{
  "id": "on-prem",
  "date": "2026-09-09",
  "scope": "Derived from published inputs on 9 September 2026. Nothing here is measured by Zero One Labs.",
  "inputs": {
    "workload": {
      "name": "AgentX v1.0, replayed Claude Code sessions",
      "source": "https://inferencex.semianalysis.com/agentx",
      "inputTokens": 142000,
      "outputTokens": 444,
      "cachedShare": {
        "full": 0.98,
        "limited": 0.95
      },
      "note": "393 sessions, median 142k input and 444 output tokens a request, 44% with subagents. Input is 98% cache hits in the full set and 95% in the 256k variant. Which variant each run replayed is not printed on the run page.",
      "retrievedAt": "2026-09-09"
    },
    "throughput": {
      "gpu": "NVIDIA B200",
      "workload": "AgentX agentic coding workload, replayed Claude Code sessions",
      "operatingPoint": 100,
      "tco": {
        "hyperscalerVolume": 1.73,
        "retail": 3.7
      },
      "retrievedAt": "2026-09-09",
      "models": [
        {
          "id": "glm",
          "name": "GLM-5.3",
          "maker": "Z.ai",
          "parameters": "744B",
          "source": "https://inferencex.semianalysis.com/run/glm-5-3-on-b200",
          "published": "2026-09-02",
          "latestRun": "2026-09-02",
          "engine": "SGLang",
          "precision": "FP4",
          "nodes": 1,
          "configNote": "5 configurations, one run, TP8 on eight physical chips, concurrency 4 to 16. The sweep stopped at 16 clients, so the 30, 50 and 75 rungs are unmeasured.",
          "ladder": [
            {
              "perUser": 100,
              "perGpu": 10968
            },
            {
              "perUser": 150,
              "perGpu": 8054
            },
            {
              "perUser": 200,
              "perGpu": 6809
            }
          ]
        },
        {
          "id": "k3",
          "name": "Kimi K3",
          "maker": "Moonshot AI",
          "parameters": "2.8T",
          "source": "https://inferencex.semianalysis.com/run/kimi-k3-on-b200",
          "published": "2026-07-29",
          "latestRun": "2026-08-28",
          "engine": "Dynamo vLLM",
          "precision": "FP4",
          "nodes": 2,
          "configNote": "13 configurations over two runs, TP8 across sixteen physical chips, concurrency 4 to 96. Moonshot's own recommendation is a supernode of 64 or more accelerators. A DGX B200 carries 1,440 GB, and 2.8T parameters in MXFP4 do not fit one.",
          "ladder": [
            {
              "perUser": 30,
              "perGpu": 6783
            },
            {
              "perUser": 50,
              "perGpu": 4561
            },
            {
              "perUser": 75,
              "perGpu": 3022
            },
            {
              "perUser": 100,
              "perGpu": 2147
            },
            {
              "perUser": 150,
              "perGpu": 1366
            },
            {
              "perUser": 200,
              "perGpu": 1007
            }
          ]
        }
      ]
    },
    "node": {
      "gpus": 8,
      "serverUsd": 400328.5,
      "serverSource": "https://www.exxactcorp.com/Exxact-TS4-169219634-E169219634",
      "serverNote": "Exxact TensorEX 10U HGX B200, listed starting price, live crawl 2026-09-08 (a cached copy read $404,618.50 earlier the same day). NVIDIA publishes no list price for a B200 system.",
      "maxPowerKw": 14.3,
      "memoryGb": 1440,
      "powerSource": "https://www.nvidia.com/en-us/data-center/dgx-b200/",
      "powerNote": "DGX B200 system power usage, ~14.3 kW max, and 1,440 GB of GPU memory. Used as the rated draw and memory of an eight-B200 node.",
      "depreciationMonths": 36,
      "residualUsd": 0,
      "hoursPerMonth": 730
    },
    "electricity": {
      "usIndustrialCentsPerKwh": 9.17,
      "usSource": "https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_5_6_a",
      "usNote": "EIA Electric Power Monthly, table 5.6.A, U.S. Total, industrial, June 2026, released 2026-08-26.",
      "norwayServicesOrePerKwh": 89,
      "norwaySource": "https://www.ssb.no/en/energi-og-industri/energi/statistikk/elektrisitetspriser",
      "norwayNote": "SSB, electricity price for services excluding taxes, 2nd quarter 2026, updated 2026-08-17."
