# Top Bittensor subnets, September 2026

Research brief for Jacob Steeves' keynote (Exploit Conference, Montreal, 28 September 2026).
Compiled 12 September 2026. Every number has a source and a date. Market figures move daily; re-check anything marked [VERIFY] on the day.

How to read this file:
- **Team-disclosed** = the team said it (tweet, dashboard, docs).
- **On-chain** = you can check it in a block explorer or a public ledger endpoint.
- **Third-party estimate** = an analyst or tracker inferred it.
- **CONFLICT** = two sources disagree. Both are shown.
- Tweets were read directly from the X API or `api.fxtwitter.com`, not from screenshots. Text is verbatim.

Key terms, defined once:
- **Subnet / netuid**: one independent market on Bittensor, identified by a number (SN51 = netuid 51).
- **Alpha**: a subnet's own token. Its price is quoted in TAO (τ). "Market cap" here = taostats definition: alpha price × (alpha in the pool + alpha staked). Other trackers use other definitions, so market caps differ by tracker (see "Tracker disagreement").
- **Emission share**: the slice of newly minted TAO a subnet receives. About 3,600 TAO/day are minted network-wide. Trackers disagree on this number too (see below).
- **External revenue**: money paid by customers outside the network (fiat or TAO deposits for a product). It is not emissions.
- **Buyback and burn**: a team uses revenue to buy its own alpha and destroy it. It is the main way revenue "flows back" into a subnet token today.
- **TEE**: trusted execution environment. A hardware-sealed part of a chip that hides data from the machine's operator.
- **TAO price used**: $234.18 (CoinGecko, 12 Sep 2026 17:35 UTC). Use "about $234".

---

## Executive summary (10 lines)

1. Lium (SN51), Chutes (SN64) and Targon (SN4) are the top three subnets by market cap and by revenue. Together they make roughly half of all external revenue on the network (SubConnect, Aug 2026).
2. Lium is the strongest customer story: $964K billed to 1,187 renters in one month (+36% month over month), and a $1,000,000 revenue-funded buyback and burn on 9 Sep 2026, both checkable on-chain.
3. Chutes is the most externally checkable: ~2.8M requests and ~31B tokens a day on a public API, a one-year production trace published with Harvard and UChicago (6.12B requests, 314,970 users), and ~$4.6M/yr of revenue per its own dashboard.
4. Engy (SN53) is the newest of Jacob's picks and the cleanest "consumer GPUs beat the labs on price" story: Kimi K3 (2.8T parameters) on 80 RTX 5090s, sold at half Moonshot's price, with a public per-epoch ledger showing $14.3K settled to miners in the week of 2–9 Sep 2026.
5. Affine, Teutonic, IOTA, Albedo, Oro and Conjectures have no paying customers. They are emissions-funded research markets. Say so plainly; their story is "open competition produces models and proofs", not revenue.
6. Teutonic-I's "most performant decentralized LLM" claim is contested by an audit from Templar (the previous SN3 team) that found its weights are ~96% correlated with Qwen3.5-9B-Base. Do not repeat the benchmark claim on stage without the caveat.
7. Total external revenue: the only full per-subnet index (SubConnect, 26 Aug 2026) says $28–35M/yr across 24–25 subnets. SubConnect is a Bittensor consultancy, not an independent auditor. Independent measurement (ownyourmind, Aug 2026) gets $1–6M; the team-disclosed, on-chain-checkable run-rate sums to roughly $17–20M.
8. Emissions are worth about $300M/yr at $234/TAO. Even the most generous revenue figure ($35M) is about 11% of that. SubConnect's own line: revenue "offsets 11.7% of annual emissions" (31 Aug 2026).
9. Jacob's 4 Dec 2025 tweet ("no subnet ... generating flow through revenue in any serious way") was close to true then. It is no longer true for Lium, Chutes, Targon and Bitcast. It is still true relative to emissions.
10. Safe stage framing: "Nine months ago I said no subnet had serious revenue. Today three or four do, about $15–20M a year that you can check on-chain, and 14 teams buy back their token with it. It is still small next to what the protocol pays out. That is the next fight."

