# Growth slides: compute, inference and data

For Jacob Steeves' keynote, Exploit Summit, Montreal, 28 September 2026. Three "up and to the right" slides built 24 September, four days out, by two workers in parallel. Each slide is drawn from `content/data.json` in the deck repository (branch `slides-site`) and lives at https://bittensor-talk.pages.dev. Each worker owns one section below; append, do not overwrite.

Terms used in this file. **Netuid**: the number of a subnet on the chain (SN75 is netuid 75). **Time series**: the same measure read on several dates. **Lower bound**: a number that is certainly at least this large. **Stock and flow**: a stock is how much exists; a flow is how fast it grows. **Erasure coding / parity shards**: a way to store a file as pieces with extra pieces, so the file survives when some nodes fail; the extra pieces use disk but are not customer data. **Egress**: data leaving a storage provider; the big clouds bill for it.

## Slide 3. Data (the data layer)

Built 24 September 2026. Live as slide 3 at https://bittensor-talk.pages.dev/#/3. PNG: `media/growth-data.png` (3200×1800). Data: `content/data.json → growthData`. Placed after the library row (slide 2) because the compute and inference growth slides were not merged when this was built; when they land, move it to follow the inference slide (`scripts/generate-slides.mjs` LAYOUTS key and the `## 3. Data` block in `content/slides.md`). Every slide from 4 on is one higher than on 23 September; every cross-reference in `slides.md` and the README was renumbered.

### Headline (the one sentence under the figure)

"Tonight: 730 terabytes stored, 2.7 million molecules in the open, 55 billion rows, a terabyte moved in three minutes forty."

Small grey title above the figure: "Data · Hippius SN75 · NOVA SN68 · Data Universe SN13 · Beam SN105 · to 24 Sep 2026". The dates sit on the panels: each panel shows its first value and month at the left and its last date at the right.

### Chart form: four panels, not one axis. Why.

The four measures are in four units: bytes, molecules, rows, bytes per transfer. They start on four dates (Aug 2024, May 2025, Apr 2026, Apr 2026). One shared log axis would put 55 billion rows a hundred thousand times above 730 terabytes and the eye would read Data Universe as the biggest thing on the data layer, which is the reverse of the truth (its collection rate has fallen about 98%). Normalising each line to its first point would hide that same stall and would make Beam (2,240× in three months of test loads) the steepest line for the least commercial reason. So: four small panels, each with one latest number in large type, one measure line, one black sparkline on its own linear scale from a zero hairline, a dot per point (hollow when unverified, larger for the last), and the start value and end date in grey. Black, white, one grey. No legend. One annotation sentence. On a phone the row becomes two by two. Checked at 1600×900 and 390×844: no overflow; slides 2, 4 and 20 re-checked after the renumber.

### The four series

| Panel | Measure | First point | Last point | Multiple | Points | Source | Status |
|---|---|---|---|---|---|---|---|
| Hippius SN75 | bytes stored on nodes (decimal TB, Arion era) | 43.6 TB · 1 Apr 2026 | 729.8 TB · 24 Sep 2026 | 16.7× in 176 days | 8, all with Hippius-chain block numbers | hipstats.com/api/storage-usage, read via the explorer 24 Sep 2026 | Verified; independent explorer run by the team |
| NOVA SN68 | molecules in the open archive (cumulative) | 38,427 · 2 May 2025 | 2,675,535 · 24 Sep 2026 | 70× in 510 days | 11 | Commit log of huggingface.co/datasets/Metanova/Submission-Archive: 13,832 commits, each "Added N molecules from epoch E to <target>.csv", summed by day | Verified; public, reproducible |
| Data Universe SN13 | rows collected (cumulative) | 1.34B · 2 Aug 2024 | 55B+ · 2026, undated | 41× | 5 (last unverified, drawn hollow) | @MacrocosmosAI posts 2 Aug 2024, 4 Oct 2024, 28 Jan 2025; substack 4 Apr 2025; datauniverse.macrocosmos.ai and docs.macrocosmos.ai read 24 Sep 2026 | Team-stated throughout; the 55B is undated |
| Beam SN105 | largest single orchestrated transfer (GB) | 0.5 GB · 2 Apr 2026 | 1,120 GB · 18 Jul 2026 | 2,240× in 107 days | 5 | @b1m_ai posts, one per point | Team-run test loads; no customer traffic |

#### Hippius SN75, the points (used storage summed over active nodes; hipstats `total_storage`, decimal terabytes)

1 Apr 2026 43.6 (block 5,165,400) · 1 May 72.8 (5,583,083) · 1 Jun 141.8 (6,029,496) · 1 Jul 241.4 (6,461,492) · 1 Aug 318.3 (6,864,689) · 1 Sep 343.4 (7,368,603) · 15 Sep 455.6 (7,555,883) · 24 Sep 729.8 (7,692,625). Network capacity the same day: 2,435.6 TB (2.44 PB), 1,680.7 TB free, 311 nodes all online (Europe 256, North America 51, Asia 4).

Other Hippius series from the same explorer, not drawn: files pinned 223.7M (1 Apr 2026) → 4,773.2M (24 Sep 2026), which after Arion are shards, so do not say "files" on stage; accounts 29,818 (23 Sep 2026; 991 on 1 May 2025, but 2,142 → 19,317 between 1 Aug and 1 Sep 2025 looks like a promotion, so do not call them customers); credits bought on-chain 73,819 in total from 8 Apr 2025 to 24 Sep 2026 (1 credit = $1 per hippius.com), 35,244 consumed, with the biggest months Aug 2025 (21,509), Sep 2025 (12,250), Apr 2026 (11,756), May 2026 (10,520); consumption jumped to 27,769 in September 2026 alone, in step with the storage jump. Cross-checks: the team said $23,000 of storage revenue on 12 Oct 2025 (x.com/Hippius_cloud/status/1977249628731908398) and TAO Daily said $46,600 on 27 Oct 2025; the explorer's cumulative purchases were about 34,000 credits by mid October 2025, between the two. The 100 TB of egress in a month (8 May 2026) stays on slide 2's card.

The IPFS era, not on the line: used storage 28.6 TB (1 Oct 2025) → peak 127.9 TB (24 Nov 2025) → 101.8 TB (1 Feb 2026), stored as five copies of every file; Arion rollout began 3 Feb 2026 (x.com/Hippius_cloud/status/2018799239619342598), the explorer fell to about zero through February and March, and the analytics page carries the team's note: "Why the drop? No worries, zero data lost. Same files now use about 40% less space." The Arion line therefore starts 1 April 2026 and is not comparable byte for byte with the IPFS era.

#### NOVA SN68, the points (cumulative molecules added to the public archive)

2 May 2025 38,427 · 1 Jul 261,173 · 1 Sep 552,748 · 1 Oct 865,232 · 1 Nov 1,786,920 · 1 Jan 2026 2,066,403 · 1 Apr 2,126,179 · 1 Jul 2,198,351 · 1 Aug 2,222,850 · 1 Sep 2,467,915 · 24 Sep 2,675,535. Nanobodies in the same archive, not drawn: 204 (1 Apr 2026) → 108,385 (24 Sep 2026); the team's 82,575 (26 Aug) falls between the archive's 62,801 (1 Aug) and 88,807 (1 Sep), so the archive tracks the team's count for nanobodies. For small molecules it does not: the team counts 11,130,975 submitted (x.com/metanova_labs/status/2092646824725406003, 26 Aug 2026) and "roughly 4.8 million molecules against 7,000 targets" by 9 May 2026 (cryptobriefing.com); the archive holds the ones the validators accepted and published per epoch, a lower bound. Shape to say aloud: 1.8M by November 2025, then almost flat for eight months (about 10,000 a fortnight), then 450,000 added since 1 August after the 14 Aug 2026 upgrade to 20× more Boltz-2 evaluations per miner (x.com/metanova_labs/status/2088264380488012175). Launch: March 2025 (taoprotocol.org). First molecules and nanobodies in a wet lab: announced 26 Aug 2026; nanobody production with Yalotein from 28 Jul 2026. Blueprint search space 61.1B → 678.5B molecules on 23 Sep 2026 (that is a search space, not molecules screened; never put it on the panel).