    },
    "rental": [
      {
        "id": "aws",
        "name": "AWS capacity block",
        "kind": "list",
        "usdPerGpuHour": 12.355,
        "source": "https://aws.amazon.com/ec2/capacityblocks/pricing/",
        "note": "p6-b200.48xlarge, US East, $98.84 an hour for eight",
        "retrievedAt": "2026-09-08"
      },
      {
        "id": "runpod",
        "name": "RunPod on-demand, secure cloud",
        "kind": "list",
        "usdPerGpuHour": 6.79,
        "source": "https://www.runpod.io/pricing",
        "note": "Community cloud is $5.98",
        "retrievedAt": "2026-09-08"
      },
      {
        "id": "lambda",
        "name": "Lambda on-demand",
        "kind": "list",
        "usdPerGpuHour": 6.69,
        "source": "https://lambda.ai/pricing",
        "note": "Eight-GPU instance",
        "retrievedAt": "2026-09-08"
      }
    ],
    "api": {
      "glm": {
        "name": "Z.ai GLM-5.3 API",
        "source": "https://docs.z.ai/guides/overview/pricing",
        "retrievedAt": "2026-09-09",
        "inputUsdPerM": 1.4,
        "cacheHitInputUsdPerM": 0.26,
        "outputUsdPerM": 4.4
      },
      "k3": {
        "name": "Kimi K3 API",
        "source": "https://platform.kimi.ai/docs/pricing/chat-k3.md",
        "retrievedAt": "2026-09-09",
        "inputUsdPerM": 3,
        "cacheHitInputUsdPerM": 0.3,
        "outputUsdPerM": 15
      }
    },
    "utilizationObservations": [
      {
        "id": "castAverage",
        "name": "Cast AI, average across tens of thousands of clusters",
        "utilization": 0.05,
        "source": "https://cast.ai/press-release/2026-state-of-kubernetes-optimization-report/",
        "date": "2026-04-21"
      },
      {
        "id": "castBest",
        "name": "Cast AI, best case, a 136-node H200 inference fleet",
        "utilization": 0.49,
        "source": "https://cast.ai/blog/kubernetes-gpu-optimization/",
        "date": "2026-07-03"
      },
      {
        "id": "ventureBeat",
        "name": "VentureBeat Research, 83% of enterprises running their own GPUs report 50% or less",
        "utilization": 0.5,
        "share": 0.83,
        "source": "https://venturebeat.com/orchestration/wall-street-is-debating-the-ai-buildout-enterprises-just-answered-86-say-their-gpus-run-at-half-capacity-or-less",
        "date": "2026-07-10",
        "note": "573 technical leaders surveyed in June 2026. The headline first said 86% and was corrected to about 83% on 14 July. Self-selected sample, read directionally."
