# Recon: the economics and the utility frame

**Train:** the cost of intelligence (an operator, 2026-09-23). **This part:** Epoch's chart and the data behind it, the other cloud price-per-capability series, our own cost inputs, a proposed cost-per-correct-answer formula for local and frontier runs, and the "intelligence as a utility" frame.
**Written:** 2026-09-23, 15:30Z to 15:55Z. Read-only apart from this file and the pinned source copies under `recon/sources-economics/`: no GPU jobs, no service changes, no keys printed.
**Evidence labels used below.**
- **VERIFIED:** read at the source today. The URL and the read time are given.
- **RECORD:** from an estate or hub file. The path is given.
- **SECONDARY:** a search-result or press summary that has not been read at its primary source. Never publish one of these as it stands.
- **ILLUSTRATION:** arithmetic on existing receipts, done to show the formula working. It is not a bench result.

---

## 0. The short version, with corrections to the brief

1. **Epoch's unit is dollars per benchmark question at a fixed accuracy.** It is not a price per token. For each benchmark and date, Epoch takes the cheapest model that reaches the accuracy, using transcripts and prices from OpenRouter's history, or rented GPUs for open models that OpenRouter doesn't carry. The headline is **47.0 % a quarter, which is 12.67× a year, over 2023–26, after adjusting for inflation**. At SOTA debut the rate is 75× a year; two years later it is 4.7× a year. (VERIFIED, §1)
2. **Correction: "a million-fold in about five years" is not what the chart's data says.** Epoch's own `Figure 1.csv` indexes the LLM line at 100 in 2021.88 and 0.001 in 2026.67. That display is rounded. Compounding 12.67× a year over 4.78 years gives a fall of about **1.9 × 10⁵** (index 0.00054). The honest wording is five orders of magnitude, or about two hundred thousand-fold. The post itself does not claim a million-fold (the fetch found no such figure). Treat "million" as wrong until someone finds it on the page. (VERIFIED, §1.3)
3. **Correction: the 2021–23 part of the LLM line is extrapolated, not measured.** The index point for 2023.50 (1.649) is the 2023–26 rate run backwards. Epoch calls that rate "plausible" back to November 2021, citing Appenzeller's 10× a year and a GPT-3-to-Llama-2-7B comparison at 31× a year. It did not fit 2021–23 data. (VERIFIED, §1.3)
4. **Correction: the brief's naming rule is out of date.** The brief says to publish the 96 GB card by class only. **D-20260917-05** (an operator, 2026-09-17T16:09Z) retired that rule: hardware is named in full on the shelf, and "RTX PRO 6000 Blackwell" already appears on 10+ hub pages. What stays fenced is box and pet names (so "the desktop's larger card" and "the desktop" stay off pages), addresses, paths, the dog and the location. The **price** is a separate matter, and a real one: the shelf holds **no price of record** for that card, and the hub rule is that "an invented one is worse than a blank" (RECORD). §3.2 sets out three ways to publish without inventing one. The best of them is the **break-even purchase price**.
5. **Our electricity input is VERIFIED today.** It is 18.34 ¢/kWh, the US residential average for June 2026 (EIA Electric Power Monthly table 5.6.A, released 2026-08-26). **The next release is tomorrow, 2026-09-24**, so the rule should say "the latest month at the reading's date". Commercial is 14.19 ¢ and industrial 9.17 ¢. Data centres buy at roughly industrial rates, so the page should carry an industrial-rate row for fairness.
6. **Card prices on record.**
   - 3090: $1,449.99 paid, renewed, May 2026 (RECORD, exhibit 46).
   - 3090 today: $1,829.99–$1,879.99 as read on 2026-09-13/17, and $1,094.44 as August 2026's average completed used sale.
   - Laptop: $3,099.99 for the whole machine, new, ordered 2026-07-03 (public form).
   - 96 GB card: none of record. SECONDARY reports put the list price at $8,565 at launch (2025-03) and about $16,000 by 2026-08/09, driven by the GDDR7 shortage.
   - **Owned hardware got more expensive in 2026, while the cloud's price per capability fell about 13× a year.** That is a headline fact in its own right, and it breaks naive mark-to-market amortisation (§3.4).
7. **Utilisation decides whether owning pays, and ours is low.** The house rule counts a card as busy when it draws over 100 W. By that rule the 96 GB card was busy **2.24 % of 2026-09-21 and 2.59 % of 2026-09-22**, from 15-second ticks (RECORD, estate-watch history). The what-it-costs piece found 6.05 % on 09-09 and 1.79 % on 09-10, and about nine tenths of the shelf's electricity was idle.
   - ILLUSTRATION: one 3090 serving one stream of gemma4:26b has to be busy about **54 %** of the time to beat renting the same model on OpenRouter. The assumptions are 3-year straight-line, no resale, board watts and 18.34 ¢.
   - At 2.6 % busy it costs about **16×** the rental price per question.
   - This is the self-refuting result the train has to be ready to publish (§4.4).
8. **Proposed formula: a levelised cost of intelligence (LCOI).** It is built on NREL's simple LCOE formula: capital × capital recovery factor, plus fixed O&M (the idle electricity), divided by busy hours, plus fuel (electricity per answer). Divide by accuracy to get cost per correct answer.
   - The frontier side is metered tokens × the list price read on a date, with every call counted, including failed and format-failed ones.
   - Keep the four cost states from exhibits 40 and 41 (metered, plan-included, no-figure-held, own-silicon). The new own-silicon rows now carry a derived dollar figure, labelled as derived and never as a charge.
9. **Precedent for benching frontier models: confirmed (RECORD).** The no-cloud-AI law governs live products, not research benches. Exhibits 40 and 41 ran frontier arms through three routes:
   - a sealed Claude CLI on a subscription, which gives no per-call receipt, only a list-rate estimate;
   - the OpenAI API, metered, with a cap set below the key's balance;
   - the ollama.com cloud shelf, plan-included, with kimi metered.

   They logged egress receipts and caps for each leg.
   - Exhibit 40's metered bill came to **$23.05** (Astra $17.79 at the cached rate, $22.83 if nothing had been cached; kimi $5.25), plus a CLI list-rate estimate of $31.44 that is not a receipt.
   - Exhibit 41's came to **$0.7652**.
10. **The utility frame writes itself, with a twist.** On-site intelligence burns grid electricity as its fuel. The fuel's price is flat to rising (nominal 13.66 ¢ in 2021, 17.3 ¢ in 2025 and 18.34 ¢ in June 2026, from Epoch's own sheet and EIA). Epoch's chart line for electricity stops in 1973, and that is the year its price stopped falling (Hirsh, *Power Loss*). §5 sets out the mapping: grid tariff = API, isolated plant = owned card, capacity factor = utilisation, LCOE = LCOI.
11. **Quick wins, all doable today without a GPU (§7):**
    - (a) Cross-walk our measured joules per token onto Epoch's per-question token counts for the four open models that appear in both our records and Epoch's data. That puts dots beside Epoch's line today, labelled.
    - (b) A daily snapshot of OpenRouter's public price JSON. That builds our own dated grid-tariff series.
    - (c) Pin Epoch's data by commit sha. Done here, but the repo has **no licence file**, so it stays internal until that is settled.
    - (d) Record which of the four 3090s cost what. The price ledger by UUID is missing.
    - (e) A plug-in wall meter for the laptop. It has no UPS, and its wall draw has never been measured.