#### Data Universe SN13, the points

2 Aug 2024 1.34B rows (x.com/MacrocosmosAI/status/1819354978504278256; the ten largest X and Reddit datasets on Hugging Face) · 4 Oct 2024 18B (status/1842252141966000406: "we were off by a factor of ten. 18bn") · 28 Jan 2025 27.7B (status/1884305311223406918) · 4 Apr 2025 40B (macrocosmosai.substack.com/p/data-universe-how-to-collect-real) · 2026 55B+ (datauniverse.macrocosmos.ai: "55+ billion rows", "up to 80 million new rows added every single day"; docs.macrocosmos.ai: "over 55 billion scraped posts and comments"; undated, read 24 Sep 2026, drawn hollow). Also dated but not drawn: 23 Jul 2024 702.1M Reddit rows on Hugging Face (Reddit only); 11 Oct 2024 about 350M rows a day; 8 Jul 2025 "another 20bn rows added since you last looked" (status/1942593700066099259), which would put the total near 60B and above the site's 55B; 23 Oct 2025 the team said it had moved to private storage and stopped uploading to Hugging Face (status/1981396190622015527: 12.37% of all Hugging Face datasets over 100M rows, 51,476 real user downloads); 3 Jun 2026 "174,932,479 data rows within May" (x.com/Data_SN13/status/2062208097695781083), about 5.8M a day, so the site's "up to 80 million a day" is a ceiling, not a rate; 10 Jul 2026 233 miners, 140 datasets; 2 Sep 2026 On-Demand API 63k jobs in 15 days. Dashboard sn13-dashboard.api.macrocosmos.ai on 24 Sep 2026: network bucket 7.58 TB, 253,968 files, 244 miners with data, 239 earning. Emission: about 0% this week under the gate (ecosystem-as-empowerment.md).

#### Beam SN105, the points (largest single orchestrated transfer)

2 Apr 2026 500 MB in 21–29 s (recalled in x.com/b1m_ai/status/2088647455512355102) · 4 Jun 10 GB, R2 → internet → R2, 21 s (status/2062329501049135318) · 27 Jun 50 GB as five concurrent 10 GB transfers, about 51 s (status/2070727855185478122) · 12 Jul one 107 GB dataset, 2 min 16 s (status/2076184463801737591) · 18 Jul 1.12 TB, 3 min 40 s, hundreds of concurrent streams (status/2078327748842463530). Sustained load, not drawn: 15 Aug "500 GB in parallel every 10 minutes", 3 TB an hour; 23 Aug "500 GB to 1 TB every 10 minutes" for over a week across Hippius, R2 and S3; 13 Sep "about 1 TB every 30 minutes" (status/2099230791850971514), 48 TB a day, which is lower than the August rate, so a throughput line would end downward. First subnet-to-subnet transfer R2 → Beam → Hippius, 107 GB in about 6 minutes, 29 Jul 2026. Live on mainnet since 10 Mar 2026 (six-month post, 10 Sep 2026). Dashboard API data.b1m.ai/api/dashboard/metrics, 24 Sep 2026: 26,480 worker keys, 25,563 registered, 267 active, 211 active orchestrators, 30,031,973 tasks, 29,006,111 completed, 13,189 failed. Beam is live and measurable, so it stays; but every byte on the panel is the team's own load. Their words: "In preparation for real client workloads" (23 Aug) and "The next phase is adoption" (10 Sep). No paying customer is evidenced (ecosystem-as-empowerment.md, 17 Sep). Say "proving ground".

### Method

1. Hippius: hipstats.com is a Next.js explorer whose `/api/*` routes refuse plain HTTP clients (403), so the history was captured from the explorer's own requests in a headless Chrome session (Playwright): `storage-usage`, `files-growth`, `account-growth`, `daily-node-count`, `purchased-consumed-credits`, each with `limit=2000`, daily from April 2025. The first-of-month reading was taken for each point; the block number on each point is the Hippius chain block the explorer recorded. The team's own RPC (`get_active_nodes_metrics_by_type`, docs.hippius.com/blockchain/api) would give the same sum live.
2. NOVA: `GET https://huggingface.co/api/datasets/Metanova/Submission-Archive/commits/main?limit=1000&p=N` for N = 0…13; commit titles parsed with the regular expressions `Added (\d+) molecules` and `Added (\d+) nanobody`; summed by commit date; four test files and their deletions ignored. The archive was created 2 May 2025 at 00:56 UTC.
3. Data Universe: X API v2 full-archive search, `(from:MacrocosmosAI OR from:Data_SN13) (rows OR billion OR bn OR dataset OR "Hugging Face")`, 33 posts; substack `datePublished`; site text.
4. Beam: X API v2 full-archive search, `from:b1m_ai (TB OR GB OR terabyte OR gigabyte OR workers OR orchestrators OR milestone OR transfer OR moved)`, 30 posts; dashboard metrics JSON.
5. X reads in total: 150 of the 200 allowed (Hippius 38, Beam 30, NOVA 34, SN13 33, Minos 15), bearer token from 1Password vault Arbos, never printed.
6. Labels are computed from the points by `compactValue` in `scripts/generate-slides.mjs`, so a re-pull changes the big numbers without editing text.

### Caveats, to say before a critic does

1. Hippius's bytes are raw disk across nodes after Arion erasure coding: they include parity shards and are not deduplicated customer bytes. Do not say "730 terabytes of customer files". Say "stored on the nodes".
2. Hippius went from 456 TB to 730 TB between 15 and 24 September, and capacity from 2.2 to 2.4 PB, with credit consumption up twenty-fold in the same month. It is on the explorer; the team has not published why. If it is a migration or a test load it may fall back. Re-pull on 27 September and, if it has fallen, draw the fall.
3. The Hippius line begins 1 April 2026 because the IPFS-era bytes (five copies) are not comparable. Do not say "seventeen times since launch"; say "since April".
4. NOVA's 2.7M is a lower bound from the public archive; the 11.1M is the team's count of everything submitted and is not independently checkable. The archive was almost flat from December 2025 to July 2026.
5. Data Universe's 55B is the team's undated site number; their own 8 Jul 2025 post implies about 60B. The collection rate fell from about 350M rows a day (Oct 2024) to about 5.8M a day (May 2026), the subnet left Hugging Face, and the emission gate pays it about nothing. The line is a stock that has stopped growing fast. It is drawn hollow at the end for that reason and the notes say so.
6. Beam's five points are demonstrations run by the team, not customer traffic, and its sustained throughput fell between August and September. It is on the slide because Jacob names it as a layer and it is live and measurable; the notes say "proving ground, not business".
7. None of the four is a paying-customer series. The closest is Hippius's credits purchased (73,819 credits since April 2025, about $74K), which is small and is in the notes, not on the screen.
8. Every number on the panel is in a count or in bytes; nothing is a dollar or a token price.

### Considered and not drawn

- **Minos SN107** (synthetic genomes and variant-calling evaluations): registered 2 Feb 2026, "officially live" per the team's 2 May 2026 post; "27,000 evaluations on chromosome 20 in the first 6 days" (6 May 2026, x.com/theminos_ai/status/2052165296375689292); "more than 37,000 variant-calling evaluations" in 11 days on chromosome 21 (8 Jun 2026, status/2064015596941959598); one challenge genome every 72 minutes, about 20 a day, stored on Hippius (SN75) as backup; a 200,000-person synthetic schizophrenia cohort (a third-party post Jacob retweeted in 2026; not team-confirmed here). Two dated points, both rates over different windows, so no honest cumulative line without the team's help; api.theminos.ai has no public counter and mcp.theminos.ai needs an MCP client. Kept to the notes of slide 3. Mamad Anghari speaks on Day 2.
- **Hippius files pinned** (224M → 4.77B): after Arion these are shards; a "files" line would overstate.
- **Hippius accounts** (29,818): a jump of 17,000 in one month in 2025 looks promotional.
- **Beam throughput per day**: falls from 72 TB (Aug) to 48 TB (Sep).
- **Data Universe rows per day**: the honest series and the one that would embarrass the panel; said in the notes instead.
- **Chutes/Gradients data products, Desearch SN22**: inference and search, not the data layer as Jacob names it.

### On screen (the line in `content/slides.md`)

White. Small grey title. Four panels in a row, each with a hairline across the top and its own scale: "Hippius · SN75 / 730 TB / bytes stored on nodes / [line] / 44 TB · Apr 2026 … 24 Sep 2026", "NOVA · SN68 / 2.7M / molecules in the open archive / [line] / 38K · May 2025 … 24 Sep 2026", "Data Universe · SN13 / 55B+ / rows collected / [line, hollow last dot] / 1.3B · Aug 2024 … 24 Sep 2026", "Beam · SN105 / 1.12 TB / largest single transfer / [line] / 500 MB · Apr 2026 … 18 Jul 2026". One sentence beneath: "Tonight: 730 terabytes stored, 2.7 million molecules in the open, 55 billion rows, a terabyte moved in three minutes forty."

