{
  "meta": {
    "title": "The napkin math is public now: the serving floor",
    "essayUrl": "https://jakelawrence.xyz/blog/the-napkin-math-is-public-now",
    "measuredAt": "2026-07-27",
    "license": "CC-BY-4.0",
    "formula": "$/MTok = $/hour / (tokens/second * 3600 * utilization) * 1,000,000",
    "note": "One line of arithmetic, three deployment shapes, three assumption levels. Hourly costs trace to public listings; the throughput column is the shakiest input, exposed as a toggle rather than asserted. Every printed figure re-derives in CI."
  },
  "model": {
    "name": "Kimi-K3",
    "totalParams": "3T",
    "precision": "native MXFP4",
    "weightsTB": 1.5
  },
  "levels": [
    {
      "id": "conservative",
      "label": "Conservative"
    },
    {
      "id": "middle",
      "label": "Middle"
    },
    {
      "id": "optimistic",
      "label": "Optimistic"
    }
  ],
  "scenarios": [
    {
      "id": "rented",
      "label": "Rented frontier node",
      "sub": "16× B200 on a GPU marketplace",
      "dollarsPerHour": 72,
      "costNote": "16 GPUs at an assumed $4.50 per GPU-hour, mid-2026 on-demand range",
      "perLevel": {
        "conservative": {
          "perMTok": 10,
          "display": "$10.00"
        },
        "middle": {
          "perMTok": 1.9,
          "display": "$1.90"
        },
        "optimistic": {
          "perMTok": 0.74,
          "display": "$0.74"
        }
      }
    },
    {
      "id": "owned",
      "label": "Owned node, amortized",
      "sub": "the same 16× B200, bought outright",
      "dollarsPerHour": 28,
      "costNote": "roughly $700k of hardware over three years plus ~20 kW of power, rounded",
      "perLevel": {
        "conservative": {
          "perMTok": 3.89,
          "display": "$3.89"
        },
        "middle": {
          "perMTok": 0.74,
          "display": "$0.74"
        },
        "optimistic": {
          "perMTok": 0.29,
          "display": "$0.29"
        }
      }
    },
    {
      "id": "sovereign",
      "label": "The sovereignty rig",
      "sub": "used quad-socket Xeon, up to 3 TB RAM, no GPU",
      "dollarsPerHour": 1.21,
      "costNote": "under $30k of used hardware over three years plus ~$48/month of electricity at 7.5¢/kWh",
      "perLevel": {
        "conservative": {
          "perMTok": 168.06,
          "display": "$168"
        },
        "middle": {
          "perMTok": 67.22,
          "display": "$67"
        },
        "optimistic": {
          "perMTok": 42.01,
          "display": "$42"
        }
      }
    }
  ],
  "cheapestFloor": {
    "conservative": {
      "perMTok": 3.89,
      "display": "$3.89"
    },
    "middle": {
      "perMTok": 0.74,
      "display": "$0.74"
    },
    "optimistic": {
      "perMTok": 0.29,
      "display": "$0.29"
    }
  },
  "closedUnknowns": [
    "total parameters",
    "active parameters",
    "serving precision",
    "cost to serve",
    "gross margin",
    "who else could serve it"
  ],
  "precedent": {
    "model": "GLM 5.2",
    "releasedOn": "2026-06-16",
    "observedOn": "2026-07-27",
    "indexStart": 100,
    "indexEnd": 55,
    "source": "OpenRouter provider listings, via the Kimi-K3 release thread",
    "dropPct": 45,
    "days": 41
  }
}