      }
    ],
    "latency": {
      "source": "https://modelstats.ai/blog/llm-api-latency-spikes-what-average-ttft-doesnt-tell-you",
      "published": "2026-04-07",
      "retrievedAt": "2026-09-08",
      "window": "24 hours, one probe every 10 minutes, one US datacenter",
      "rows": [
        {
          "model": "Gemini 2.5 Flash Lite",
          "p50": 311,
          "p95": 1142,
          "p99": 3080,
          "max": 3334
        },
        {
          "model": "Gemini 2.5 Flash",
          "p50": 477,
          "p95": 776,
          "p99": 1309,
          "max": 1879
        },
        {
          "model": "GPT-4o Mini",
          "p50": 536,
          "p95": 844,
          "p99": 4461,
          "max": 5946
        },
        {
          "model": "Claude Haiku 4.5",
          "p50": 541,
          "p95": 1591,
          "p99": 4117,
          "max": 11611
        },
        {
          "model": "GPT-4o",
          "p50": 639,
          "p95": 1144,
          "p99": 4517,
          "max": 7010
        },
        {
          "model": "GPT-4.1 Mini",
          "p50": 661,
          "p95": 1087,
          "p99": 2169,
          "max": 4279
        },
        {
          "model": "GPT-4.1",
          "p50": 726,
          "p95": 1220,
          "p99": 1946,
          "max": 4813
        },
        {
          "model": "Claude Sonnet 4.6",
          "p50": 941,
          "p95": 2939,
          "p99": 4787,
          "max": 5453
        },
        {
          "model": "DeepSeek V3",
          "p50": 1479,
          "p95": 1926,
          "p99": 2492,
          "max": 3105
        },
        {
          "model": "Claude Opus 4.6",
          "p50": 1626,
          "p95": 2420,
          "p99": 3797,
          "max": 5291
        }
      ]
    }
  },
  "derived": {
    "operatingPoint": 100,
    "nodeUsdPerMonth": 12077,
    "depreciationUsdPerMonth": 11120,
    "kwhPerMonth": 10439,
    "powerUsdPerMonthUs": 957,
    "powerNokPerMonthNorway": 9291,
    "ownedUsdPerGpuHour": 2.07,
    "models": [
      {
        "id": "glm",
        "name": "GLM-5.3",
        "nodes": 1,
        "gpus": 8,
        "tokensPerSecondPerGpu": 10968,
        "tokensPerDayAtFull": 7581081600,
        "tokensPerMonthAtFull": 230591232000,
        "ownedUsdPerMonth": 12077,
        "ownedUsdPerMillionAtFull": 0.0524,
        "apiListUsdPerMillion": 1.409,
        "apiCachedUsdPerMillion": 0.273,
        "apiTracedUsdPerMillion": 0.296,
        "apiTracedLimitedUsdPerMillion": 0.33,
        "lambdaUsdPerMillion": 0.169,
        "crossovers": {
          "list": {
            "utilization": 0.0372,
            "tokensPerDay": 281738354,
            "usdPerMonth": 12077
          },
          "cached": {
            "utilization": 0.1919,
            "tokensPerDay": 1454971726,
            "usdPerMonth": 12077
          },
          "traced": {
            "utilization": 0.1772,
            "tokensPerDay": 1343110358,
            "usdPerMonth": 12077
          },
          "tracedLimited": {
            "utilization": 0.1588,
            "tokensPerDay": 1204234172,
            "usdPerMonth": 12077
          },
          "lambda": {
            "utilization": 0.3091,
            "tokensPerDay": 2343521702,
            "usdPerMonth": 12077
          }
        }
      },
      {
        "id": "k3",
        "name": "Kimi K3",
        "nodes": 2,
        "gpus": 16,
        "tokensPerSecondPerGpu": 2147,
        "tokensPerDayAtFull": 2968012800,
        "tokensPerMonthAtFull": 90277056000,
        "ownedUsdPerMonth": 24155,