---

## Snapshot: the 15 subnets in this brief

Market cap and price from taostats.io embedded data, block 9,053,036, 12 Sep 2026 17:36 UTC. USD at $234.18. Emission share: bittensor.ai 7-day average / SubnetRadar, both 12 Sep 2026 (they disagree; see "Tracker disagreement"). Revenue: see each section.

| # | Subnet | Team | One-line product | Paying customers? | Alpha mcap (TAO / USD) | Mcap rank | Emission share (bittensor.ai 7d / SubnetRadar) |
|---|---|---|---|---|---|---|---|
| 1 | Lium SN51 | Datura AI Corp ("Fish") | GPU rental marketplace | Yes, $964K/month | 453,230 τ / $106M | 1 | 24.1% / 13.7% |
| 2 | Chutes SN64 | Rayon Labs (Chutes Global Corp) | Serverless inference in TEEs | Yes, ~$4.6M/yr | 411,410 τ / $96M | 2 | 9.3% / 9.7% |
| 3 | Targon SN4 | Manifold Labs (Austin) | Confidential GPU/CPU cloud | Yes, est. $5.5–6M/yr | 329,328 τ / $77M | 3 | 3.6% / 7.5% |
| 4 | Affine SN120 | Affine Foundation (Jacob founded; China-based team since Jun 2026) | Open post-training arena | No | 205,817 τ / $48M | 4 | 3.2% / 7.1% |
| 5 | Score SN44 (extra) | Score Technologies / Manako | Computer vision from camera feeds | Yes, $1–2M/yr est. | 203,352 τ / $48M | 5 | n/a / 5.3% |
| 6 | Teutonic SN3 | Jacob + Dendrite (unarbos) | Permissionless LLM pre-training | No | 163,404 τ / $38M | 6 | 1.8% / 4.1% |
| 7 | Engy SN53 | Hanlin AI (Ning Ren, SF) | Verified inference on consumer GPUs | Yes, ~$2K/day settled | 137,327 τ / $32M | 8 | 1.6% / 3.3% |
| 8 | IOTA SN9 | Macrocosmos Ltd (London) | Decentralized pre-training | Not yet | 131,633 τ / $31M | 9 | 9.4% / 2.8% |
| 9 | gm SN28 (extra) | Taostats team (London) | TEE gateway to Claude/GPT below list price | Yes, GMV only; est. $2.5–3M/yr | 97,946 τ / $23M | 12 | 1.3% / 2.4% |
| 10 | Bitcast SN93 (extra) | Bitcast | Creator marketing for brands | Yes, $390K cumulative | 96,164 τ / $23M | 13 | n/a / 3.1% |
| 11 | Hippius SN75 (extra) | The Nerve Lab | S3-style decentralized storage | Yes, small; mostly other subnets | 82,010 τ / $19M | 17 | n/a / 2.2% |
| 12 | Ridges SN62 (extra) | Ridges AI + Latent Holdings | Autonomous coding agents | Beta; $9.99/month plan | 59,074 τ / $14M | 19 | n/a / 1.1% |
| 13 | Oro SN15 | ORO AI (Vancouver; YC F26) | AI shopping-agent competition | No | 53,246 τ / $12M | 22 | 1.8% / 0% |
| 14 | Albedo SN97 | Dendrite (Jacob contributes) | King-of-the-hill coding-model arena | No | 36,974 τ / $8.7M | 29 | 1.5% / 3.1% |
| 15 | Conjectures SN66 | Pseudonymous (unarbos org) | Bounties for Lean-verified math proofs | No | 12,989 τ / $3.0M | 76 | 0.17% / 0.01% |

Total alpha market cap of all 128 subnets in the same snapshot: 4,841,631 τ ≈ $1.13B (taostats definition). SubnetRadar shows $1.33B the same day with a wider definition. Say "over a billion dollars".