---

## 1. Epoch AI: "The plunging price of thought"

### 1.1 The post (VERIFIED, read 2026-09-23 ~15:32Z)
- **URL:** https://epoch.ai/publications/the-plunging-price-of-thought
- **Authors and date:** Luke Emberson and David Roodman, **published 2026-09-22**, one day before this train started.
- **Licence:** "Creative Commons Attribution" (CC BY 4.0). The citation Epoch asks for:
  > Luke Emberson and David Roodman (2026), "The plunging price of thought". Published online at epoch.ai. Retrieved from 'https://epoch.ai/publications/the-plunging-price-of-thought' [online resource]. Accessed 23 Sept 2026.
- **Unit:** the **cost per question** of reaching a fixed accuracy on a benchmark. The worked example is o3 at about $0.30 a question for 75 % on GPQA Diamond (2025-01-31), against GPT-5.6 Luna at $0.0004 a question for about the same score, a year and a half later. That is a 725-fold fall in under 18 months.
- **Cost frontier:** C(t, a) = min over models released by date t with accuracy ≥ a of the model's cost c. It is the empirical Pareto staircase.
- **Budgets:** a CAISI-style (NIST Center for AI Standards and Innovation) truncation. Epoch replays each model's high-budget transcripts under hypothetical per-question output-token caps, so one run yields a whole cost–accuracy curve for that model.
- **Prices:** `price_source` in the data is mostly `openrouter-wayback`, meaning OpenRouter's price at the time, taken from the Wayback Machine. Some rows are first-party prices. For open models that OpenRouter doesn't carry, Epoch used **`gpu-effective`**: the authors "used the cost of their rented hardware", optimising the hardware and settings. Checked against five open models served through an API, the gap was under 30 %.
  - **For us this matters:** Epoch's line already includes open models on **rented** GPUs. Our slice would be open models on **owned** GPUs.
- **Benchmarks:** five primaries (AIME / OTIS Mock, Chess Puzzles, FrontierMath tiers 1–3, GPQA Diamond, Mystery Game Puzzles), plus secondaries (MATH L5, SimpleQA Verified, SWE-Bench Verified, other FrontierMath cuts).
- **Model:** a Box–Tidwell (Box–Cox transforms) frontier fit, with time counted in years since the GPT-3 release. A model-free staircase check gives the same answer.
- **Headline rates:**
  - The model-free average is **47.1 % a quarter** (43.0–53.1 % across the five primaries). The Box–Tidwell average is **47.0 %**. Either way it is about **13× a year**.
  - **No standard errors are reported. The authors say so in plain words.**
- **Rate by age of the capability:** **66 % a quarter (75× a year) when a level first debuts as SOTA**, and **32 % a quarter (4.7× a year) two years later**.
- **Comparisons (from the post):** 4× faster than DNA sequencing, 6× faster than compute, 18× faster than lithium batteries, 54× faster than electricity. These ratios are of log-point rates (Figure 1.csv, §1.3).
- **Caveats the authors list:**
  - benchmaxxing, with the secret Mystery Game falling a little slower (38.7–44.0 %);
  - a benchmark is not useful work;
  - **no real user stays on the cost frontier**;
  - three years is a short window;
  - the averaging choice matters (grid schemes give 42.9–58.0 %);
  - the data are incomplete and noisy.

  Their verdict: "reasonably representative … should not be read as exact."
- **Earlier work cited in the post:**
  - Appenzeller 2024 at a16z: 10× a year.
  - Epoch, March 2025: 9–900× a year.
  - Tveit Ihle, September 2025: coding challenges halving every 1.4–2 months, i.e. 64–380× a year. SECONDARY, since only Epoch's citation was read.
  - GPT-3 against Llama-2-7B on MMLU: $60 to $0.20 per M tokens, i.e. 31× a year.

### 1.2 Data and code (VERIFIED, pinned)
- **Repo:** https://github.com/droodman/inference-cost. Created 2026-08-17. HEAD at the time of reading: **`a89f968bcb1f07b5c72e83fce615376a07700c66`** (2026-09-22T21:23:23Z, "Update Cost declines by tech.xlsx").
  - Code is R (plus a little Stata and Julia). The viewers are at https://droodman.github.io/inference-cost/, including `viewers/frontier_plot_viewer.html`, `regression_table_viewer.html` and `record_timeline_viewer.html`.
- **⚠ Licence gap:** GitHub reports **no licence** on the repo (`license: null`). The post is CC BY, and it says "Epoch AI's work is free to use … under the Creative Commons Attribution license". Whether that covers the data files as republished in one of our kits is **not settled**. Recommendation: cite and link, and keep our copies internal until Epoch confirms.
- **Our copies** are in `recon/sources-economics/epoch-inference-cost/`, with shas in `recon/sources-economics/SHA256SUMS`.

| file | sha256 (first 16) | what it is |
|---|---|---|
| `Figure 1.csv` | `74781708f25944586` | **The chart's data.** Six technologies, their date ranges, units, compounded rates a quarter, fastest 3-year decline, log points a quarter, and the LLM index series. Hand-made, per the README. |
| `Cost declines by tech.xlsx` | `b1ae3c72395c9d06` | Sheets: Figure, Figure data, Electricity, Illumination, Lithium batteries, Compute, DNA sequencing, CPI, Sources. **It holds the full comparator series and their sources** (§1.4). |
| `cost_truncated_curves.csv` | `578091c3ee84dffe` | 15,959 rows, 225 models, 11 benchmarks. Columns: run_id, model, benchmark, effort, budget_axis, budget, acc, mean_tokens_used, mean_output_tokens, mean_input_tokens, mean_cached_input_tokens, mean_cache_write_tokens, **cost_per_task_usd**, full_run_acc, n_samples, price_date, price_source. |
| `Figure 2.csv` | `00d51d32b0ef9a8b` | The cost-record holders at 25 % and 75 % accuracy, by benchmark: model, accuracy, date, cost. |
| `staircase_check.csv` | `c152f6c2e378d421` | The model-free decline rate by benchmark, e.g. GPQA 46.97 %, AIME 47.14 %, SWE-Bench Verified 27.49 %, FM tier 4 (2025-07) 26.05 % a quarter. |
| `Model versions-Grid view.csv` | `168f7bd5a2a882519` | 950 model versions, with release date, **knowledge cutoff date** and HF id. **Useful to the contamination recon.** |
| `benchmarks_runs-PUBLIC VIEW.csv`, `edi_scores.csv` | `fd0b99097b04eda9`, `a8561616e57ea916` | Run log and benchmark difficulty index. |

- **Models in Epoch's data that we run or could run** (a basis for cross-walks). Each is given as benchmark, accuracy, mean output tokens a question, and $ a question with its price date and source:
  - `gemma-4-26b-a4b-it` (minimal effort): GPQA 0.7323, 1,370.2 tokens, **$0.000484** (2026-08-03, openrouter-wayback); AIME 0.8222, 5,673 tokens, $0.001942.
  - `gemma-4-31b-it` (minimal): GPQA 0.7576, 847.6 tokens, $0.000313.
  - `mistral-small-3.2-2506`: GPQA 0.4905, 876.7 tokens, **$0.001895 (gpu-effective, i.e. rented GPU)**.
  - `gemma-3-27b-it`: GPQA 0.4773, 737.9 tokens, $0.000353.
  - `qwen3.6-27b` and `qwen3.6-35b-a3b`: GPQA 0.8586 and 0.8384 with thinking, at 8,250 and 8,498 tokens.
  - `llama-2-7b-chat` and `Llama-2-13b-chat`: chess only.
  - `gpt-oss-20b`, `openai/gpt-oss-120b` and several 2023–25 open models are also present.