### Speaker beats (Jacob's voice; also in the deck notes)

- 2:30. One breath a panel, left to right. The data layer is what a mind eats and where it keeps what it learns. Four markets, four units, so four panels on their own scales; a byte is not a molecule is not a row.
- Hippius: seven hundred and thirty terabytes on three hundred and eleven nodes tonight; forty-four on the first of April. Egress free. Every byte on a public explorer with a block number.
- NOVA: two point seven million molecules in an archive anyone can download, published one epoch at a time since May last year; the team counts eleven million submitted; the first are in a wet lab since August. Micaela Bazo speaks at 11:20; hand it to her.
- Data Universe: fifty-five billion rows, once the largest open social dataset on Hugging Face. Same breath: they took it private last October, the gate cut them to almost nothing, the flow is a trickle. The stock exists; the flow does not.
- Beam: six months old; five hundred megabytes in April, a terabyte in three minutes forty in July, a terabyte every half hour all September. All of it Beam's own load. "Proving ground", not "business".
- Caveats in one breath (see above), then: "Four markets for data, each with a door open tonight and a number you can check." Advance on "check".

### On 27 September

1. Hippius: open hipstats.com, read "Used" on the home tile (or the `storage-usage` row for the day) and replace the last point in `growthData.panels[0].points` with the value, date and block. If it has fallen from 730, keep the fall.
2. NOVA: re-run the commit-log sum (method step 2) and replace the last point; add nanobodies to the note if the team publishes a new total.
3. Data Universe: check datauniverse.macrocosmos.ai and @Data_SN13 for a dated total; if one appears, set `verified: true` and remove `floor`.
4. Beam: check @b1m_ai for a transfer larger than 1.12 TB or a first customer; if a customer is named, change the measure line and the caveat.
5. `npm run slides`, `npm run build`, open slide 3 at 1600×900 and 390×844, `npm run deploy`.

---

## Slide 3. Rent, revenue, burn (compute: Lium SN51)

Built 24 September 2026 by the compute-and-inference worker. Live as slide 3 at https://bittensor-talk.pages.dev/#/3. PNG: `media/growth-lium.png` (3200×1800). Data: `content/data.json → growthLium`. After the merge the order is compute (3), inference (4), data (5); every slide from 6 on is three higher than on 23 September, and every cross-reference in `content/slides.md`, the LAYOUTS table and the README was renumbered. The Data section above still says "slide 3" and "from 4 on"; it was written before the merge.

Terms used in this section. **Trailing 30 days**: the sum of the last thirty days, read once a week, so the line is a month's billing that slides forward. **Credit spend**: what renters' prepaid credits are charged per hour while a GPU is rented; Lium's Grafana shows it per GPU type in dollars an hour. **Integrated**: added up over time (dollars an hour × hours = dollars). **Buyback and burn**: the team buys its own subnet token (alpha) on the open market with revenue and destroys it. **Extrinsic**: one transaction on the Bittensor chain, named by block number and index. **Grafana**: an open-source dashboard tool; Lium runs one in public at grafana.lium.io with anonymous read access.

### Headline (the one sentence under the figure)

"Rent → revenue → burn → price → miners add GPUs."

Small grey title above: "Lium · SN51 · September 2025 to September 2026". The values sit on the lines: $423K → $1.20M billed in a trailing 30 days; 1,061 → 1,458 GPUs online; six burn bars labelled $50K, $665K, $150K, $571K, $400K, $966K.

### Chart form, and why

Two lines and six bars on one time axis, no gridlines, no value axis. Each line is on its own linear scale from zero, with its value written at both ends, so the shapes are true ratios (a line that ends seven times higher than it started is seven times higher on the screen). The black line stops at the top; the grey one stops at 76 percent so two rising lines do not end on the same pixel. The bars hang below the baseline: what came in above, what was burned below. Solid bar = amount read from the chain; hollow bar = the team's word only. A shared log axis (as on slide 1) was tried and rejected: five decades of axis compress a 7× rise into a fifth of the height, and the slide is about the rise.

The shape is a U, not a ramp. Both lines fall from October 2025 to a trough in March 2026 and then rise 6–7× to today. The chart shows the whole year; the notes say the dip first. "Up and to the right" is true for the last six months, not the last twelve.

### The three series

| Series | First point | Trough | Last point | Multiple from trough | Points | Source | Status |
|---|---|---|---|---|---|---|---|
| Billed, trailing 30 days (USD) | $422,781 · 15 Oct 2025 | $158,728 · 1 Apr 2026 | $1,202,479 · 23 Sep 2026 | 7.6× in 25 weeks | 50 weekly | grafana.lium.io, table `prod_revenue_history`, `AVG(revenue)` per GPU type per hour, summed and integrated | Verified against Lium's own figure (below) |
| GPUs online (count) | 1,061 · 29 Sep 2025 | 218 · 2 Apr 2026 | 1,458 · 24 Sep 2026 | 6.7× in 25 weeks | 53 weekly | grafana.lium.io: `prod_executors` (validator-scored GPUs, noon sample) to 8 Dec 2025, then `gpu_inventory` (weekly mean of hourly snapshots of `total_gpus`); last point lium.io/api/executors/stats | Verified; two tables, see method |
| Bought back and burned from revenue (USD per event) | $50K · 13 Nov 2025 | — | $966K · 9 Sep 2026 | — | 6 events | Chain extrinsics for two; @lium_io and @fish_datura posts for four | 2 verified on-chain, 4 team-stated (hollow) |

#### Billed, trailing 30 days (weekly, Wednesdays)

15 Oct 2025 $423K · 22 Oct $415K · 29 Oct $415K · 5 Nov $422K · 12 Nov $419K · 19 Nov $426K · 26 Nov $420K · 3 Dec $390K · 10 Dec $375K · 17 Dec $323K · 24 Dec $278K · 31 Dec $255K · 7 Jan 2026 $226K · 14 Jan $221K · 21 Jan $224K · 28 Jan $217K · 4 Feb $213K · 11 Feb $206K · 18 Feb $193K · 25 Feb $182K · 4 Mar $178K · 11 Mar $169K · 18 Mar $167K · 25 Mar $165K · 1 Apr $159K · 8 Apr $160K · 15 Apr $161K · 22 Apr $174K · 29 Apr $191K · 6 May $221K · 13 May $283K · 20 May $340K · 27 May $352K · 3 Jun $353K · 10 Jun $309K · 17 Jun $253K · 24 Jun $246K · 1 Jul $291K · 8 Jul $371K · 15 Jul $447K · 22 Jul $521K · 29 Jul $564K · 5 Aug $642K · 12 Aug $731K · 19 Aug $793K · 26 Aug $856K · 2 Sep $929K · 9 Sep $971K · 16 Sep $1,069K · 23 Sep $1,202K.

Calendar months, same source: Sep 2025 (from the 11th) $284K · Oct $430K · Nov $402K · Dec $267K · Jan 2026 $222K · Feb $169K · Mar $165K · Apr $193K · May $367K · Jun $284K · Jul $604K · Aug $918K · Sep (to the 24th, 564 hours) $1,004K.

**Check against Lium's own numbers.** The 30 days to 9 Sep 2026 integrate to $970,790; Lium posted "$964k billed to 1187 renters (36% MoM growth)" for its month to 9 Sep (x.com/lium_io/status/2097824624117473549). The 30 days to 9 Aug integrate to $697,878, so the dashboard's month-over-month is +39% against the team's +36%. May 2026 integrates to $367K against the single-sourced "$432K" in `top-subnets-2026.md`. The Grafana line is therefore Lium's billed revenue to within about five percent, read without asking anyone.

#### GPUs online (weekly)

29 Sep 2025 1,061 · 6 Oct 1,151 · 13 Oct 1,000 · 20 Oct 824 · 27 Oct 953 · 3 Nov 1,159 · 10 Nov 898 · 17 Nov 765 · 24 Nov 797 · 1 Dec 631 · 8 Dec 534 (all `prod_executors`, noon) · 11 Dec 579 · 18 Dec 422 · 25 Dec 378 · 1 Jan 2026 361 · 8 Jan 464 · 15 Jan 416 · 22 Jan 366 · 29 Jan 358 · 5 Feb 405 · 12 Feb 422 · 19 Feb 321 · 26 Feb 312 · 5 Mar 292 · 12 Mar 275 · 19 Mar 251 · 26 Mar 227 · 2 Apr 218 · 9 Apr 261 · 16 Apr 341 · 23 Apr 336 · 30 Apr 334 · 7 May 416 · 14 May 472 · 21 May 423 · 28 May 371 · 4 Jun 414 · 11 Jun 480 · 18 Jun 508 · 25 Jun 636 · 2 Jul 650 · 9 Jul 842 · 16 Jul 934 · 23 Jul 1,034 · 30 Jul 1,158 · 6 Aug 1,265 · 13 Aug 1,178 · 20 Aug 1,117 · 27 Aug 1,191 · 3 Sep 1,300 · 10 Sep 1,366 · 17 Sep 1,425 (all `gpu_inventory`, weekly mean) · 24 Sep 1,458 listed, 1,268 rented (lium.io/api/executors/stats, 11:48 UTC). Rented GPUs from the same inventory table, weekly mean: 315 (11 Dec 2025) → 160 (2 Apr 2026) → 1,213 (24 Sep 2026); utilisation 54% → 71% → 83%.

