{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "electricity-per-million-tokens",
 "title": "Electricity per million tokens, card by card",
 "dek": "What a million tokens of writing costs in electricity on each graphics card this site has measured, at one dated national price, from board watts and speeds already published: about three and a half cents on a laptop card to fifty-eight on a pair of used 10 GB cards, idling billed separately.",
 "published": "2026-10-30",
 "series": [
  "bench"
 ],
 "licence": "CC BY 4.0",
 "status": "pending-judge",
 "pack_note": "no judge is seated, so every chip is held",
 "notice": "items with published:false are held and are not this page's published words",
 "source": {
  "url": "https://research.strata2signal.com/electricity-per-million-tokens/",
  "md_url": "https://research.strata2signal.com/electricity-per-million-tokens/index.md",
  "md_sha": "bba7712d2c41a1eb58a4ca2037ee46395e37d10c36d988b157e9b69dd323d765",
  "html_sha": "54271f87a9968702e340145eb5bbc40eb4844904ba097b04dce955fe7eab3d0e",
  "anchors_sha": "1228df3eb93cfba8b3c2494e79f22d4f3d91f24bbdc780fc8f478fadde086aa7"
 },
 "short": {
  "paragraph": "A token is a piece of a word: in this site's own prose, about four-fifths of a word by the count on How the speed tool works (343 words came to 415 to 420 tokens), so a million tokens is roughly 800,000 words of answers. Writing a million of them with the mixture-of-experts model this workshop runs (it uses about 4 billion of its 26 billion parameters for each token) cost 3.42 cents of board electricity on an NVIDIA GeForce RTX 5090 Laptop GPU (24 GB) held to 95 W, 8.39 to 9.74 cents on an NVIDIA GeForce RTX 3090 Ti (24 GB) and 8.94 to 9.32 cents on an NVIDIA GeForce RTX 3090 (24 GB) depending on the power cap, a limit set on what a card may draw, and 12.87 to 15.23 cents on two used NVIDIA GeForce RTX 3080 (10 GB) cards sharing the model. With a dense 24-billion-parameter model, one that uses every parameter for every token, every figure is three to five times higher: 15.93 cents on the laptop card, 57.91 on the pair. The price is 18.31 cents a kilowatt-hour, the U.S. residential average for July 2026, the rate this site's running-costs page uses. Nothing was measured for this page; every figure is a published one put through one formula. Board power, not the wall; one caller; and idling is its own bill, printed separately: a 3090 Ti holding a model and doing nothing costs about $30 a year, and the box that answers around the clock here about $17 a month at the meter.",
  "from_dek": false,
  "counts": {
   "words": 3491,
   "minutes": 16,
   "tables": 1,
   "kit": false
  },
  "bullets": []
 },
 "sections": [
  {
   "id": "what-is-being-counted-and-at-what-price",
   "heading": "What is being counted, and at what price",
   "level": 2,
   "span": [
    2593,
    4175
   ],
   "chunks": [
    [
     2593,
     4175
    ]
   ],
   "chars": 1582,
   "digest": "the article explains how joules per thousand tokens are calculated by multiplying mean board watts by runtime and dividing by tokens. it uses an electricity price of 18.31 cents a kilowatt-hour, which is the average American home price in July 2026.",
   "digest_skipped": null
  },
  {
   "id": "the-arithmetic-once",
   "heading": "The arithmetic, once",
   "level": 2,
   "span": [
    4175,
    4800
   ],
   "chunks": [
    [
     4175,
     4800
    ]
   ],
   "chars": 625,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "the-table",
   "heading": "The table",
   "level": 2,
   "span": [
    4800,
    10877
   ],
   "chunks": [
    [
     4800,
     10877
    ]
   ],
   "chars": 6077,
   "digest": "a table presents data for various cards and power settings using ollama 0.32.13 and specific models. it lists tokens a second, mean board watts, joules per thousand tokens, and cents per million tokens for mixture and dense models across different hardware.",
   "digest_skipped": null
  },
  {
   "id": "what-the-table-says",
   "heading": "What the table says",
   "level": 2,
   "span": [
    10877,
    13224
   ],
   "chunks": [
    [
     10877,
     13224
    ]
   ],
   "chars": 2347,
   "digest": "the laptop is the cheapest writer for the mixture model. the dense model costs three to five times more than the mixture model on any one card. power caps can move the bill in different directions, and two 3080s spend more electricity per token than a 3090.",
   "digest_skipped": null
  },
  {
   "id": "the-cards-the-table-cannot-hold-and-three-other-instruments",
   "heading": "The cards the table cannot hold, and three other instruments",
   "level": 2,
   "span": [
    13224,
    16334
   ],
   "chunks": [
    [
     13224,
     16334
    ]
   ],
   "chars": 3110,
   "digest": "the section covers data not in the main table, including a 16 GB card, a 12 GB card, a 96 GB card, a machine with no graphics card read at the wall, and multiple callers at once.",
   "digest_skipped": null
  },
  {
   "id": "what-idling-costs-separately",
   "heading": "What idling costs, separately",
   "level": 2,
   "span": [
    16334,
    18443
   ],
   "chunks": [
    [
     16334,
     18443
    ]
   ],
   "chars": 2109,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-this-page-does-not-count",
   "heading": "What this page does not count",