### 1.3 The chart's numbers, straight from `Figure 1.csv` (VERIFIED)

All rates are "Price drops expressed as compounded quarterly rates, **after inflation adjustment**."

| technology | date range | unit | compounded a quarter | log points a quarter | × a year | LLM rate ÷ this one's (log) |
|---|---|---|---|---|---|---|
| Electricity | 1892–1973 | $/watt-hour | 1.2 % | 1.2 % | 1.05× | 54 |
| Lithium batteries | 1991–2024 | $/watt-hour | 3.6 % (fastest 3 years: 7.6 %) | 3.6 % | 1.16× | 18 |
| Illumination | 1800–2023 | £/lumen-hour | 0.9 % | 0.9 % | 1.04× | 68 |
| Compute | 1940–2001 | $/calculation | 9.9 % | 10.4 % | 1.51× | 6 |
| DNA sequencing | 2001–22 (the post says 2001–25) | $/human genome | 14.2 % (fastest 3 years: 37.9 %) | 15.3 % | 1.84× | 4 |
| **LLM inference** | **2023–26** | **"$, given benchmark score"** | **47.0 %** | **63.5 %** | **12.67×** | 1 |

- **The LLM index series** (Year, t, index): 2021.88 → 0.00 → 100; 2023.50 → 1.62 → 1.649; 2026.67 → 4.78 → 0.001.
  - 100 × 12.67^(−1.62) = 1.64, which matches 1.649: **the 2021–23 segment is the 2023–26 rate run backwards.**
  - 100 × 12.67^(−4.78) = **0.00054**, which displays as 0.001. The fall over the whole span is about **1.87 × 10⁵**, not 10⁶.
- **The sheet carries six comparators.** The brief lists four. Illumination (Fouquet's series through OWID) is in the data. Check whether the published figure draws it.
- **"Inflation adjustment"** means CPI: FRED CPIAUCSL monthly, and HSUS for the early years. **Any comparison of ours with this line must say nominal or real.** For a single 2026 reading it makes no difference. For a multi-year series of our own, keep a CPI-U deflated column.

### 1.4 The comparators' sources (VERIFIED, from the xlsx "Sources" sheet)
- **Electricity:** Richard F. Hirsh, *Power Loss: The Origins of Deregulation and Restructuring in the American Electric Utility System* (MIT Press, 1999), p. 48 (https://archive.org/details/powerlossorigins0000hirs/page/48/mode/2up).
  - The sheet also cites the Edison Electric Institute, and notes "Nominal prices received from Richard Hersh [sic], Virginia Tech, September 11, 2026."
  - The sheet carries **US residential ¢/kWh, nominal, from 1892 to 2025**: 22 ¢ in 1892; 2.09 ¢ in 1969, the low; 2.38 ¢ in 1973; 13.66 ¢ in 2021; 16.00 ¢ in 2023; 16.48 ¢ in 2024; **17.3 ¢ in 2025**. It carries **real (1982–84 $) values to 2003**: 242.6 in 1892, 5.36 in 1973, 4.73 in 2003. **The real price stopped falling around 1973, and that is why the chart's line ends there.**
- **Compute:** Nordhaus, "The Progress of Computing" (NBER SI 2001): https://users.nber.org/~confer/2001/si2001/nordhaus.pdf
- **Lithium batteries:** OWID price-of-lithium-ion-battery-cells: https://ourworldindata.org/grapher/price-of-lithium-ion-battery-cells
- **Illumination:** OWID, the price of lighting in the UK: https://ourworldindata.org/grapher/the-price-for-lighting-per-million-lumen-hours-in-the-uk-in-british-pound
- **DNA sequencing:** NHGRI cost data table: https://www.genome.gov/sites/default/files/media/files/2023-05/Sequencing_Cost_Data_Table_May2022.xls

---

## 2. Other cloud price-per-capability series, 2021–2026

| series | window | rate | unit and method | URL | status |
|---|---|---|---|---|---|
| Appenzeller (a16z), "Welcome to LLMflation" | Nov 2021 → Nov 2024 | **10× a year**. $60 → $0.06 per M tokens for GPT-3-level quality (MMLU ≈ 42) | cheapest per-token list price for a model at or above a fixed MMLU score | https://a16z.com/llmflation-llm-inference-cost/ | SECONDARY, read through search. Epoch cites it, so the figure is corroborated. Read the page itself before quoting. |
| Epoch AI (Cottier, Snodin, Owen, Adamczewski), "LLM inference prices have fallen rapidly but unequally across tasks" | ~2022 → Mar 2025 | **9–900× a year** across six benchmarks; GPT-4-level GPQA: **40× a year** | lowest $ per M tokens at a fixed score | https://epoch.ai/data-insights/llm-inference-price-trends (2025-03-12) | VERIFIED (fetch). No download link seen. |
| Stanford HAI AI Index 2025 | Nov 2022 → Oct 2024 | **280× in about two years**: GPT-3.5-level (MMLU 64.8) from $20 to $0.07 per M tokens. Hardware cost −30 % a year, energy efficiency +40 % a year | $ per M tokens at a fixed MMLU | https://hai.stanford.edu/ai-index/2025-ai-index-report | SECONDARY (search summaries, and the 2026 Index said to carry it forward). Read the report before quoting. |
| Gundlach, Lynch, Mertens, Thompson, "The Price of Progress: Price Performance and the Future of AI" | to 2025/26 | **5–10× a year** at fixed performance for frontier models; **algorithmic efficiency ≈ 3× a year**, with open models used as the proxy after controlling for hardware and competition; **the price of running frontier models is rising 3–18× a year** | benchmark-run pricing dataset | https://arxiv.org/abs/2511.23455 (v1 2025-11-28, v2 2026-03-23) | VERIFIED (abstract) |
| Epoch 2026 (§1) | 2023–26, extended to 2021 | 12.67× a year (47 % a quarter), real | $ a question at a fixed score | above | VERIFIED |

**What the series say together, and why it matters for us:**
- At a **fixed capability**, the cloud price falls 5–13× a year.
- At the **frontier**, the price of the best model does not fall and may rise (Gundlach).
- The part of the fall that comes from **algorithms** is about 3× a year. That is roughly the part an owner of fixed hardware can capture by loading newer open weights. The rest comes from hardware price and performance, datacentre scale and batching, and price competition, and the owner of a card bought in 2026 gets none of it, which is why this part of the argument can go against owning.
- This gives the cost-of-intelligence write-up its thesis to test, and Bench 3 a prediction to pre-register (§4.4).