Where the two tables overlap (8 Dec 2025 to 8 Jan 2026) they agree to within about ten percent (534 vs 579; 366 vs 422; 394 vs 378; 352 vs 361; 394 vs 464), so the join is honest but not seamless; the notes round to "a thousand last October, two hundred and twenty in March".

#### Burns (revenue-funded buyback and burn of SN51 alpha)

| Date | Amount | TAO | Alpha burned | Evidence | Status |
|---|---|---|---|---|---|
| 7–13 Nov 2025 | >$50,000 | ≈143 TAO at $348 | — | x.com/lium_io/status/1986626278448112056 (7 Nov: "buying over $10,000 per day into SN51 token... Already over 30 tao"); status/1988776686087991368 (13 Nov: "Over $50,000 so far") | Team-stated, hollow |
| 5 Apr 2026 | ≈$665,000 | 2,207 TAO | 50,409.49 | taomarketcap.com/extrinsics/7898274-0011 (`add_stake_limit`, 2,207 TAO, netuid 51, block 7,898,274) and 7898314-0023 (`burn_alpha`, 50,409.49 alpha, block 7,898,314); TAO $301.34 that day (CoinGecko). Funding source is the team's word: x.com/lium_io/status/2049346952614007286 (29 Apr: "revenue bought back + burned") | On-chain amount, solid |
| 4 May 2026 | >$150,000 | ≈518 TAO at $289 | — | x.com/lium_io/status/2051112124328374707: "This money comes from people buying credits on lium" | Team-stated, hollow |
| 22 Jun 2026 | ≈$571,000 | 2,500 TAO (team) | — | x.com/lium_io/status/2069199663530471602: "buyback and burn of 2500 $TAO... All funds came from revenue from credit purchases"; x.com/fish_datura/status/2069120925287518597: "over $500k in revenue just from tao deposits" since the last buyback; TAO $228.53 | Team-stated TAO, hollow (extrinsic not located) |
| 5 Aug 2026 | ≈$400,000 | ≈2,046 TAO at $196 | — | x.com/lium_io/status/2084900760010043595: "~$400,000 in revenue from just Tao deposits. All of it was used to buyback and burn" | Team-stated, hollow |
| 9 Sep 2026 | ≈$966,000 | 3,737.69 TAO | 40,737.93 | taomarketcap.com/extrinsics/9032919-0005 (`add_stake_burn`, block 9,032,919); TAO $258.45; x.com/fish_datura/status/2097814414447481250 ("$1,000,000.00") | On-chain, solid |

Sum of the six: about $2.8M since November 2025, against about $5.9M billed in the same period. Not on the chart, because they were not revenue-funded: 174,210 Celium tokens ≈ 14,109 TAO burned 3 Jul 2025 and "over $3,000,000 worth" of owner tokens burned 18 Sep 2025 (@fish_datura), and 1,412 owner alpha a day auto-burned from 5 Apr 2026 (x.com/fish_datura/status/2040623836052300008). Between 13 Nov 2025 and 5 Apr 2026 the team says buying was continuous and automatic; no amounts were posted, so the chart has a gap there rather than a guess.

#### SN51 alpha price and pool, read from the chain (not drawn; the "price" step of the flywheel)

Archive node https://archive.chain.opentensor.ai, `state_getStorage` of `SubtensorModule.SubnetTAO`, `SubnetAlphaIn`, `SubnetAlphaOut` for netuid 51; price = SubnetTAO ÷ SubnetAlphaIn, in TAO; market value = price × (alpha in + alpha out), the taostats definition. 1 Sep 2025 (block 6,346,976) 0.0572 τ, 139,822 τ · 1 Oct 0.0608, 174,318 · 1 Nov 0.0552, 182,406 · 1 Dec 0.0586, 217,341 · 1 Jan 2026 0.0551, 223,060 · 1 Feb 0.0424, 181,842 · 1 Mar 0.0415, 186,344 · 1 Apr 0.0408, 192,618 · 1 May 0.0614, 300,289 · 1 Jun 0.0508, 257,356 · 1 Jul 0.0537, 282,518 · 1 Aug 0.0662, 365,100 · 1 Sep 0.0827, 478,417 · 23 Sep 2026 (block 9,133,376) 0.0892 τ, 531,706 τ. So the token's price in TAO bottomed in the same fortnight as billing and GPUs (late March to early April) and has since risen 2.2×; TAO in the pool 95,476 → 159,128. Emission share is not drawn: trackers disagree (13.7% vs 24.1% on 12 Sep, `top-subnets-2026.md`).

### Method

1. grafana.lium.io (Grafana 12.4.2) allows anonymous Viewer access and exposes one datasource, a read-only TimescaleDB/Postgres (`grafana_public` user, uid `eflbe16w0qigwd`). `GET /api/search?type=dash-db` lists ten dashboards; `GET /api/dashboards/uid/<uid>` returns every panel's SQL. `POST /api/ds/query` with `rawSql` runs read-only SQL as the anonymous viewer. Tables read: `prod_revenue_history` (11 Sep 2025 → now, 3.70M rows; `time`, `gpu_type`, `revenue` = $/hr), `gpu_inventory` (10 Dec 2025 → now; `snapshot_time`, `gpu_type`, `total_gpus`, `rented_gpus`, `available_gpus`), `prod_executors` (25 Sep 2025 → now; `time`, `validator_hotkey`, `executor_id`, `gpu_count`, `score`), `rental_history` (4 Sep 2025 → now, 42,910 rentals; `customer_id`, `price_per_gpu`, start and end).
2. Billed: the dashboard's own "Credit spend ($/hr)" panel is `SUM(AVG(revenue) per gpu_type per hour)`; that was integrated hour by hour into dollars a day, then summed over trailing 30-day windows ending each Wednesday. Two hours on 15–16 Nov 2025 report the RTX A4000 type at $166,667/hr (a data error, $610K in two days against $14K on the days around); any gpu_type-hour above $10,000 is excluded.
3. GPUs: `gpu_inventory` summed over GPU types per snapshot, then averaged per week (168 hourly snapshots). Before 10 Dec 2025 the `Total GPUs` stat panel's method was used: per timestamp and validator, the sum of each executor's max `gpu_count` where `score <> 0`, the max over validators, sampled between 12:00 and 12:20 UTC each Monday (a full-table scan times out at the gateway).
4. Renters and rentals, for the notes: `rental_history` gives 7,549 rentals and 900 distinct `customer_id`s in August 2026 and 11,465 / 630 in September to the 24th; the team's month to 9 Sep said 7,697 rentals and 1,187 renters, so the table is complete for rentals and undercounts renters (the ids may be per key, not per person). Median rental 0.9–1.5 hours since December 2025, matching the team's 1.15.
5. Burns: extrinsic pages on taomarketcap.com carry the decoded call in their HTML; TAO prices from api.coingecko.com `/coins/bittensor/history`.
6. X reads: 7 (two timelines of @lium_io and @fish_datura since June 2025, 306 posts). Bearer token from 1Password vault Arbos, never printed. Everything else was public HTTP.
7. Every point in `content/data.json → growthLium` carries `source`; the file's `source` string records the check against Lium's own figure.

### Caveats, to say before a critic does

1. **It is a U.** Billing fell 62% and GPUs fell 79% from October 2025 to the end of March 2026, then rose 7.6× and 6.7×. The team has not published a reason for the winter; on the record is that rental failures ran at 30% in April and 3% by July (x.com/lium_io/status/2077877053496467804). Say the dip first.
2. **Billed is gross.** Providers take most of it; nobody publishes Lium's take. Do not say "revenue" for the black line without "billed to renters".
3. **Some demand is circular.** Affine and Teutonic miners rent Lium GPUs with emission-funded alpha ("subnet 3 pretraining relaunches -> lium revenue up 20%", x.com/lium_io/status/2045540119952695455, 18 Apr 2026). The split is not disclosed.
4. **Emissions still exceed revenue.** At 13.7–24.1% of about 3,600 TAO a day and $287, SN51 receives $140K–250K a day in emissions against about $40K a day billed. The flywheel is real; it is not yet self-funding.
5. **Four of the six burns rest on the team's word.** The TAO amounts for 22 Jun and 5 Aug were not located on-chain in the time available; the April and September burns were.
6. **The GPU line joins two tables** with different definitions (validator-scored GPUs vs marketplace inventory); the overlap agrees to about ten percent.
7. **The Grafana is Lium's own.** It is public and read-only, and it reproduces the team's posted billing to five percent, but it is not an independent audit.