   "level": 2,
   "span": [
    18443,
    20255
   ],
   "chunks": [
    [
     18443,
     20255
    ]
   ],
   "chars": 1812,
   "digest": "the article notes it does not count wall power, multiple callers, or idling in the table cells. it also mentions that mixture model joules are thin and that the price is a national average.",
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work",
   "heading": "How to check our work",
   "level": 2,
   "span": [
    20255,
    20887
   ],
   "chunks": [
    [
     20255,
     20887
    ]
   ],
   "chars": 632,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "who-ran-this-and-thanks",
   "heading": "Who ran this, and thanks",
   "level": 2,
   "span": [
    20887,
    22927
   ],
   "chunks": [
    [
     20887,
     22927
    ]
   ],
   "chars": 2040,
   "digest": null,
   "digest_skipped": "credits"
  }
 ],
 "chips": [
  {
   "id": "c-d3b86f53",
   "q": "how is the cost of electricity calculated per token?",
   "a": "The calculation multiplies the mean board watts by the runtime seconds, divides by the tokens written, and converts the resulting joules to kilowatt-hours. This value is then multiplied by the electricity rate, which is set at 18.31 cents a kilowatt-hour based on the July 2026 U.S. residential average.",
   "cites": [
    "what-is-being-counted-and-at-what-price"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-f15bf365",
   "q": "what formula converts joules to cents per million tokens?",
   "a": "The formula used is joules per thousand tokens multiplied by 1,000, divided by 3,600,000, and then multiplied by 18.31 cents. This converts the energy used during the decode clock into a monetary cost based on the kilowatt-hour rate.",
   "cites": [
    "the-arithmetic-once"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-1a0a224b",
   "q": "what specific conditions were used for the benchmark runs?",
   "a": "Every row used ollama 0.32.13, Q4_K_M weights, an 8-bit cache, and 256 tokens out. The data represents the median of three scored runs with board power read from the driver at 2 Hz.",
   "cites": [
    "the-table"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-dff4bdfc",
   "q": "how does the model type affect the electricity bill?",
   "a": "The dense model moves the bill three to five times more than the mixture model. For example, on a laptop at 95 W, the dense model costs 15.93 cents per million tokens compared to 3.42 cents for the mixture model.",
   "cites": [
    "what-the-table-says"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-e593bb2d",
   "q": "what happens to costs when using a machine without a graphics card?",
   "a": "A mini PC with no graphics card drew a mean of 75.6 W at the battery backup unit's meter, resulting in approximately 32 cents a million tokens for the mixture model.",
   "cites": [
    "the-cards-the-table-cannot-hold-and-three-other-instruments"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-dc70dfc9",
   "q": "how much does it cost to keep a model resident in memory?",
   "a": "Idling costs vary by hardware. A 3090 Ti with the mixture model resident at a 300 W cap costs about $30 a year, while two 3080s holding the same model at 250 W each cost about $137 a year.",
   "cites": [
    "what-idling-costs-separately"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-7fbe5e36",
   "q": "what elements are excluded from the board power measurements?",
   "a": "The measurements exclude the processor, system memory, case fans, and the power supply. They also do not account for the power drawn at the wall, which would be higher due to power supply losses.",
   "cites": [
    "what-this-page-does-not-count"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "two-used-3080s-priced",
   "why": "shares ground with § What the table says · § At a glance"
  },
  {
   "slug": "one-3090-against-one-3090-ti",
   "why": "shares ground with § What the table says · § At a glance"
  },
  {
   "slug": "six-cards-the-same-questions",
   "why": "shares ground with § What the table says · § Writing: how fast each one answers"
  }
 ],
 "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nThe figures are eight pages' work and not this one's: [*A laptop, asked the desktop's questions*](/a-laptop-asked-the-desktops-questions/) for the laptop rows and the streams; [*RTX 3090 vs RTX 3090 Ti*](/one-3090-against-one-3090-ti/) for the 24 GB cards' ladders and idle watts; [*Two 3080s against one 3090, and the cap decides*](/two-used-3080s-priced/) for the pair and the 250 W rows; [*RTX 5080 vs RTX 3080 Ti*](/rtx-5080-vs-rtx-3080-ti/) for the 16 GB and 12 GB cards; [*What it costs to run a high-end AI rig at home*](/what-it-costs-to-run/) for the price, the 96 GB card's probe and the box at the meter; [*Two Hours at 12 tok/s — On Battery*](/two-hours-on-battery/) for the mini PC; [*A short history of Mistral*](/a-short-history-of-mistral/) for the 96 GB card's readings on these two models; and [*How the speed tool works*](/how-the-speed-tool-works/) for the count of words in a token. The U.S. Energy Information Administration publishes the price monthly where anyone can read it.