---

## 3. Our cost inputs

### 3.1 Electricity (VERIFIED, and RECORD for the house rule)
- **Rate:** **18.34 ¢/kWh**, US total, residential, June 2026, from EIA *Electric Power Monthly* **table 5.6.A**: https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_5_6_a
  - Read 2026-09-23 ~15:36Z. The page says "Data for June 2026 · Release Date: August 26, 2026 · Next Release Date: September 24, 2026".
  - Same row: June 2025 residential 17.47 ¢; commercial 14.19 ¢; industrial 9.17 ¢; all sectors 14.48 ¢.
  - Copy: `sources-economics/prices/eia-epm-table-5-6-a-read-2026-09-23.html` (sha `7d5bcbe3…`).
- **The house rule, already in print** (the mistral page, and the what-it-costs piece v10): "a national figure and not this workshop's bill". **The location law forbids the local utility's rate.** Keep this rule as it is.
- **Registration proposal:**
  - Primary: the latest EIA month at the reading's date, printed with its month.
  - Sensitivity rows at commercial and **industrial**. The industrial row is the fairness row: what a datacentre pays is closer to industrial.
  - Tomorrow's release (July 2026) will move the number. Freeze the rate per reading, and never reprice old readings silently. That is the reproducible-rule law applied to money.
- **Carbon, if wanted:** the what-it-costs piece used "~400 g CO₂e per kWh, a round mixed-grid intensity, **not measured here**." Keep that label, or cite EPA eGRID or EIA properly.
- **Measurement layers.** Every energy row names its layer:
  1. **board**: `nvidia-smi power.draw`, 1 Hz in the Jev harness and 15 s in the ledger;
  2. **box at the UPS output**: `ups.realpower` over NUT, 60 s cadence (the desktop CP2000PFCRM2U; benchbox and the desktop bench box PR1500; the mini node);
  3. **wall**: the UPS input. **Nothing measures this.** UPS losses are unseen, and so is the room's air conditioning.
  - The house's derived "at the meter" rule is box idle + (card draw − card idle) ÷ 0.9 for the PSU, × seconds (RECORD, the what-it-costs v10 methods).
  - **The laptop has no UPS.** The ledger holds the laptop's card only, not the box. Its wall draw has never been measured (§7).

### 3.2 The three reference cards: prices of record

| card | price, as the shelf holds it | kind | source |
|---|---|---|---|
| RTX 3090 (EVGA XC3 Ultra, renewed) | **$1,449.99** each, **paid** May 2026 (two cards) | historic cost | RECORD: hub roster, `home-inference-and-diffusion-rig-cost` (exhibit 46) |
| RTX 3090, the same listing | $1,849.99 renewed (2026-09-13 06:02–06:05Z); cheapest new, a ZOTAC Trinity OC, $1,829.99; renewed $1,879.99 (2026-09-17 15:00–15:30Z) | replacement-cost readings | RECORD: hub roster (exhibit 46; `two-used-3080s-priced`; `one-3090-against-one-3090-ti`) |
| RTX 3090, used market | **$1,094.44** (August 2026's average completed sale); $908.00 (the past year's average) | resale / market | RECORD: hub roster (exhibit 46) |
| ~~RTX 3090, two more EVGA XC3 bought new~~ | ~~**$1,899 each**, ordered about 2026-08-24, as spares~~ — **STRUCK 2026-09-23** (PLAN-v2 §5; `panel/article.md` M-2): the boards never arrived — the seller was a scammer and the money was refunded (`~/estate/inventory/hardware.yml` `did_not_come:`; `bench/ampere-ladder-2026-09/PRICES.md` §3, struck for exhibit fifty). Never a historic cost, never a pair, never on a page. | — | the two memory notes are superseded by the inventory ledger; do not re-import |
| RTX 5090 Laptop GPU | **$3,099.99**, the whole MSI Raider 16 Max HX (SKU B2WJ-002US), new, ordered 2026-07-03, before tax | historic cost of the **whole machine**. The GPU cannot be bought on its own. | RECORD: hub roster (exhibit 57). The order total with tax is **internal only** (D-20260922-005). |
| RTX PRO 6000 Blackwell (96 GB), the workstation card | **none of record** | — | RECORD: `compare_cards_outside.json` ("no price is printed for this card, by a standing rule of this workshop: it holds none of record, and an invented one is worse than a blank"); the mistral page says the same |
| RTX PRO 6000 Blackwell, market | launch list price **$8,565** (2025-03). About **$13,250** (2026-06) and **$16,000** (NVIDIA's own listing, 2026-08/09). Amazon about $14,979.99. Cause: the GDDR7 shortage. | replacement cost | **SECONDARY** (TechPowerUp, VideoCardz, wccftech, Thunder Compute; search summaries). Two primary fetches were refused (403/402). The shelf's standing rule is that **an outside price is read at the listing's own product page, never a results page**. That reading is owed. |

**Three ways to put the 96 GB card on a page without inventing a price:**
- **(a)** an operator supplies the price paid, and rules whether its public form (before tax, no place) can print. This follows the laptop's precedent.
- **(b)** A dated product-page reading, printed as "listed at", as a replacement cost. Never print it as "paid".
- **(c)** **Print the break-even instead of the price.** For example: "at the utilisation measured, the card would have to cost less than $X to beat renting the same model." That needs no price at all, and it is the more interesting number. **This is the recommendation**, with (a) kept internal for the private ledger.

**Gap:** there is **no ledger of which 3090 cost what by UUID**. The candidate reference cards are <withheld: GPU UUID of board E>, <withheld: GPU UUID of board A> and <withheld: GPU UUID of board F>, and the desktop's 3090 (<withheld: GPU UUID of board G>) is not a candidate. The shelf knows "two renewed at $1,449.99 (the pair in the darkroom's box, benchbox)" ~~and memory knows "two new at $1,899"~~ (struck 2026-09-23: that pair never arrived — `hardware.yml` `did_not_come:`; PLAN-v2 ⧖-11). Which UUID is which purchase is owed from an operator (§7).

**The fact that feeds the write-up:** the owned cards **appreciated** in 2026. A renewed 3090 went from $1,449.99 in May to ~~$1,849.99~~ **$1,879.99** in September (corrected 2026-09-23: the $1,849.99 reading was struck for exhibit fifty, `PRICES.md` §3; the figure of record is the 2026-09-17 15:00–15:30Z listing in the table above, exhibit 46), and the workstation card roughly doubled against its launch list price (SECONDARY — and **that clause never travels to a page**: P-6 forbids any dollar figure for the 96 GB card, `panel/article.md` M-1). Over the same months the cloud's price for a fixed capability fell about 13× a year. That is a strange world, and the page should say so plainly: the cost of owning came from memory scarcity, not from silicon.

