ],
   "unit": "$ per GPU, all-in system incl. networking and storage",
   "evidence": "estimate (analyst)",
   "source": "SemiAnalysis estimate (20 Aug 2025): GB200 NVL72 ~$3.9m all-in incl. networking and storage / 72 GPUs; HSBC/Wolfe ~$3.0m rack only"
  },
  "useful_life_years": {
   "value": 6,
   "range": [
    3,
    6
   ],
   "unit": "years",
   "evidence": "filing (policy)",
   "source": "CoreWeave depreciates technology equipment over six years (10-K policy; see research/neocloud-gtm/data/neocloud-financials.csv)"
  },
  "cost_of_capital": {
   "value": 0.09,
   "range": [
    0.06,
    0.12
   ],
   "unit": "annual rate on average balance",
   "evidence": "assumption bounded by filings",
   "source": "see research/neocloud-gtm/data/neocloud-financials.csv for disclosed interest expense; assumption"
  },
  "opex_pct_of_capex": {
   "value": 0.05,
   "range": [
    0.03,
    0.08
   ],
   "unit": "share of capex per year",
   "evidence": "assumption",
   "source": "assumption: staff, support, software, spares, storage and network transit"
  },
  "utilisation": {
   "value": 0.7,
   "range": [
    0.3,
    1.0
   ],
   "unit": "share of hours serving billable load at SLO",
   "evidence": "assumption",
   "source": "assumption; no provider discloses inference utilisation"
  },
  "slo_tokens_per_s_per_gpu": {
   "value": 4130,
   "range": [
    941,
    8064
   ],
   "unit": "tokens/s per GPU at 125 tok/s/user",
   "evidence": "independent (via search)",
   "source": "SemiAnalysis InferenceX, DeepSeek R1 0528 FP4, 1k/1k, disaggregated Dynamo + TRT-LLM, GB200 NVL72, 22 May 2026. Whether this counts input+output or output only is unconfirmed; the model treats it as total tokens."
  },
  "price_usd_per_m_tokens": {
   "value": 0.75,
   "range": [
    0.1,
    1.0
   ],
   "unit": "$ per M tokens, 1:1 input:output blend",
   "evidence": "list price",
   "source": "DeepSeek reference price $0.30 in / $1.20 out per M (DeepSeek V4.1 Flash endpoints on Artificial Analysis, 4 Nov 2026), blended 1:1 to match the 1k/1k benchmark"
  },
  "gpu_hour_price": {
   "value": 4.0,
   "range": [
    3.79,
    8.5
   ],
   "unit": "$ per GPU-hour",
   "evidence": "scenario within observed range (via search)",
   "source": "B200 on-demand $3.79 (Lambda, conflicting) to $8.50 (Nebius from 1 Oct 2026); no public GB200 per-GPU price found"
  },
  "list_input_price": {
   "value": 0.3,
   "range": [
    0.15,
    1.4
   ],
   "unit": "$ per M input tokens",
   "evidence": "list price",
   "source": "Artificial Analysis, DeepSeek V4.1 Flash reference price, 4 Nov 2026"
  },
  "list_output_price": {
   "value": 1.2,
   "range": [
    0.6,
    4.4
   ],
   "unit": "$ per M output tokens",
   "evidence": "list price",
   "source": "Artificial Analysis, DeepSeek V4.1 Flash reference price, 4 Nov 2026"
  },
  "prefill_to_decode_speed": {
   "value": 4.0,
   "range": [
    2,
    10
   ],
   "unit": "input tokens per output token of GPU time",
   "evidence": "assumption anchored to list prices",
   "source": "set equal to the market's output/input list price ratio (1.20/0.30); real ratio depends on context length and batch"
  },
  "cached_input_cost_factor": {
   "value": 0.1,
   "range": [
    0.05,
    0.3
   ],
   "unit": "cost of a cached input token vs uncached",
   "evidence": "assumption",
   "source": "assumption; prefix-cache hits skip prefill compute"
  },
  "cached_price_factor": {
   "value": 0.25,
   "range": [
    0.1,
    0.5
   ],
   "unit": "price of cached input vs list",
   "evidence": "assumption",
   "source": "assumption in line with common cache-hit discounts; check per provider"
  },
  "resale_share": {
   "value": 0.5,
   "range": [
    0,
    1
   ],
   "unit": "share of freed capacity resold at base revenue",
   "evidence": "assumption",
   "source": "assumption for the demand-constrained case"
  },
  "model_layers": {
   "value": 80,
   "range": [
    80,
    80
   ],
   "unit": "layers",
   "evidence": "spec",
   "source": "Llama 3.1 70B config, meta-llama/llama-models sku_list.py; Llama 3 paper Table 3 (arXiv 2407.21783)"
  },
  "model_kv_heads": {
   "value": 8,
   "range": [
    8,
    8
   ],
   "unit": "KV heads",
   "evidence": "spec",
   "source": "as above"
  },
  "model_head_dim": {
   "value": 128,
   "range": [
    128,
    128
   ],
   "unit": "dims",
   "evidence": "spec",
   "source": "as above"
  },
  "kv_bytes": {
   "value": 2,
   "range": [
    1,
    2
   ],
   "unit": "bytes per value (BF16=2, FP8=1)",
   "evidence": "spec",
   "source": "as above"
  },
  "reserved_rev_per_it_mw": {
   "value": 20000000.0,
   "range": [
    12000000.0,
    40000000.0
   ],
   "unit": "$ revenue per MW of IT load per year, take-or-pay",
   "evidence": "filing (company metric)",
   "source": "Nebius Q2 2026 shareholder letter (6-K EX-99.2, 0001104659-26-094568): ACV per MW 2026 base $12M, Q2 deals >$20M, Q3 short-term capacity deals >$40M; IREN FY26 8-K EX-99.1 (0001878848-26-000051): recent 3-year contracts >$20m revenue per MW (IT), ~2 year payback"
  }
 }
}