### Considered and not drawn

- **Renters a month** (`rental_history` distinct customers: 102 in Dec 2025 → 900 in Aug 2026): a clean 9× line, but the ids undercount the team's 1,187 and the table is empty before December; kept to the notes.
- **Rentals a month** (972 → 11,465): the same shape as billing; redundant on screen.
- **Alpha price in TAO** (0.041 → 0.089): the "price" step, but a token price on the third slide of the deck breaks the opening's rule (nothing on screen is a price); said aloud instead.
- **TAO burned instead of dollars**: 2,207 and 3,738 TAO are exact, the other four would be conversions of dollar posts; dollars are the unit the team posted in five of six cases.
- **A shared log axis**: tried; flattens the rise.

### On screen (the line in `content/slides.md`)

White. Small grey title: "Lium · SN51 · September 2025 to September 2026". One chart, no gridlines, no value axis, months along the bottom (Oct 2025, Jan 2026, Apr, Jul, Today). A black line, "Billed, trailing 30 days": $423K at the left, a trough in March, $1.20M at the right. A grey line beneath it, "GPUs online": 1,061 at the left, a trough in March, 1,458 at the right; each line on its own scale from zero, its value written at both ends. Along the baseline, six thin black bars, "Bought back and burned from revenue", each with its dollar amount: $50K (Nov), $665K (Apr), $150K (May), $571K (Jun), $400K (Aug), $966K (9 Sep); the two read from the chain are solid, the four that rest on the team's word are hollow. One sentence beneath: "Rent → revenue → burn → price → miners add GPUs."

### Speaker beats (Jacob's voice; also in the deck notes)

- 2:30. Take the first card and put a year under it. Everything on this screen comes from Lium's own public dashboard and from the chain; nobody at Lium sent me a number.
- The black line is dollars renters paid for GPU time in the trailing thirty days. Four hundred and twenty thousand a year ago. A hundred and sixty thousand at the bottom, in March. One point two million today. Seven times in six months. Their own month-in-review said nine hundred and sixty-four thousand to the ninth of September; the dashboard says nine hundred and seventy-one thousand for the same thirty days. Same number, two doors.
- Say the dip before anyone else does. The grey line is GPUs online. A thousand last October. Two hundred and twenty in March. I do not have Lium's reason for the winter on the record. What is on the record: in April three rentals in ten failed; by July it was three in a hundred, and the team says renters came back when the nodes stopped failing. Supply followed: fourteen hundred and fifty-eight this morning, twelve hundred and sixty-eight of them rented. Nobody told a provider to plug in a card. The price told them.
- The bars are what the team bought back and burned with that revenue. Ten thousand a day when the program started in November. Six hundred and sixty-five thousand in April. Nine hundred and sixty-six thousand on the ninth of September: three thousand seven hundred and thirty-eight TAO in one transaction, block 9,032,919. Check it. The hollow bars are the team's word; the solid ones are on-chain.
- Read the sentence under the chart. Rent, revenue, burn, price, miners add GPUs. Lium's token, priced in TAO, went from 0.041 in April to 0.089 this week, and the GPUs came with it. "Let the flywheel commence. This is how bittensor succeeds." [@fish_datura, 7 Nov 2025] That is the protocol doing what it was designed to do. I did not do it. Fish did not do it. The mechanism did.
- Do not say "up only". It fell by half over the winter and came back seven times. Say the three honest things: billing is gross and providers take most of it; part of the demand is other subnets' miners paying with emissions, Fish says so himself; and the protocol still pays SN51 more in emissions than renters pay Lium. The flywheel is real. It is not yet self-funding.
- Advance on "tokens".

### On 27 September

1. Re-run the three Grafana queries (method steps 2–3) with `POST https://grafana.lium.io/api/ds/query` (body: `{"queries":[{"refId":"A","datasource":{"type":"grafana-postgresql-datasource","uid":"eflbe16w0qigwd"},"rawSql":"<sql>","format":"table","rawQuery":true}],"from":"now-5y","to":"now"}`); append the new Wednesday to `growthLium.series[0].points` (billed) and `series[1].points` (GPUs); replace the last GPU point with `lium.io/api/executors/stats` summed.
2. Check @lium_io and @fish_datura for a new burn; if one has an extrinsic, add it solid with the TAO amount, else hollow with the dollar figure.
3. If Lium posts a "month in review" for September, put its number in the notes beside the dashboard's trailing-30 for the same end date, and keep both.
4. `npm run slides`, `npm run build`, open slide 3 at 1600×900 and 390×844, `npm run deploy`.

---

## Slide 4. Tokens a day (inference: Chutes SN64, Engy SN53, SayGM SN28)

Built 24 September 2026 by the compute-and-inference worker. Live as slide 4 at https://bittensor-talk.pages.dev/#/4. PNG: `media/growth-inference.png` (3200×1800). Data: `content/data.json → growthInference`. Jacob named chutes.ai/app/research as the source for Chutes; it is the primary source here and is cited as "chutes.ai/app/research, 24 Sep 2026".

Terms used in this section. **Token**: the unit a language model reads and writes, about three quarters of a word. **Tokens a day**: input plus output tokens, summed over a day. **7-day mean**: the average of seven daily values, ending on the date given; it smooths weekends. **Epoch (Engy)**: one settlement week, Wednesday 08:00 UTC to Wednesday 08:00 UTC, published as a signed ledger. **Epoch (SayGM)**: one 72-minute settlement window on saygm.com/miners. **Settled**: what a ledger says miners earned for the traffic they served, at the team's rate card; not audited revenue. **Free tier**: traffic served for no money.

### Headline (the number, the date and the method)

**About 41 billion tokens a day across three inference subnets, in the seven days to 23 September 2026.** Chutes 26.4B (measured, chutes.ai/app/research, 7-day mean 17–23 Sep) + Engy 12.1B (measured, signed ledger epoch 14, 16–23 Sep, 84.67B tokens ÷ 7) + SayGM about 2.5B (approximated, range 2–3B, see method) = 41.0B. Requests: Chutes 1.95M a day + Engy 1.00M a day + SayGM about 0.29M a day = about 3.25M a day. Confidence: the first two are read from the teams' own published counters and are exact for what they count (Chutes counts public LLM endpoints only; Engy counts miner-served traffic only); the third is an estimate and moves the total by at most ±1B. Say "about forty billion" if "forty-one" feels too exact.

The one sentence under the figure: "Two of the three did not exist in April; the third cut its free tier in March."

### The honest shape

This is **not** an up-and-to-the-right series. The total peaked at 151B tokens a day in the week to 25 June 2025, when Chutes alone served that much, nearly all free; the daily peak was 159.5B on 20 June 2025. The total is 41B today, 27% of the peak. Chutes fell to 26B (17% of its peak) after cutting free-tier traffic in March 2026 (tokens −45%, revenue −25%, by the team's account; revenue per trillion tokens at a record $340K in July, `top-subnets-2026.md`). Engy (from 21 Jul 2026) and SayGM (from 28 May 2026) have grown from nothing to about 14.6B a day together, and the total has been roughly flat at 41–58B since August 2026. The chart draws the whole history from February 2025 with the peak ringed and labelled, so the room sees the fall before a critic names it; the on-screen sentence and the beats explain it. If Jacob wants a rising line on this slide, the honest one is the two entrants alone (0.9B → 14.6B a day in nine weeks); the total cannot be drawn rising without hiding 2025.

### The four series

| Series | First point | Peak | Last point (7 days to 23 Sep 2026) | Points | Source | Status |
|---|---|---|---|---|---|---|
| All three (total) | 1.07B · 12 Feb 2025 (Chutes only) | 151.4B · 25 Jun 2025 | 41.0B | 85 weekly | sum of the three below | Solid to May 2026; hollow and dashed from June 2026 because it includes the SayGM approximation |
| Chutes SN64 | 1.07B · 12 Feb 2025 | 151.4B · 25 Jun 2025 | 26.4B; 1.95M requests | 85 weekly (7-day means ending Wednesdays) | chutes.ai/app/research, 24 Sep 2026: its JSON `/api/research/stats?type=llm&period=1y` (daily `tokens`, `count`, `input_tokens`, `output_tokens` since 30 Jan 2025). Identical, day for day, to api.chutes.ai/invocations/stats/llm | Verified |
| Engy SN53 | 0.87B · 22–29 Jul 2026 | 12.1B · 23 Sep 2026 | 12.1B; 1.00M requests; $34,870 settled; 82 miners | 9 weekly | provider.engy.ai/api/subnet/v1/epoch/0 … /14, `tokens_total` summed over miners ÷ days | Verified (signed ledger) |
| SayGM SN28 | ≈0.1B · Jun 2026 | — | ≈2.5B (range 2–3B); ≈0.29M requests | 4 monthly, all hollow | @say_gm_ posts of 7 Sep 2026; saygm.com/miners ledger; saygm.com/models pages | Approximation |