### 3.3 Idle draw and duty: what an owned card costs to keep (RECORD)

| quantity | value | window and source |
|---|---|---|
| the 96 GB card, card mean | 44.9–75.1 W daily mean; minimum about 16.7 W since 2026-09-16 (27 W before) | ledger `~/.local/state/estate-ledger/power/days/2026-09-06 … 2026-09-22.json`, `gpus.desktop:0` |
| the 96 GB card, share of 15 s ticks above 100 W | **2.24 %** (09-21), **2.59 %** (09-22); above 200 W 0.50 % and 0.68 %. VRAM in use about **66,262–66,444 MiB** (the live seats resident), so about **30 GiB free** | estate-watch tick history `~/.local/state/estate-watch/history/2026-09-2{1,2}/` (computed today; 5,759 and 5,746 ticks) |
| the answer box's card above 100 W | 6.05 % (2026-09-09); 1.79 % (2026-09-10) | what-it-costs v10, the numbers table |
| the answer box (the desktop) at the UPS output | idle 105–119 W; daily mean 121–157 W; about 118 W live | ledger `boxes.desktop`; what-it-costs v10 |
| the darkroom's box (benchbox) at the UPS output | idle 91–99 W, with the two 3090s at about 23–24 W each by their own reading | exhibit 46; what-it-costs v10; ledger |
| a 3090, card idle | 22.26–22.39 W median, 52.76 W mean with the model resident (the mean carries the clocked-up tail after work) | hub roster, `one-3090-against-one-3090-ti` |
| the 5090 Laptop GPU, card idle | 7.14–7.18 W empty; 9.03–14.42 W resident; 11.57 W in the shipped posture | hub roster (exhibit 57) |
| the laptop, card share of 15 s ticks above 100 W | 1.46 % (09-21, a bench day); 0.97 % (09-22) | estate-watch history (computed today) |
| the shelf, idle share of electricity | about **nine tenths** (5.1 of 5.55 kWh a day over three clean days) | what-it-costs v10 |
| the rig's idle month (two 3090s in their box) | **$12.10–$12.28** a month at 18.34 ¢ | hub roster (exhibit 46), window 2026-09-12 → 2026-09-15 UTC |
| the series' standing usage figure | **six hours a day** (D-20260912-18) | what-it-costs piece |

### 3.4 An amortisation convention (proposal, for the hardening panel)

The shape is NREL's *simple* LCOE, which makes the analogy exact. NREL's documentation (https://www.nrel.gov/analysis/tech-lcoe-documentation, which now also resolves at `nlr.gov`; SECONDARY, read through search because the fetcher's DNS failed) gives:
> sLCOE = {(overnight capital cost × capital recovery factor + fixed O&M cost) / (8760 × capacity factor)} + (fuel cost × heat rate) + variable O&M cost
> CRF = i(1+i)ⁿ / ((1+i)ⁿ − 1)

The proposed defaults. Each is a registered choice, and each has a sensitivity row.
- **Capital C: historic cost as paid, before tax.** Sensitivity: replacement cost (a dated product-page reading). For the laptop, the whole machine, because the GPU cannot be bought alone. A second row for the laptop gives C = 0 as the "marginal" case: the machine is bought for development anyway, and the GPU work rides a sunk cost. For the 96 GB card, the break-even price (§3.2 (c)).
- **Life n: 3 years for the primary figure.** It is conservative and matches how fast the model curve moves. Sensitivity: 5 years. The IRS MACRS class for computers is 5-year property (from memory; verify in IRS Pub. 946). Hyperscalers book servers at 5–6 years: Microsoft and Alphabet 6, Meta 5.5, Amazon cut a subset to 5 in 2025 (SECONDARY; cite the 10-Ks if printed). Note that a 3090 from 2020 is still earning in 2026, so the realised life of a 3090 is already more than 6 years.
- **Resale R: 0 for the primary figure.** Sensitivity: today's used-market average ($1,094.44 in August 2026 for a 3090). ⚠ **Don't mark to market in the primary figure.** Resale values rose in 2026, which would make depreciation negative and the owned card look free. The honest economist's figure is the resale value's fall plus forgone interest. Print it as a sensitivity row with its sign, never as the headline.
- **Discount rate i: 0 for the primary figure (straight line).** Sensitivity: 5 %. On a $1,449.99 3090 over 3 years the annual capital charge is $483.43 at i = 0, $532.45 at 5 % and $562.64 at 8 % (arithmetic).
- **Utilisation u (the capacity factor).** Three registered scenarios, and one solved number:
  - **U100**, busy all the time at the bench's measured throughput. This is the card's ceiling, the "utility's best case".
  - **U25**, six hours a day (the series' standing figure, D-20260912-18).
  - **Umeasured**, the ledger's duty for that card over a stated window, using the >100 W rule and printing that rule.
  - The solved number is the **break-even u\*** at which owning costs the same as renting (a) the same open model and (b) the frontier.
- **Fixed O&M:** the card's idle watts × the hours it isn't busy × the price. Report it two ways: **card only**, and **box**, meaning the box's idle at the UPS output shared across its cards by a stated rule (split evenly, or by card idle).
  - The laptop's box idle is unmeasured, so it carries card-only rows until a wall meter arrives.
  - The 96 GB card is a special case. It sits beside live 24/7 seats, so its idle is **not caused by the bench**. The primary figure charges the bench **only its marginal energy** above the pre-run idle reading (the Jev harness's `idle_reading` → `net_joules_per_decision`), and the gross figure prints beside it.
- **Concurrency:** busy seconds per answer = the answer's tokens ÷ the card's **aggregate** throughput at the served concurrency c. Register c = 1 (a person at a keyboard) and c = the card's best measured value. The cloud's price is set by very large batches, so a one-stream owner is structurally disadvantaged, and the page must say which c each row assumes.
  - Existing receipts: the laptop does 354.050 tok/s summed over four callers against 155.3 for one; a 3090 does about 210 tok/s at two streams (RECORD, exhibit 57 and `never-split-a-model-over-two-cards.md`).

### 3.5 The frontier side: prices and our own bills

**List prices read today** (2026-09-23):

| provider · model | input / cached input / output, $ per M | batch | read |
|---|---|---|---|
| Anthropic · Claude Fable 5.1 | $10 / cache hit $0.25 (0.025×) / $50; 5-minute cache write $12.50 | 50 %: $5 / $25 | VERIFIED https://platform.claude.com/docs/en/about-claude/pricing.md (~15:35Z) |
| Anthropic · Claude Opus 5.5 / Opus 5 | $4 / $0.20 / $20 · $5 / $0.50 / $25 | $2 / $10 · $2.50 / $12.50 | same |
| Anthropic · Claude Sonnet 5 / Haiku 4.5 | $2 / $0.20 / $10 (the planned rise to $3 / $15 cancelled) · $1 / $0.10 / $5 | $1 / $5 · $0.50 / $2.50 | same. Also on that page: a 1.1× multiplier for US-only `inference_geo`; the newer tokenizer "≈30 % more tokens for the same text" |
| OpenAI · GPT-6 Astra / Sol / Luna | $10 / $1 / $50 · $2 / $0.20 / $10 · **$0.10 / $0.01 / $0.50** | 50 % (Batch); Flex at batch rates | https://developers.openai.com/api/docs/pricing, through a **fetcher summary**. Re-read the page before printing. Astra's $10 / $1 / $50 matches exhibit 40's roster, read 2026-09-05. |
| OpenAI · GPT-5.6 Sol / Terra / Luna | $4 / $0.40 / $20 · $2 / $0.20 / $12 · $0.20 / $0.02 / $1.20 | 50 % | same |
| OpenRouter, the market list for open weights we could own | gemma-4-26b-a4b-it **$0.09 / $0.30**; gemma-4-31b-it $0.09 / $0.34; mistral-small-3.2-24b $0.094 / $0.25; qwen3.6-27b $0.32 / $2.70; qwen3.6-35b-a3b $0.15 / $1.00; gpt-oss-20b $0.018 / $0.09; gpt-oss-120b $0.15 / $0.60; llama-3.3-70b $0.10 / $0.32; glm-5.3 $0.84 / $2.64; kimi-k3 $3 / $15; deepseek-v4-pro-0813 $0.462 / $1.386; nemotron-3-ultra $0.60 / $2.40; qwen3.5-397b $0.55 / $3.50 (these OpenRouter lines are input / output only) | `:batch` variants listed where offered | VERIFIED: the public JSON `https://openrouter.ai/api/v1/models`, pulled 2026-09-23T15:37:08Z, 457 models. Copy at `sources-economics/prices/openrouter-models-2026-09-23T1537Z.json` (sha `9c6e2f1a…`) |