\n\n**EVGA** made the 24 GB boards and one of the 10 GB ones, **Gigabyte** the other 10 GB board, **ZOTAC** the 16 GB card, **NVIDIA** the Founders Edition, the laptop part, the workstation card and the driver every watt on this page was read from, and **MSI** the laptop that carried its board. The writing was done on **ollama**, and under it **llama.cpp** and **NVIDIA's CUDA**. The models are Google's **Gemma 4**, **Mistral Small 3.2** from Mistral AI, and **Qwen 3.5** from Alibaba Cloud on the smaller cards' bank. Each source page names the licence it found for the model it ran; this page adds none, because it ran nothing.\n\nA small human team owns the cards, set the rules before the runs and signed the figures on each source page; a fleet of AI agents ran the harnesses and did the division here under that team's rulings. Thanks to the readers who asked, on every forum this workshop reads, what the electricity actually comes to: this page is the division they asked for.",
 "kit": null,
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 0,
  "digests_written": 7,
  "digests_kept": 5,
  "chips_written": 7,
  "chips_kept": 7,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 0,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0,
   "identifier": 1,
   "bare_figure": 0,
   "house_voice": 0,
   "editorial": 0,
   "hand-edit": 4
  },
  "new_noun_tokens_checked": [
   "0.32.13",
   "000",
   "1",
   "12",
   "136.618",
   "137",
   "14",
   "15.93",
   "16",
   "18.31",
   "2",
   "2026",
   "240.49",
   "250",
   "256",
   "26b",
   "3",
   "3.42",
   "30",
   "300",
   "3080s",
   "3090",
   "32",
   "350",
   "600",
   "75.6",
   "758.13",
   "760",
   "8",
   "8-bit",
   "8.94",
   "95",
   "96",
   "American",
   "GB",
   "Hz",
   "July",
   "PC",
   "Q4KM",
   "Ti",
   "U.S",
   "W",
   "gemma4"
  ],
  "new_noun_tokens_withheld": 0,
  "figure_definition": "v3",
  "figure_boundary": "guarded",
  "drops": [
   {
    "kind": "bullet",
    "reason": "identifier",
    "item": "the laptop card with gemma4:26b resident held to 95 W costs about $14",
    "detail": "gemma4 is in the cited section only as code or in a table cell"
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 41,
    "text_sha": "0de8dabc87ed",
    "figure": "136.618",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-10-29",
    "why": "prints 136.618 with no unit and no model; the page gives it as tokens a second for a 3090 at its own 350 W limit writing the mixture model, gemma4:26b.",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-30: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-29 (matched on its line key). The earlier record reads: removed from this pack by hand on 2026-10-29 (UTC), after the writer ran and after the gates had counted: prints 136.618 with no unit and no model; the page gives it as tokens a second for a 3090 at its own 350 W limit writing the mixture model, gemma4:26b. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 80,
    "text_sha": "b71cc5cca0c9",
    "figure": "18.31",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-10-29",
    "why": "prints the price as 18.31 cents with no unit; the page's price is 18.31 cents a kilowatt-hour, the U.S. residential average for July 2026.",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-30: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-29 (matched on its line key). The earlier record reads: removed from this pack by hand on 2026-10-29 (UTC), after the writer ran and after the gates had counted: prints the price as 18.31 cents with no unit; the page's price is 18.31 cents a kilowatt-hour, the U.S. residential average for July 2026. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 225,
    "text_sha": "43d530e9b945",
    "section": "the-arithmetic-once",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-29",
    "why": "says 240.49 W at 136.618 tokens a second results in 1,758.13 joules per thousand tokens; the page divides them to 1.76 joules a token and gives 1,758.13 as its source page's own figure, read from the runtime's decode clock.",
    "detail": "Carried forward by --carry-hand-edits on 2026-10-30: the writer produced this line again and the pack on disk records it removed by hand on 2026-10-29 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-10-29 (UTC), after the writer ran and after the gates had counted: says 240.49 W at 136.618 tokens a second results in 1,758.13 joules per thousand tokens; the page divides them to 1.76 joules a token and gives 1,758.13 as its source page's own figure, read from the runtime's decode clock. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 201,
    "text_sha": "018f44400a66",
    "section": "what-idling-costs-separately",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-30",
    "why": "says a 3090 Ti at 300 W idles for about $30 a year, which reads as a card drawing 300 W; the page's 300 W is the power cap, and the $30 is its 18.74 W idle median with the mixture model held in memory and nothing generating.",
    "detail": "removed from this pack by hand on 2026-10-30 (UTC), after the writer ran and after the gates had counted: says a 3090 Ti at 300 W idles for about $30 a year, which reads as a card drawing 300 W; the page's 300 W is the power cap, and the $30 is its 18.74 W idle median with the mixture model held in memory and nothing generating. Not a gate drop, and recorded here so the ledger adds up."
   }
  ]
 },
 "generated_utc": "2026-10-30T01:31:12Z"
}