#### Chutes SN64 (chutes.ai/app/research, 24 Sep 2026), monthly means of the weekly points, tokens a day

Feb 2025 3.2B · Mar 11.9B · Apr 34.9B · May 73.5B · Jun 134.0B · Jul 111.0B · Aug 137.4B · Sep 119.1B · Oct 104.5B · Nov 102.9B · Dec 103.7B · Jan 2026 107.1B · Feb 104.4B · Mar 63.1B · Apr 76.2B · May 53.0B · Jun 47.2B · Jul 42.7B · Aug 44.0B · Sep (to the 23rd) 34.8B. Last eight weeks (7-day mean, tokens a day / requests a day): 5 Aug 43.3B / 6.68M · 12 Aug 42.5B / 6.99M · 19 Aug 42.8B / 6.09M · 26 Aug 47.3B / 6.64M · 2 Sep 48.4B / 6.53M · 9 Sep 35.0B / 3.61M · 16 Sep 29.3B / 2.37M · 23 Sep 26.4B / 1.95M. Requests fell by two thirds and tokens by 45% in the three weeks after 2 September; tokens per request rose from 7,400 to 13,600. Lifetime on the page: 45.54T tokens (42.67T input, 2.87T output), 6.66B requests, 30 Jan 2025 to 24 Sep 2026; 801,938 users (`/api/research/stats?type=users`, cumulative to 24 Sep 2026); overview tile: 4,312 GPUs online, 491 public chutes, 80,427 requests in the last hour (`/api/research/overview`, 11:50 UTC). The research page says its token chart is "smoothed as a 7-day average", which is why the deck uses 7-day means too. The arXiv paper (2608.13573, trace Apr 2025 – Apr 2026): 6.12B requests, 35.8T input tokens, 314,970 users; consistent with the page's cumulative numbers for that window.

#### Engy SN53 (provider.engy.ai signed ledger), per epoch

Epochs 0–6 were daily (22–29 Jul 2026): 0.22, 0.54, 0.47, 0.94, 1.14, 1.56, 1.24B tokens; 6.1B in the week, 0.87B a day; 1.87M requests; $682 settled; drawn as one point at 29 Jul. Weekly epochs, tokens a day / requests a day / settled / miners: 7 (29 Jul–5 Aug) 1.46B / 273K / $945 / 105 · 8 (5–12 Aug) 0.85B / 64K / $350 / 154 · 9 (12–19 Aug) 2.26B / 241K / $3,039 / 171 · 10 (19–26 Aug) 5.33B / 271K / $4,467 / 199 · 11 (26 Aug–2 Sep) 7.68B / 609K / $15,712 / 163 · 12 (2–9 Sep) 8.00B / 509K / $14,318 / 77 · 13 (9–16 Sep) 9.50B / 544K / $12,726 / 75 · 14 (16–23 Sep) 12.10B / 1,005K / $34,870 / 82. Epoch 14 model mix: 64 miners on Qwen3.6-35B-A3B; the Kimi K3 miners carry most of the tokens (one Kimi miner alone: 316,692 requests, $6,448 settled). Each epoch JSON is signed by the owner hotkey (`signed_hotkey`, `signature`, `digest`), so any validator can recompute it. "Settled" excludes first-party-served models and is a rate-card value, not audited revenue (`top-subnets-2026.md`).

#### SayGM SN28 (approximation; every point hollow)

What SayGM publishes: per 72-minute epoch, requests, dollar traffic value and miners (saygm.com/miners: epoch 25,310 finalised 24 Sep 2026 11:47 UTC, 16,372 requests, $639.67, 70 miners; the twenty epochs from 23 Sep 12:56 to 24 Sep 11:47 UTC total 292,266 requests and $11,523); per model, "Measured over N windows: R requests and T input-side tokens" (saygm.com/models/<slug>, 47 models with data, 24 Sep 2026). What it posted: "GM is approaching 100 billion recorded network tokens processed. We're at 96.48B. Including more than 30B output tokens generated" (x.com/say_gm_/status/2096886191081615619, 7 Sep 2026); "August's finalized network token volume grew 545% over July. Output tokens alone grew 412%, reaching 20.44 billion in August" (status/2096886191794668016). It does not publish tokens per epoch or per day.

Method, monthly points: let A be August's tokens and J July's; A = 6.45 J. Cumulative to 7 Sep = 96.48B = June + J + A + (1–7 Sep). Taking June as about a third of July (beta went live 28 May) and 1–7 Sep at August's daily rate gives J ≈ 10.5B, A ≈ 67.6B (2.2B a day; output 20.44B is then 30% of the total, and the cumulative output share of 31% agrees). If June were zero, A ≈ 69.9B; the August point moves by ±0.1B a day either way. Points drawn: Jun 2026 ≈ 0.1B a day; Jul ≈ 0.34B; Aug ≈ 2.2B; 23 Sep ≈ 2.5B.

Method, the last point: the latest epoch's requests by model (from the miners page) multiplied by each model's input tokens per request (from its model page; 6,430 weighted over the epoch's mix, driven by GLM-5.3-Flash at 47,600, DeepSeek V4.1 Flash at 50,900 and GPT-5.4-mini at 25,400 tokens a request) gives 96.4M input tokens in the epoch, 1.93B a day over twenty epochs; adding output at the August share (30%) gives about 2.75B a day. The simpler fleet-wide ratio (4,147 input tokens a request over 41,588 sampled requests) gives 1.2B input, about 1.7B total. The chart uses 2.5B and the doc says 2–3B. Cross-check: 2.5B a day for the 17 days after 7 Sep would put SayGM past 100B on about 9 Sep; the team has not posted the crossing.

#### The total

Weekly, Chutes + Engy (where its epoch ends on the same Wednesday) + SayGM (linear between the monthly points; zero before 28 May 2026). Last eight weeks: 5 Aug 46.4B · 12 Aug 45.4B · 19 Aug 47.3B · 26 Aug 54.9B · 2 Sep 58.5B · 9 Sep 45.4B · 16 Sep 41.3B · 23 Sep 41.0B. Hollow and dashed from June 2026 because about 6% of it is the SayGM approximation.

### The comparison anchor

OpenRouter publishes tokens per model, not a platform total, without an API key. Its rankings page (openrouter.ai/rankings, "Top models today, as text", usage through 23 Sep 2026) lists the ten busiest models on the most recent complete day: DeepSeek V4.1 Flash 2.81T, GLM 5.3 Flash 2.41T, Hy4 preview 1.63T, GPT-5.6 Luna 1.24T, DeepSeek V4 Flash 0731 1.2T, MiMo-V2.6-Flash 567B, DeepSeek V4 Flash 0423 515B, Nemotron 3 Ultra (free) 483B, Hy3 456B, GLM 5.3 451B: **11.76 trillion tokens a day for the top ten alone**, about 290 times the three subnets together. The tenth-busiest model on OpenRouter (451B a day) does eleven times the three subnets. OpenRouter's own blog (Aug 2026): one model, GLM-5.3-Flash as "Ox Alpha", processed over 20 trillion tokens in six days on OpenRouter. The anchor Jacob can say: "OpenRouter's ten busiest models alone did eleven point eight trillion tokens on the twenty-third of September. We are a rounding error to OpenRouter." Alternative anchors, with sources: Chutes' lifetime 45.5T tokens (chutes.ai/app/research) against the 100T-token dataset OpenRouter and a16z analysed in "State of AI 2025" (openrouter.ai/data).

### Method

1. Chutes: the research page is a Next.js client app; its bundle calls `/api/research/overview`, `/api/research/stats?type=llm&period=<period>` and `/api/research/stats?type=users` on chutes.ai. `period=1y` returned the full daily series from 30 Jan 2025 (603 days). Every daily value was compared with the per-chute rows of api.chutes.ai/invocations/stats/llm summed by day (92,370 rows): identical. 7-day means end on Wednesdays so they line up with Engy's epochs.
2. Engy: `GET https://provider.engy.ai/api/subnet/v1/epoch/<n>` for n = 0…14 (`/epoch/latest` wraps the same object in `result_json`); `tokens_total`, `requests`, `settled_micro` summed over `miners`; divided by the epoch length in days.
3. SayGM: saygm.com/miners and saygm.com/miners/<epoch> (server-rendered; the twenty-epoch table is in the HTML); saygm.com/models/<slug> pages need a browser user agent (plain clients get 403); the input-tokens-per-request line is "Measured over N windows: R requests and T input-side tokens"; @say_gm_ timeline (196 posts since 28 May 2026) via the X API.
4. OpenRouter: openrouter.ai/rankings HTML, the "as text" tables.
5. X reads in total for this section: 4 (one timeline). Bearer token from 1Password vault Arbos, never printed. Total X reads by this worker: 11.
6. Labels are computed from the points by `formatGrowth` in `src/figures.ts`; the headline object in `content/data.json → growthInference.headline` carries the three components and the window.