**Our bills of record for frontier arms (RECORD):**
- **Exhibit 40** (`~/rs-two-frontiers-2026-09-05/results/bill.json`, window 2026-09-05 14:13:33Z → 18:15:00Z):
  - GPT-6 Astra, **metered**: 1,572,354 prompt tokens (559,416 of them cached) and 142,086 completion tokens (92,960 of them reasoning), priced from `https://developers.openai.com/api/docs/pricing` as read 2026-09-05. That comes to **$17.79**, or $22.83 had nothing been cached, over 235 records: **$0.0757 a record**.
  - The Claude Fable 5.1 CLI, **no-figure-held** (a subscription): the CLI's own list-rate estimate was **$31.44 over 242 records**, "not a receipt".
  - kimi-k3: **metered $5.25** against a $5 cap that did not fire.
  - Five shelf seats: **plan-included, $0.00\***.
  - The local seat: **own-silicon**, "no dollar is charged and none is invented".
  - Metered total **$23.05**. an operator: "WELL worth the $23.05".
- **Exhibit 41** (`results/cove/bill.json`, 2026-09-06 13:27:57Z → 17:03:52Z): metered **$0.7652** (Astra $0.5183, kimi $0.2469), against a registered estimate of $0.703 and a ceiling of $0.75, which it went over "by a cent and a half, and printed rather than rounded".
  - The CLI arm printed **two** estimates: $0.6461 (the transport's own figure) and $1.5590 (the registered basis).
  - The balances at the time: OpenAI key $17.71, ollama.com $17.17 (an operator, 09-06; memory `two-frontiers-train.md`).

**The precedent, confirmed.** `no-cloud-ai-in-live-products.md` rules on *live products*. Benchmarks that call frontier models ran under an operator's asks and standing goes:
- exhibit 40: an operator's ask, 09-04/05;
- exhibit 41: D-20260906-27, "you have my approval to run it".

What made those runs acceptable, and what this train should reuse:
- sealed transports (`claude -p --safe-mode`, pinned channels, telemetry off);
- per-leg **caps set below the balance**;
- an **egress receipt** (the bench children reached only the API hosts);
- the **cost-state vocabulary**;
- the served model recorded per call.

**Exhibit 40's transport finding also matters for money.** The CLI silently served `claude-opus-5` in place of Fable on 19 of 19 corrupt-corpus calls (MODEL-FALLBACK). A per-call price row has to be keyed on the model **actually served**, not the model asked for.

**A decision owed.** Exhibit 40 had "no API key, so no second road" for Anthropic. The standing instrument will want **metered receipts** for the Claude comparator. Either an operator adds an Anthropic API key, with a cap below its balance, or the Claude row stays a labelled list-rate estimate forever. That is an operator's call.

---

## 4. The proposed formula: cost per correct answer

Notation: task set T (frozen), card k, model m, reading date D. A task set has N items, of which n_correct are correct. Format failures score as wrong. Instrument artefacts (transport errors, harness faults) are excluded **and counted in a column of their own**.

### 4.1 Local, own silicon

For each item i, from the harness's rows:
- **E_i**, the board energy. It is the trapezoid integral of `power.draw` over the item's wall time, prefill and decode together, read by a sampler pinned to the card's UUID. This is the Jev harness's `Watts`/`Boards` (1 Hz).
  - The gross figure is the raw integral.
  - The net figure is the raw integral − P_idle,pre × t_i, where P_idle,pre is the ten-second reading taken before the arm (`idle_reading`).
- **E_i at the box**: per the house "at the meter" rule, box idle × t_i + (E_i − P_card_idle × t_i) ÷ 0.9. Or, better, from the UPS series over the whole arm, minus a matched idle window, divided across the arm's items. The 60 s UPS cadence is too coarse for one item, but fine for an arm.
- **t_i^busy**: the item's own seconds at concurrency 1; at concurrency c, the arm's wall time × c ÷ N, or tokens ÷ aggregate tok/s.

The four cost components:
- **Fuel** (electricity): `fuel_i = E_i[J] / 3.6e6 × p_e`, with p_e in $/kWh from EIA, month stated.
- **Capital**: `cap_i = (C − R) × CRF(i, n) / (8766 h × u) × t_i^busy / 3600`.
- **Idle** (fixed O&M): `idle_i = P_idle × (1 − u) / u × t_i^busy / 3600 [h] / 1000 × p_e`. Use card idle for card rows and box idle for box rows.
- **Total**: `cost_per_correct(T, k, m, D, u, c, layer) = Σ_i (fuel_i + cap_i + idle_i) / n_correct`.

**Three bands print for every reading:**
1. **fuel only**: what the meter sees, i.e. the marginal cost of an answer;
2. **levelised at U100, U25 and Umeasured**: what owning costs;
3. **the break-even u\* and the break-even purchase price C\***: solve the levelised cost equal to the rental price of the same model, and to the frontier's.

**Every input and its source:**
- p_e: EIA table 5.6.A.
- C and R: the hub price rows (§3.2), or the break-even.
- n and i: the registered convention (§3.4).
- u: the ledger and history, with the >100 W rule stated.
- P_idle: the harness's pre-arm reading, or a roster idle row.
- E_i and t_i: the harness.
- The 0.9 PSU factor: the house rule, marked as assumed.

### 4.2 Frontier, metered or list-rate

For each call j, **including retries, failed calls and format-failed answers** (the tokens were billed):
```
cost_j = (in_uncached × p_in + in_cache_write × p_cw + in_cache_read × p_cr + out_total × p_out)
         × m_geo × m_batch + tool_fees_j
out_total = visible output + reasoning tokens (billed as output)
cost_per_correct(T, model, D) = Σ_j cost_j / n_correct
```
- Prices come from the provider's page, as a block in the roster: `{url, read_date, input_usd_per_mtok, cached_input_usd_per_mtok, output_usd_per_mtok}`. This is the exhibit 40 shape.
- The token counts are the **endpoint's own usage counters**. For the CLI, "prompt = input + cache_creation + cache_read" (exhibit 40, A8).
- Cost states carry over unchanged from exhibits 40 and 41:
  - **metered**: a receipt;
  - **no-figure-held**: a subscription, whose list-rate estimate prints labelled and is never added into a total;
  - **plan-included**: $0.00\*, with a second row at a named public list price (OpenRouter) labelled "what these tokens list at";
  - **own-silicon**: gains the derived dollars of §4.1, labelled as derived and never as a charge.
- **Batch rows:** the task sets are not interactive, so a batch-priced frontier row (×0.5) is the fair comparator for a card that runs queued work. Print both.