        "ownedUsdPerMillionAtFull": 0.2676,
        "apiListUsdPerMillion": 3.037,
        "apiCachedUsdPerMillion": 0.346,
        "apiTracedUsdPerMillion": 0.4,
        "apiTracedLimitedUsdPerMillion": 0.48,
        "lambdaUsdPerMillion": 0.866,
        "crossovers": {
          "list": {
            "utilization": 0.0881,
            "tokensPerDay": 261452359,
            "usdPerMonth": 24155
          },
          "cached": {
            "utilization": 0.7737,
            "tokensPerDay": 2296386070,
            "usdPerMonth": 24155
          },
          "traced": {
            "utilization": 0.6695,
            "tokensPerDay": 1987070998,
            "usdPerMonth": 24155
          },
          "tracedLimited": {
            "utilization": 0.557,
            "tokensPerDay": 1653075812,
            "usdPerMonth": 24155
          },
          "lambda": {
            "utilization": 0.3091,
            "tokensPerDay": 917494729,
            "usdPerMonth": 24155
          }
        }
      }
    ]
  },
  "ladder": [
    {
      "model": "glm",
      "perUser": 100,
      "perGpu": 10968,
      "engine": "SGLang",
      "precision": "FP4",
      "usdPerMillion": 0.044
    },
    {
      "model": "glm",
      "perUser": 150,
      "perGpu": 8054,
      "engine": "SGLang",
      "precision": "FP4",
      "usdPerMillion": 0.06
    },
    {
      "model": "glm",
      "perUser": 200,
      "perGpu": 6809,
      "engine": "SGLang",
      "precision": "FP4",
      "usdPerMillion": 0.071
    },
    {
      "model": "k3",
      "perUser": 30,
      "perGpu": 6783,
      "engine": "Dynamo vLLM",
      "precision": "FP4",
      "usdPerMillion": 0.071
    },
    {
      "model": "k3",
      "perUser": 50,
      "perGpu": 4561,
      "engine": "Dynamo vLLM",
      "precision": "FP4",
      "usdPerMillion": 0.105
    },
    {
      "model": "k3",
      "perUser": 75,
      "perGpu": 3022,
      "engine": "Dynamo vLLM",
      "precision": "FP4",
      "usdPerMillion": 0.159
    },
    {
      "model": "k3",
      "perUser": 100,
      "perGpu": 2147,
      "engine": "Dynamo vLLM",
      "precision": "FP4",
      "usdPerMillion": 0.224
    },
    {
      "model": "k3",
      "perUser": 150,
      "perGpu": 1366,
      "engine": "Dynamo vLLM",
      "precision": "FP4",
      "usdPerMillion": 0.352
    },
    {
      "model": "k3",
      "perUser": 200,
      "perGpu": 1007,
      "engine": "Dynamo vLLM",
      "precision": "FP4",
      "usdPerMillion": 0.477
    }
  ],
  "prices": [
    {
      "id": "aws",
      "name": "AWS capacity block",
      "kind": "list",
      "usdPerGpuHour": 12.355,
      "source": "https://aws.amazon.com/ec2/capacityblocks/pricing/",
      "note": "p6-b200.48xlarge, US East, $98.84 an hour for eight",
      "retrievedAt": "2026-09-08"
    },
    {
      "id": "runpod",
      "name": "RunPod on-demand, secure cloud",
      "kind": "list",
      "usdPerGpuHour": 6.79,
      "source": "https://www.runpod.io/pricing",
      "note": "Community cloud is $5.98",
      "retrievedAt": "2026-09-08"
    },
    {
      "id": "lambda",
      "name": "Lambda on-demand",
      "kind": "list",
      "usdPerGpuHour": 6.69,
      "source": "https://lambda.ai/pricing",
      "note": "Eight-GPU instance",
      "retrievedAt": "2026-09-08"
    },
    {
      "id": "owned",
      "name": "Owned, small buyer, server and power only",
      "kind": "cost",
      "usdPerGpuHour": 2.07,
      "source": "https://www.exxactcorp.com/Exxact-TS4-169219634-E169219634",