### Caveats, to say before a critic does

1. **The total is below the 2025 peak** (41B vs 151B). Chutes cut free-tier traffic in March 2026; tokens fell 45% and revenue 25% by the team's account. Never say "we serve a hundred and twenty billion tokens a day"; that was mid-2025 (`factcheck-slides-v2.md`, "Do not say" 24).
2. **Chutes fell again in September**: 48.4B (week to 2 Sep) → 26.4B (week to 23 Sep), requests 6.5M → 1.95M. No team statement found for it. The new CEO was named 7 Sep. Re-pull on the 27th; if it has fallen further, keep it.
3. **Chutes counts public LLM endpoints only**; image, speech and private chutes are excluded, so 26B is a floor for Chutes and the arXiv paper's 35.8T input tokens over a year is a different window.
4. **Engy's "tokens" are miner-served tokens** at the team's rate card; first-party-served models are excluded; one validator does the verification (`top-subnets-2026.md`). Nine weeks of history, so the multiple (14×) is from a small base.
5. **SayGM is an estimate**, ±1B a day, and SayGM resells the labs' own APIs, so its tokens are Claude, GPT and Gemini tokens routed through a TEE, not open-model inference on Bittensor hardware. Say "routed", not "served", for SayGM.
6. **Requests are not comparable across the three**: Chutes' 13,600 tokens a request, Engy's 12,000, SayGM's 6,400 input.
7. **Nothing here is revenue.** Engy's $34,870 settled in a week is on the ledger; Chutes' $1.13M per 90 days is the team's dashboard; SayGM's $11.5K a day is gross traffic value at list minus discount. Revenue is on slide 24, not here.

### Considered and not drawn

- **Tokens on a log axis**: makes the two entrants look like the story and hides the fall; rejected for this slide, though slide 1 uses one.
- **Stacked areas**: the deck draws lines and bands, not fills; and a stack would need three tones.
- **Paid tokens only**: Chutes does not publish a paid/free split by day; the team's statements are ratios, not series.
- **Requests a day**: down for Chutes (25M a day in Sep 2025 → 1.95M), and it is what `opening-slides.md` already rejected for slide 1.
- **OpenRouter's Chutes-as-provider volume**: OpenRouter's provider pages need a key for history; the public rankings are by model, not provider.
- **Targon SN4 and KubeTEE SN90**: compute, not inference; no public token counters.

### On screen (the line in `content/slides.md`)

White. Small grey title: "Tokens a day · Chutes SN64, Engy SN53, SayGM SN28 · Feb 2025 to Sep 2026". One chart on one linear scale from zero, months along the bottom (Apr 2025, Jul, Oct, Jan 2026, Apr, Jul, Today). A black line, "All three · 41B": Chutes alone until 2026, a peak ringed and labelled "151B at the peak" in June 2025, a long slope down, then a floor that turns up in August 2026; the line is dashed from June 2026 because it includes an approximation. Three grey lines beneath: "Chutes · 26B" (the same shape), "Engy · 12B" (from nothing on 29 July 2026, straight up), "SayGM (approx.) · 2.5B" (four hollow dots from June). One sentence beneath: "Two of the three did not exist in April; the third cut its free tier in March."

### Speaker beats (Jacob's voice; also in the deck notes)

- 3:30. Now inference. The question I get asked is: how many tokens are you actually serving? Here is the honest answer, with the method on the slide. About forty-one billion tokens a day across three inference subnets, in the week to the twenty-third of September. Twenty-six billion on Chutes, from its own public research page. Twelve billion on Engy, from a ledger its validators sign every week. About two and a half billion on SayGM, and that one is an estimate; I will tell you how in a second. Three and a quarter million requests a day between them.
- Say the peak first. That ring is June 2025: a hundred and fifty-one billion tokens a day, Chutes alone, almost all of it free. In March the team cut the free tier. Tokens fell by nearly half and revenue fell by a quarter, by their own account, and revenue per trillion tokens hit a record in July. The line went down because they stopped giving it away. I would rather show you that than hide it.
- Now the two lines that did not exist in April. Engy launched on the twenty-first of July. Its first week served nine hundred million tokens a day. Last week, twelve billion a day, seven million requests, eighty-two miners, thirty-five thousand dollars settled to them, mostly on gaming cards. Fourteen times in nine weeks. Every number is in a signed ledger you can pull. SayGM went live in May and reported ninety-six billion tokens on the seventh of September; the dots are hollow because SayGM publishes requests and dollars per epoch, not tokens, so I multiplied requests by the tokens per request its own model pages show.
- The comparison, so nobody has to make it for me. OpenRouter's ten busiest models alone did eleven point eight trillion tokens on the twenty-third of September. We are a rounding error to OpenRouter. Fifteen months ago Chutes alone did a hundred and fifty billion a day, free. It stopped giving it away. What is on this screen now is paid, or on its way to paid.
- One line for the room: the volume that left was free. The volume that came is paid, on hardware nobody licensed, with a receipt for every answer.
- Do not say "we serve a hundred and twenty billion tokens a day". That was mid-2025. Do not add SayGM's number without the word "about".
- Advance on "paid". The data layer is next.

### On 27 September

1. Chutes: `GET https://chutes.ai/api/research/stats?type=llm&period=1y`; take the seven days to Wednesday 30 Sep (or the last full day); append a point to `growthInference.series[1].points` with `value` = mean `tokens`, `requests` = mean `count`.
2. Engy: `GET https://provider.engy.ai/api/subnet/v1/epoch/15` (ends 30 Sep 08:00 UTC); `tokens_total` summed ÷ 7; append to `series[2].points`.
3. SayGM: read the twenty latest epochs on saygm.com/miners for requests a day; if the team posts a new cumulative or monthly token figure, recompute the monthly derivation above; append a hollow point to `series[3].points`.
4. Recompute the total (`series[0].points`) and `headline`; re-read the OpenRouter "Top models today" table for the anchor.
5. `npm run slides`, `npm run build`, open slide 4 at 1600×900 and 390×844, `npm run deploy`.

---

## Slide 6. Twelve markets (the metrics grid)

Built 24 September 2026 by the metrics-grid worker. Live as slide 6 at https://bittensor-talk.pages.dev/#/6. PNG: `media/metrics-grid.png` (3200×1800). Data: `content/data.json → metricsGrid`. Placed immediately after the Data slide (5), so every slide from 7 on is one higher than before (Hello, one year is now 7; Close is 33). Every cross-reference in `content/slides.md`, the LAYOUTS table in `scripts/generate-slides.mjs` and the README was renumbered; the `[VERIFY]` and `[JACOB]` lists in the production notes were rebuilt from the slide headers.

Terms used in this section. **Panel**: one small box on the slide with one subnet, one number and one line. **Sparkline**: a small line chart with no axes; here it starts from a zero hairline and each panel has its own scale. **Hollow point**: a ring instead of a dot, marking a value that came from a team post, not from a ledger, chain, API or dashboard. **Rate**: a value per day, week or month. **Stock**: how much exists at one moment. **Paper positions**: trades that are scored but never placed on an exchange. **Simulation (sim)**: a robot policy scored inside a physics simulator (MuJoCo), not on a physical robot.

### Headline (the one sentence under the grid)

"Twelve markets, twelve teams, and not one of these lines is mine to draw."

Small grey title above: "Twelve markets · to 24 Sep 2026". Small grey footnote beneath the sentence: "Sim and paper: OpenRoboto scores are MuJoCo simulation; Vanta volume is paper positions scored by validators, not exchange trades. Hollow points are team posts, not read from a ledger (IOTA, Feb and Apr 2026)."

### Chart form

The same panel as slide 5, twelve times: grey mono head (name · netuid), one big latest number with its unit, a one-line label, a sparkline from a zero hairline, the start value and month at the left, the end date at the right. Four across and three down at 16:9; two across and six down on a phone (portrait). Row 1 is the four customer-dollar series (Bitcast, Engy, gm, Almanac); row 2 is output and usage (Gradients, Vanta, UR, Hippius egress); row 3 is proof, holders and downloads (OpenRoboto, Conjectures, TAO holders, IOTA). Black, white, one grey; the τ mark and counter as on every other slide.

To fit twelve panels legibly nothing was dropped; the sparklines were made flatter instead (viewBox 320×58 against 320×150 on slide 5, smaller dots) and the type set from the stage height. Measured at 1600×900: head, label and range 15.3 px, footnote 14.4 px, big number 40 px, title 22 px, sentence 34 px; no element leaves the stage. At 390×844: head, label and range 9 px, footnote 8.2 px, big number 19.5 px; no overflow; the grid ends 12 px above the τ mark. A series with more than twenty points (Vanta 26, TAO 43, UR 60) gets a dot at its ends only, so the line stays readable; hollow rings are always drawn.