### 4.3 The metrics each reading carries (the economics lens only)

For each task × card × model × date:
- accuracy with its denominator;
- format-failure count, and instrument-artefact count kept separate;
- latency (TTFT and total) and tok/s at c = 1 and at the best c;
- J per correct at the board, and at the box where measured, with the layer named;
- $ per correct in the three bands (§4.1);
- $ per question at a fixed accuracy, which is Epoch's unit and equals $ per correct × accuracy for a single model;
- **the ratio**: best local score per card ÷ best frontier score, per task;
- the **cost ratio**: local levelised ÷ frontier list, at U25 and at Umeasured;
- **u\*** and **C\***.

**Across readings:** the point-to-point factor per bar (×/year) between dated readings. **Do not fit a slope to two or three points.** The threshold ladder's "cheapest (model, card) that clears bar B at date D" is the direct analogue of Epoch's C(t, a).

### 4.4 A worked ILLUSTRATION: arithmetic on existing receipts, not a result

The model is gemma4:26b at one stream, on one GPQA-shaped question as Epoch measured it for `gemma-4-26b-a4b-it` (minimal effort): 1,370.2 output and 253.2 input tokens (Epoch data). Decode energy only; the prefill (253 tokens) is ignored; 18.34 ¢/kWh.

**Fuel only, board** (J/1k tokens from the RECORD rows):

| card | J/1k tokens | fuel per question |
|---|---|---|
| 3090 at a 250 W cap, 126.12 tok/s (`m5-3090.json`, the mistral page; at 350 W: 1,823 J/1k, 136.49 tok/s) | 1,632 | ≈ $0.000114 |
| laptop, boost | 787.76 | ≈ $0.000055 |
| laptop, 95 W | 672.50 | ≈ $0.000047 |
| workstation card at 420 W, 205.85 tok/s (`m5-6000-live.json`, the live seat; floor included) | 887 | ≈ $0.000062 |

**Renting the same model:** OpenRouter's list today gives 253.2 × $0.09/M + 1,370.2 × $0.30/M ≈ **$0.000434**. Epoch's own figure is $0.000484, at OpenRouter prices as of 2026-08-03.

**Levelised on a 3090** ($1,449.99, 3 years, no resale, i = 0, card idle 23 W, 250 W cap, 10.9 s a question at 126.12 tok/s):

| utilisation | $ per question | against renting |
|---|---|---|
| U100 | $0.00028 | 0.6× the rental |
| U25, six hours a day | $0.00082 | 1.9× |
| 6 % busy | $0.0031 | 7.1× |
| 2.6 % busy | $0.0070 | 16× |

**Break-even u\* ≈ 54 %** (single stream, board watts, the primary convention).

**What that shows:**
- Fuel is cheap. Owning costs **capital divided by utilisation**, and our utilisation is in single digits.
- Resale moves this a lot. At the August 2026 used price the net capital is roughly a quarter of the price paid, and u\* falls sharply.
- So does concurrency. Two streams roughly halve the capital per answer.
- **Pre-register the uncomfortable direction.** With the primary convention and the measured duty, owned hardware may lose on cost to renting the same open weights. It will very likely beat the frontier on cost only where a cheap open model clears the bar. The case for owning then rests on what it buys other than cost: sovereignty, privacy, no meter, and staying up in an outage. The page must be willing to print exactly that (self-refutation honoured).

---

## 5. The utility frame: intelligence from the grid, or made on site

**The mapping.** Most of it is analogy. The last row is literal.

| electricity | intelligence |
|---|---|
| the kWh, a unit of *useful* energy | **the correct answer at a fixed bar**. Not the token. A token is closer to "volts": a count of work that varies 6× between models on the same question (Epoch's data: 1,370 tokens for gemma-4-26b against 8,250 for Qwen3.6-27B on GPQA). |
| the utility tariff, ¢/kWh | the API list price, $/M tokens → $ per correct |
| the central station (Insull's load diversity, and the capacity factor that makes it cheap) | the cloud: many users' requests batched onto one GPU, and high utilisation |
| the **isolated plant** (a factory's own dynamo), later the diesel genset or rooftop PV | the owned card |
| capacity factor | **utilisation u**. The estate measures 2–6 % on the busiest card. |
| LCOE (NREL's sLCOE) | **LCOI**, the levelised cost of intelligence (§4.1) |
| fuel cost × heat rate | the electricity price × joules per answer. **The recursion: the fuel for on-site intelligence *is* grid electricity.** |
| flat-rate lamp contracts before the meter (early utilities); net metering (today) | the subscription plan ($0.00\*); selling idle GPU hours back on a marketplace (an idea only, and one that sits badly with the ethos) |
| reliability, islanding, running through an outage | on-prem: nothing phones home, the UPS holds 60–90 minutes, no vendor can reprice or deprecate the model |

**The history worth telling. Read the sources before quoting.**
- Early electrification was a contest between **isolated plants and central stations**. The central station won on **load diversity**: many customers' peaks don't coincide, so shared generators run at a high capacity factor. Sources: Thomas P. Hughes, *Networks of Power* (1983); David Nye, *Electrifying America* (1990); Hirsh 1999.
- The price of electricity fell for eighty years as scale grew. Then, around 1970, the learning curve stalled; Hirsh calls it "technological stasis". Epoch's line ends in 1973 for that reason.
- **The utility-computing framing** is old: John McCarthy's 1961 MIT remark that computing may someday be organised as a public utility, and Nicholas Carr's *The Big Switch* (2008), on data centres as the isolated plants of computing. Both are well known; verify the exact wording before quoting.
- **This workshop is, deliberately, an isolated plant.** The piece's honest question is the one the isolated plants faced: what do you get for running your own generator, and what does it cost you at the capacity factor you actually run?

**The electricity line on the same chart.** The data exist, pinned. Draw three series on one log axis (index = 100 at the first point, real $):
1. **Epoch's LLM line**: index 100 → 0.00054, from 2021.88 to 2026.67, with the 2021–23 segment marked as extrapolated.
2. **Electricity**, drawn two ways: Epoch's comparator (1892–1973, 1.05× a year, real), and the **same series carried on to 2026** from Epoch's own sheet (nominal to 2025, real to 2003) plus EIA 2026. The second shows the fuel price going flat, then rising.
3. **Our dots**: $ per correct on each reference card at each dated reading, drawn as **our own index, never on Epoch's y-axis values**. The units differ: our task sets are not their benchmarks. The honest overlay is *rates*, meaning slopes. Levels are only placed side by side via the cross-walk (§7, item 1), labelled.

The dataviz skill applies when the chart is built. Every point carries its source and window, per an operator's show-the-data trait and the units-at-point-of-use law.

---

## 6. What this lens recommends the plan carry (the economics only)

1. **Register the formula of §4** and the convention of §3.4 in the prereg: primary choices, sensitivity rows, the layer and concurrency named per row. A reading may be recomputed later under a new convention only if it is **printed as a new row** and the old one is kept. That is the reproducible-rule law.
2. **Freeze prices per reading:** the EIA month, the provider page and read time, the OpenRouter snapshot sha. Never reprice history silently.
3. **Print the break-evens (u\*, C\*) as the headline economic numbers.** They need no private price, they move over time (that is Bench 3), and they are the honest answer to "does owning ride the curve?"
4. **Pre-register Bench 3's prediction** from §2:
   - the owned card's $ per correct at a fixed bar falls at about the **algorithmic rate** (≈3× a year, Gundlach) plus runtime gains;
   - the cloud's falls at 5–13× a year;
   - so the cost ratio drifts **against** ownership unless utilisation rises.

