{
  "title": "GPU lifecycle: workload economics",
  "as_of": "2026-11-13",
  "timezone": "Europe/London",
  "page": "/ai/analysts/cases/gpu-lifecycle-workloads-20261113.html",
  "classification": "editorial analysis with synthetic sensitivity model",
  "inherited_material": {
    "title": "GPU Lifecycle v1",
    "date": "2026-11-09",
    "status": "Illustrative model; none of 46 numerical inputs verified; prior NPV, resale and token-cost outputs not adopted as market facts"
  },
  "sources": [
    {
      "id": "S1",
      "title": "MLCommons Inference Datacenter",
      "url": "https://mlcommons.org/benchmarks/inference-datacenter/",
      "checked_on": "2026-11-13",
      "scope": "Quality/scenario normalization; no extracted system performance ranking",
      "evidence_type": "primary_public"
    },
    {
      "id": "S2",
      "title": "NVIDIA T4",
      "url": "https://www.nvidia.com/en-us/data-center/tesla-t4/",
      "checked_on": "2026-11-13",
      "scope": "16 GB capacity and 70 W GPU specification; no whole-system power or claimed speedups used",
      "evidence_type": "primary_public"
    },
    {
      "id": "S3",
      "title": "NVIDIA video support matrix",
      "url": "https://developer.nvidia.com/video-encode-decode-support-matrix",
      "checked_on": "2026-11-13",
      "scope": "T4 H.264/HEVC and no AV1 encode; A100 no NVENC",
      "evidence_type": "primary_public"
    },
    {
      "id": "S4",
      "title": "NVIDIA A100",
      "url": "https://www.nvidia.com/en-gb/data-center/a100/",
      "checked_on": "2026-11-13",
      "scope": "40/80 GB alternatives; no derived measured throughput",
      "evidence_type": "primary_public"
    },
    {
      "id": "S5",
      "title": "CUDA Toolkit 13.0",
      "url": "https://developer.nvidia.com/blog/whats-new-and-important-in-cuda-toolkit-13-0/",
      "checked_on": "2026-11-13",
      "scope": "Pre-SM7.5 offline compilation support removed; existing applications are a separate matter",
      "evidence_type": "primary_public"
    },
    {
      "id": "S6",
      "title": "vLLM v0.17.1 quantization",
      "url": "https://docs.vllm.ai/en/v0.17.1/features/quantization/",
      "checked_on": "2026-11-13",
      "scope": "Versioned hardware/kernel compatibility; no universal quantization speed claim",
      "evidence_type": "primary_public"
    },
    {
      "id": "S7",
      "title": "QLoRA",
      "url": "https://arxiv.org/abs/2305.14314",
      "checked_on": "2026-11-13",
      "scope": "Capacity feasibility in tested settings; no inference about lifecycle ROI or universal quality parity",
      "evidence_type": "primary_public"
    }
  ],
  "model": {
    "equations": [
      "Q = 3600 * qualified_rate * effective_utilization",
      "unit_cost = complete_hourly_cost / Q",
      "old_hourly_cost_ceiling = new_hourly_cost * Q_old / Q_new"
    ],
    "input_units": {
      "cost": "USD/calendar hour",
      "rate": "matched qualified units/production second",
      "utilization": "percent of calendar time"
    },
    "default_synthetic_inputs": {
      "oldCost": 0.8,
      "newCost": 2.4,
      "oldRate": 1000,
      "newRate": 2000,
      "oldUtil": 60,
      "newUtil": 60
    },
    "expected_default_outputs": {
      "oldCostPerMillion": 0.37037037037037035,
      "newCostPerMillion": 0.5555555555555556,
      "oldHourlyCostCeiling": 1.2,
      "oldUtilizationFloorPercent": 40
    },
    "limits": [
      "No measured benchmark results or live prices",
      "No automatic power-capacity, demand elasticity, failure hazard or NPV calculation",
      "Feasibility gates must be established outside calculator",
      "Costs held fixed per comparison; update them when utilization changes"
    ]
  },
  "corpus_changes": {
    "real_questions": 0,
    "analyst_profiles": 0,
    "thought_cards": 0
  },
  "next_evidence": [
    "Matched workloads and quality/SLO measurements",
    "Realized rental and resale transactions",
    "Whole-system power and maintenance cohorts"
  ]
}