---

## Part 1. Subnet profiles

### Lium (SN51)

- **Netuid / name**: 51, "lium.io". Formerly Celium; rebranded Aug 2025.
- **Team**: Datura AI Corp. Founder is pseudonymous: "Fish" (@fish_datura), a former large Bittensor miner. No disclosed HQ or funding round. Datura also runs Desearch (SN22).
- **What it produces**: A no-KYC marketplace where anyone can list NVIDIA GPUs (RTX 3090 up to B300) and anyone, human or agent, can rent them per second.
- **Customers**: 1,187 paying renters in the month reported 9 Sep 2026; 63% of rentals started by software agents; AI startups, researchers and other subnets' miners (Affine, Teutonic miners). No named enterprise customer. Not on OpenRouter (it sells GPU hours, not tokens).
- **Why it matters for the talk**: The cleanest "real customers, permissionless supply" story on the network. Anyone can add a GPU; anyone can rent one; the revenue buys and burns the token on-chain.

Numbers:

| Metric | Value | Source | Date |
|---|---|---|---|
| Monthly billing | $964K to 1,187 renters, +36% MoM | @lium_io tweet (team-disclosed) | 9 Sep 2026 |
| Annualized run-rate | ~$11.6M if flat (12 × $964K) | arithmetic on the above | 9 Sep 2026 |
| Index estimate | $8–10M/yr, "high confidence", #1 on index | SubConnect Revenue Index Vol. I | 26 Aug 2026 |
| Revenue-funded buyback and burn | $1,000,000 in one transaction (extrinsic 9032919-0005) | @fish_datura, taomarketcap (on-chain) | 9 Sep 2026 |
| Earlier revenue burns | $700K (5 Apr), >$150K (4 May), 2,500 TAO (22 Jun), ~$400K (5 Aug) | @fish_datura / @lium_io tweets | Apr–Aug 2026 |
| Signups / retention | 1,908 new signups; ~80% MoM retention; 7,697 rentals; median 1.15 h | @lium_io | 9 Sep 2026 |
| Supply | 68 datacenters, 21 countries | @lium_io | 9 Sep 2026 |
| GPUs online | 1,393 GPUs across 18 types; 1,188 rented (85% utilisation) | lium.io/api/executors/stats (live) | 12 Sep 2026 |
| Benchmark | Qwen3.6-35B at 14,499 tok/s on one B200 = $0.107 per M output tokens, "85% cheaper than OpenRouter" | @lium_io | 6 Sep 2026 |
| Market cap | 453,230 τ ≈ $106M; #1 non-root subnet | taostats snapshot | 12 Sep 2026 |
| Emission share | 24.1% (bittensor.ai 7d) / 13.7% (SubnetRadar) CONFLICT | trackers | 12 Sep 2026 |
| Exchange | SN51 alpha listed on Kraken | Kraken blog | 9 Jul 2026 |

Best recent posts (X: https://x.com/lium_io, 4,939 followers):

1. 9 Sep 2026, https://x.com/lium_io/status/2097824624117473549
   > https://lium.io Month in review:
   > $964k billed to 1187 renters (36% MoM growth).
   > 1,908 new signups, ~80% MoM user retention
   > 63% of rentals programatically initiated (agents)
   > 7,697 rentals; medium rental time: 1.15 hours.
2. 9 Sep 2026, founder, https://x.com/fish_datura/status/2097814414447481250
   > The masculine urge to buyback + burn $1,000,000.00 worth of my token with our product revenue.
   > https://taomarketcap.com/extrinsics/9032919-0005
3. 6 Sep 2026 (1,163 likes), https://x.com/lium_io/status/2096626330808828216
   > We just ran Qwen3.6 35B at 14,499 tokens per second.
   > on 1 lium GPU.
   > 85% cheaper than Openrouter.
4. 22 Jun 2026, https://x.com/lium_io/status/2069199663530471602
   > We just completed a buyback and burn of 2500 $TAO into subnet 51. All funds came from revenue from credit purchases for gpus.

Links: https://lium.io · https://docs.lium.io · https://github.com/Datura-ai/lium-io

Watch out: pseudonymous founder and offshore-style entity (friction for enterprises). Part of demand is other subnets renting GPUs (emission-funded, circular); the split is not disclosed. $964K is billings, not profit; providers take a share. Emissions to SN51 (hundreds of TAO/day) still exceed revenue.