   If the data refute this, print that. If they confirm it, print that too.
5. **Keep the cost states** and extend own-silicon to "derived $, three bands".
6. **The 96 GB card's economics are marginal.** It is bought, it is carrying 24/7 seats, and it has about 30 GiB free. Charge the bench net energy only, with zero capital (primary), and give C\* as the publishable figure. Any seats-down window is an outage (the 24/7 law), and that is not an economics question.

---

## 7. Gaps, decisions owed, quick wins, and what to plan for

**Quick wins, now, with no GPU:**
1. **The cross-walk** (about 30 minutes, the lead or a lane). For `gemma-4-26b-a4b-it`, `gemma-4-31b-it`, `mistral-small-3.2-2506` and `gemma-3-27b-it` (all four are in Epoch's data and in our records): multiply Epoch's mean tokens per question by our measured J/token per card to get joules and fuel $ per Epoch question. Print it beside Epoch's cost per task: rented GPU for mistral-small-3.2, OpenRouter for the Gemma models.
   - Label it "their accuracy and token counts, our joules, not a run of their benchmark".
   - This puts our three cards next to Epoch's line **today**, before any bench runs, and is the natural first figure for the write-up.
2. **A daily price snapshot** (about 20 minutes to add): pull `https://openrouter.ai/api/v1/models` (public, no key) once a UTC day into `bench/cost-of-intelligence-*/prices/`, with a sha per file. It becomes our own grid-tariff series, for every open model we can run and every frontier comparator, with receipts. Epoch had to use the Wayback Machine for the same thing.
   - Add a monthly re-read of the Anthropic and OpenAI pricing pages at the pages themselves.
3. **EIA tomorrow, 2026-09-24:** re-read table 5.6.A (July 2026). Register "the latest month at the reading's date".
4. **Pin Epoch:** done. Commit `a89f968b…`; files and shas under `recon/sources-economics/`. **Ask Epoch, or decide, whether the data repo's missing licence blocks republishing the CSVs in a kit.** Until then, cite and link only.

**Owed from an operator, as decisions (render each in chat with options):**

5. **Which 3090 cost what, by UUID** (<withheld: GPU UUID of board E> · <withheld: GPU UUID of board A> · <withheld: GPU UUID of board F>), and which is the reference card. ~~Without that the historic-cost row is ambiguous between $1,449.99 and $1,899.~~ (Struck 2026-09-23: the $1,899 pair never arrived, so there is no ambiguity — the only card-only line is the LISTING, and the UUID ledger fills the private record only; PLAN-v2 ⧖-11.)
6. **The 96 GB card's price:** (a) the price paid, private, with a ruling on its public form; (b) a product-page reading as replacement cost; (c) the break-even only (recommended).
7. **The Claude comparator:** add a metered Anthropic API key with a cap, or stay with the subscription CLI's list-rate estimate, which is no-figure-held forever.
8. **The amortisation convention** (§3.4): 3 or 5 years, resale 0 or market, i = 0 or 5 %. The recommendation is 3 / 0 / 0 as primary, with the others as sensitivity rows.
9. **The brief's "class only" line versus D-20260917-05.** Confirm that full hardware names are fine on this piece. The pet names stay fenced in either case.

**Instrument gaps:**

10. **No wall reading for the laptop.** It has no UPS, and the ledger covers the card only. A plug-in meter would close it. So would a battery-discharge proxy (upower on battery), which is marked as a proxy.
11. **UPS output is not the wall.** A plug-in meter upstream of one UPS, read once, would calibrate the loss. The what-it-costs piece has owed this since 09-12.
12. **Utilisation is read at a 15 s cadence with a >100 W rule.** That is fine for duty. For the formula's u, register the rule and the window. Consider reporting both "any activity" (utilization.gpu > 0: 26 % of ticks on the 96 GB card) and ">100 W" (2.6 %), because they differ tenfold.
13. **Cooling:** the room's air conditioning is unmetered (what-it-costs says so). Say so on the page. Datacentres report PUE; a home's summer "PUE" is higher and unknown.

**Plan for:**

14. **Bench 4, "frontier enough"** (sketch only): cost per *accepted* dev task, meaning a fix that passes the repo's tests on estate repos after the model's cutoff, at a measured share of the cloud frontier. The unit is $ per merged fix, and the utility analogy is "enough power for the workshop".
15. **The flat-rate lens:** effective $ per correct for plan-included subscriptions, i.e. the plan price per month ÷ correct answers actually drawn per month. This is the flat-rate tariff analogy, but it needs the plan prices, which are private and an operator's.
16. **Real versus nominal:** once there are two or more readings a year apart, add a CPI-U deflated column (FRED CPIAUCSL) to match Epoch's real-dollar basis.
17. **Keep the curves apart.** Bench 1's backfill (2023 models on 2026 cards) measures *model generations at fixed hardware and fixed 2026 electricity*. It is **not** what 2023 intelligence cost a 2023 owner. Label it so, or the page will draw a false curve beside Epoch's market curve.

---

## 8. Receipts: where every number above lives

**Pinned sources:** `recon/sources-economics/SHA256SUMS` covers the eight Epoch files, the OpenRouter snapshot (2026-09-23T15:37:08Z) and the EIA page as read today.

**Estate and hub records read:**
- `~/s2s-research-hub/site/hardware/rows.json`: the price and idle rows.
- `~/s2s-research-hub/tools/compare_cards_outside.json`: the 96 GB card's price rule.
- `~/s2s-research-hub/site/a-short-history-of-mistral/index.html`: the precedent for the electricity-only $ per M, and the no-invented-price sentence.
- `~/estate/docs/article/what-it-costs-to-run-v10.md`: duty, idle, and the "at the meter" rule.
- `~/estate/DECISIONS.md`: D-20260917-05 (hardware named in full).
- `~/rs-two-frontiers-2026-09-05/results/bill.json`, `results/cove/bill.json` and `prereg/rosters.json`: the frontier bills and the pricing block.
- `~/estate/bench/jev-2026-09-21/run.py`: the energy integration, idle subtraction and `joules_per_decision`.
- `~/.local/state/estate-ledger/power/days/*.json` and `~/.local/state/estate-watch/history/2026-09-2{0,1,2}/`: duty, computed today.
- Memory: `laptop-5090-bench-2026-09-21.md`, `price-series-rig-piece.md`, `two-frontiers-train.md`, `no-cloud-ai-in-live-products.md`, `no-hardware-specifics-published.md` (superseded on card names by D-20260917-05), `hardware-roster-train.md`.

**Web sources read today, all 2026-09-23 UTC:**
- epoch.ai (the post, and the 2025 data insight);
- github.com/droodman/inference-cost, through the API and raw files;
- platform.claude.com pricing;
- developers.openai.com pricing (fetcher summary);
- openrouter.ai `/api/v1/models`;
- eia.gov EPM table 5.6.A;
- arxiv.org/abs/2511.23455.

**SECONDARY only:** Appenzeller (a16z), the Stanford AI Index figures, the RTX PRO 6000 market prices, the hyperscaler depreciation lives, and NREL's formula text.