### The twelve panels

| # | Panel | On screen | First point | Last point | Points | Source (route) | Grade |
|---|---|---|---|---|---|---|---|
| 1 | Bitcast SN93 | $63.1K · brand revenue a month | $2,093 · May 2025 | $63,073 · Aug 2026 | 16 monthly | bitcast.network/revenue (ledger object in the Next.js chunk) | Team ledger, per transaction; customer dollars |
| 2 | Engy SN53 | $34.9K · settled to miners a week | $682 · week to 29 Jul 2026 | $34,870 · week to 23 Sep 2026 | 9 weekly epochs | provider.engy.ai/api/subnet/v1/epoch/6…14, `settled_micro` summed | Signed ledger (owner hotkey) |
| 3 | gm SN28 | $13K · traffic routed a day | $359 · Jun 2026 (26–30 Jun) | $12,976 · Sep 2026 (1–24 Sep) | 4 monthly means | saygm.com/miners/<epoch>, 469 epoch pages sampled (every 4th, 23440–25314) | Team pages, sampled |
| 4 | Almanac SN41 | $40.7K · terminal fees a month | $1,490 · Feb 2026 | $40,689 · Sep 2026 | 8 monthly | almnc.ai/revenue (monthly audit table; two public wallets on taostats) | Team page; wallets public |
| 5 | Gradients SN56 | 565 · models published a month | 4 · Nov 2025 | 565 · Aug 2026 | 10 monthly | huggingface.co/api/models?author=gradients-io-tournaments (+ datasets), counted by `createdAt` | Public API |
| 6 | Vanta SN8 | $1.46B · paper volume, 30 days | $55.9M · 1 Apr 2026 | $1,456M · 23 Sep 2026 | 26 weekly | vantanetwork.io/dashboard (server action, key `volumeData`) | Team dashboard; paper positions |
| 7 | UR SN25 | 206K · devices connected | 120,484 · 21 Jul 2026 | 205,897 · 23 Sep 2026 | 60 daily | grafana.bringyour.com public dashboard ea86caca…, panel 4 | Operator Grafana |
| 8 | Hippius SN75 | 200 TB · S3 egress a month | 15.6 TB · Mar 2026 | 200.3 TB · Sep 2026 (24 days) | 7 monthly | hipstats.com/api/s3-bandwidth-stats (header `X-Requested-With: XMLHttpRequest`) | Team explorer |
| 9 | OpenRoboto SN80 | 90% · LIBERO-Pro champion, sim | 50.3% · 29 Jul 2026 (baseline) | 89.7% · 28 Aug 2026 (gen 11) | 12 | api.openroboto.ai/api/v1/champions/history?competition=1 | Team API; simulation only |
| 10 | Conjectures SN66 | 31 · Lean-verified proofs | 1 · 5 Aug 2026 | 31 · 21 Sep 2026 | 13 | conjectures.io/results | Team results page, every submission dated |
| 11 | TAO · chain | 386K · accounts holding TAO | 14,124 · 31 Mar 2023 | 386,072 · 24 Sep 2026 | 43 monthly | taostats.io/accounts (`statsHistory` payload) | Taostats index of the chain |
| 12 | IOTA SN9 | 45.9K · Train-at-Home downloads | 2,100 · 24 Feb 2026 (post, hollow) | 45,916 · 24 Sep 2026 | 5 (2 hollow) | iota-web.api.macrocosmos.ai/mainnet/v1/stats_overview_metrics; Wayback 14 May and 7 Jun 2026; @IOTA_SN9 posts 24 Feb and 20 Apr | Mixed; two team posts |

Every point in the JSON came from `docs/metrics-gallery.json` (the twelve with `shortlistRank` 1–12) unchanged, with two omissions:

- **Bitcast, September 2026** ($58,341 to 17 Sep) is a partial month and would have drawn as a fall. It is off the line; August is the last complete month and the big number. The note on the panel carries the September figure.
- **UR, 10 August 2026** (21,906) is a telemetry gap on the operator's Grafana, not churn (the gallery's note). It is off the line so the panel does not show a one-day collapse to zero.

Labels are computed by `gridValue` in `scripts/generate-slides.mjs` (three significant figures: $63.1K, $1.46B, 386K, 200 TB, 90%), so a re-pull changes the numbers without editing text. Month-only series show the end as a month ("Aug 2026"); dated series show the day.

### Caveats, to say before a critic does (in this order)

1. **Two panels are not money and not customers.** OpenRoboto's 90% is a robot policy in simulation (MuJoCo, 6 tasks × 6 trials), no physical robot yet; Vanta's $1.46B is paper positions scored by validators, not exchange trades. The real money at Vanta is the payout ledger: $759K to 104 traders since April (USDC on Base). Both carry "sim" and "paper" on the panel label and in the footnote.
2. **Two IOTA points are team posts** (2,100 on 24 Feb; 23,000 on 20 Apr), drawn hollow. Downloads are not active nodes (134 nodes on the v9 run).
3. **Small dollars.** Bitcast $63K and Almanac $41K a month; Engy $35K a week; gm $13K a day of gross traffic at list price, of which gm's take is not published.
4. **Not all twelve are up at the last reading.** gm's weekly line is flat for three weeks; Bitcast's September is partial; UR's 23 Sep (205,897) is below its 15 Sep peak (227,481). The shapes are drawn as they are.
5. **TAO holders is a stock**, which only falls when wallets empty; new holders a month peaked at 20.6K in April 2026 and were 8.8K in August. One exchange wallet is one holder.
6. **Gradients counts repos, not jobs or dollars**; 712 more in September to the 24th are not on the line because the month is not complete.
7. **Hippius egress is free**, so it is usage, not revenue; September's 200 TB is 24 days and already the high; a July counter glitch (petabytes a day) is excluded upstream.
8. **Conjectures is cumulative** (rises by construction); the honest signal is the rate: 19 of 25 bounties in the 13 days to 21 Sep. Bounties are paid in SN66 alpha; the $118K is the page's conversion.
9. **Nothing on the screen is a token price.**

### Speaker beats (Jacob's voice; also in the deck notes)

- 5:30. Last of the opening charts. Do not read it panel by panel; the room can see twelve lines go up. Say what it is: twelve markets, twelve teams, twelve doors, and the same rule as every chart tonight: the number is on a page you can open. I did not draw one of these lines. I do not run one of these teams. Most of them I have never met.
- Caveats first, in one breath (items 1–3 above), then: and every one of them is on a page you can open tonight.
- Say three aloud. **Bitcast** (top left): brands paying for creator reach, two thousand dollars a month in May last year, sixty-three thousand in August, from their per-transaction ledger; customer dollars, not emission. **Conjectures** (bottom row): thirty-one proofs of open conjectures the Lean kernel accepted, the first on 5 August, nineteen of the twenty-five bounties paid in the last two weeks. **Gradients** (second row): four fine-tuned models published in November, five hundred and sixty-five in August, seven hundred and twelve so far this month, every one a repo on Hugging Face.
- If there is time for a fourth: **Engy** (top row), a ledger signed by the owner hotkey every Wednesday; $682 settled to miners in its first week, $34,870 last week, on gaming cards; the same subnet as the twelve billion tokens on slide 4, here in dollars.
- The point: nobody planned this grid; there is no roadmap with these twelve names on it. Each team found a thing people would pay for, or a thing worth measuring, and the market paid them to keep going. That is what "incentive computing" means when it works: I am not the one drawing the lines.
- Do not say "twelve businesses" (four are customer dollars; the rest are usage, output or proof), "growth" for TAO holders, or "up only". Advance on "drawing the lines". Then black (slide 7).

### On 27 September

1. Re-pull the twelve with the routes in `docs/metrics-gallery.md` → "Re-pull on 27 September" (the same routes as the table above). For each, edit `content/data.json → metricsGrid.panels[i].points`: append or replace the last point; keep `verified: true` only for a value read from the route, `false` for a post.
2. Bitcast: if the September ledger total to the 27th is above August's $63,073, add `{ "date": "2026-09-01", "value": <total>, "verified": true }` and change the panel note; if it is still below, leave it off and keep the note.
3. Gradients: count September repos by `createdAt`; if the month is to be shown incomplete, do not add it; mention the count in the notes instead.
4. UR: append the daily readings to 27 Sep; leave out any single-day gap below 50K and note it.
5. gm: recompute the September mean from the latest twenty epochs on saygm.com/miners and replace the last point's value and date.
6. OpenRoboto: if season 2 (`competition=2`) has a higher champion than 0.897, decide whether to switch the panel to season 2 (new base 0.286); if so, replace the points and change the measure line to say "season 2".
7. Check the footnote still names every simulation/paper series and every hollow point.
8. `npm run slides`, `npm run build`, open slide 6 at 1600×900 and 390×844 (nothing may cross the τ mark or the counter), `npm run deploy`; verify https://bittensor-talk.pages.dev/#/6.