      "note": "Derived below: the listed server over 36 months plus power at the US industrial rate. No colocation, staff, spares or networking.",
      "retrievedAt": "2026-09-09"
    },
    {
      "id": "hyperscaler",
      "name": "Owned at hyperscaler volume, SemiAnalysis estimate",
      "kind": "cost",
      "usdPerGpuHour": 1.73,
      "source": "https://inferencex.semianalysis.com/run/glm-5-3-on-b200",
      "note": "SemiAnalysis AI Cloud TCO model, 'Owning at Large Hyperscaler Volume'",
      "retrievedAt": "2026-09-09"
    }
  ],
  "crossings": [
    {
      "model": "glm",
      "usdPerMillion": 0.273,
      "ownedUsdPerMonth": 12077,
      "tokensPerDayAtFull": 7581081600,
      "crossover": {
        "utilization": 0.1919,
        "tokensPerDay": 1454971726,
        "usdPerMonth": 12077
      },
      "number": 1
    },
    {
      "model": "k3",
      "usdPerMillion": 0.346,
      "ownedUsdPerMonth": 24155,
      "tokensPerDayAtFull": 2968012800,
      "crossover": {
        "utilization": 0.7737,
        "tokensPerDay": 2296386070,
        "usdPerMonth": 24155
      },
      "number": 2
    }
  ],
  "utilizationBars": [
    {
      "id": "castAverage",
      "utilization": 0.05,
      "usdPerMillion": 1.048,
      "anchor": "Cast AI average"
    },
    {
      "id": "crossApi",
      "utilization": 0.1919,
      "usdPerMillion": 0.273,
      "anchor": "breaks even with the cached API"
    },
    {
      "id": "crossLambda",
      "utilization": 0.3091,
      "usdPerMillion": 0.169,
      "anchor": "breaks even with renting"
    },
    {
      "id": "half",
      "utilization": 0.5,
      "usdPerMillion": 0.105,
      "anchor": "the surveys' ceiling"
    },
    {
      "id": "full",
      "utilization": 1,
      "usdPerMillion": 0.052,
      "anchor": "never sleeps"
    }
  ],
  "latency": [
    {
      "model": "Gemini 2.5 Flash Lite",
      "p50": 311,
      "p95": 1142,
      "p99": 3080,
      "max": 3334,
      "p99OverP50": 9.9,
      "maxOverP50": 10.72
    },
    {
      "model": "Gemini 2.5 Flash",
      "p50": 477,
      "p95": 776,
      "p99": 1309,
      "max": 1879,
      "p99OverP50": 2.74,
      "maxOverP50": 3.94
    },
    {
      "model": "GPT-4o Mini",
      "p50": 536,
      "p95": 844,
      "p99": 4461,
      "max": 5946,
      "p99OverP50": 8.32,
      "maxOverP50": 11.09
    },
    {
      "model": "Claude Haiku 4.5",
      "p50": 541,
      "p95": 1591,
      "p99": 4117,
      "max": 11611,
      "p99OverP50": 7.61,
      "maxOverP50": 21.46
    },
    {
      "model": "GPT-4o",
      "p50": 639,
      "p95": 1144,
      "p99": 4517,
      "max": 7010,
      "p99OverP50": 7.07,
      "maxOverP50": 10.97
    },
    {
      "model": "GPT-4.1 Mini",
      "p50": 661,
      "p95": 1087,
      "p99": 2169,
      "max": 4279,
      "p99OverP50": 3.28,
      "maxOverP50": 6.47
    },
    {
      "model": "GPT-4.1",
      "p50": 726,
      "p95": 1220,
      "p99": 1946,
      "max": 4813,
      "p99OverP50": 2.68,
      "maxOverP50": 6.63
    },
    {
      "model": "Claude Sonnet 4.6",
      "p50": 941,
      "p95": 2939,
      "p99": 4787,
      "max": 5453,
      "p99OverP50": 5.09,
      "maxOverP50": 5.79
    },
    {
      "model": "DeepSeek V3",
      "p50": 1479,
      "p95": 1926,
      "p99": 2492,
      "max": 3105,
      "p99OverP50": 1.68,
      "maxOverP50": 2.1
    },
    {
      "model": "Claude Opus 4.6",
      "p50": 1626,
      "p95": 2420,
      "p99": 3797,
      "max": 5291,
      "p99OverP50": 2.34,
      "maxOverP50": 3.25
    }
  ]
}