---

### Chutes (SN64)

- **Netuid / name**: 64, "Chutes".
- **Team**: Rayon Labs; legal entity Chutes Global Corp (Nevis IBC). Founder Jon Durbin (@jon_durbin). New CEO Christian Hyland (ex-Coinbase, Toku, First Wave AI) named 7 Sep 2026. Bootstrapped, no VC round disclosed.
- **What it produces**: An OpenAI-compatible API for open-weight models (Kimi K3, DeepSeek V4, Qwen3.8, image/video/speech models), every model served inside a hardware-attested TEE by permissionless GPU miners.
- **Customers**: Developers and agent builders; roleplay/chat apps; OpenRouter users (20–25% of volume); other subnets (Affine, Albedo). 24,408 new subscribers in the 90 days to 30 Aug 2026 (team via Discord). Research partners: Harvard (Prof. Juncheng Yang) and UChicago. On OpenRouter: yes; ranked #1 provider in early 2026.
- **Why it matters for the talk**: The one Bittensor product whose demand outsiders can measure from a public API, co-authored a production-trace paper with Harvard, and it is a direct, cheaper, private competitor to the labs' inference APIs.

Numbers:

| Metric | Value | Source | Date |
|---|---|---|---|
| Revenue, last 90 days | $1.13M ($4.58M annualized); $214.6K of it subscriptions | team in Discord via SubnetRadar | 30 Aug 2026 |
| Index estimate | $4–5M/yr, "high confidence", #3 on index | SubConnect | 26 Aug 2026 |
| Earlier disclosure | >$1.3M revenue Nov 2025–Jan 2026, ~$820K fiat | team; repeated in SubConnect | Feb 2026 |
| Revenue vs costs | revenue "offsets roughly thirty to forty percent of operating expenses" | SubConnect quoting team | Aug 2026 |
| Revenue-funded burns | 20,944.85 alpha (~$320K) mid-Jul to mid-Aug 2026; 260,770 alpha (~$4M) cumulative | SubConnect (on-chain burn_alpha extrinsics) | Aug 2026 |
| Revenue per trillion tokens | $340K (all-time high); 90-day avg $282K | @chutes_ai "Week 10" | 21 Jul 2026 |
| Requests / tokens per day | 7-day avg 2.8M requests, 31.5B tokens (30 Jan 2025–12 Sep 2026 lifetime: 6.64B requests, 45T tokens, public LLM endpoints only) | api.chutes.ai/invocations/stats/llm (live) | 12 Sep 2026 |
| Peak tokens/day | 159.5B on 20 Jun 2025 | same endpoint | 12 Sep 2026 |
| One-year trace paper | 6.12B requests, 35.8T input tokens, 314,970 users, 9,174 models | arXiv 2608.13573 (Harvard/UChicago/Chutes) | trace Apr 2025–Apr 2026 |
| Market cap | 411,410 τ ≈ $96M; #2 | taostats snapshot | 12 Sep 2026 |
| Emission share | 9.3% (bittensor.ai 7d) / 9.7% (SubnetRadar) | trackers | 12 Sep 2026 |
| Pricing | Kimi K3 $3.00 in / $15.00 out per M tokens; cheapest tier $0.0245 in | @chutes_ai | Jul 2026 |

Best recent posts (X: https://x.com/chutes_ai, 11,553 followers):

1. 7 Sep 2026, https://x.com/chutes_ai/status/2096954255236042897
   > Today the Chutes team is happy to introduce our new CEO, Christian Hyland @cjohan247. [...] With a strong background across Coinbase, Toku, and First Wave AI, paired with an extensive skill set, he was a clear choice for CEO.
2. 26 Aug 2026, https://x.com/chutes_ai/status/2092605409429491730
   > >6 billion requests, trillions of tokens across over 9000 models.
   > Here's what we learned, in research led by Harvard assistant professor Juncheng Yang @1a1a11a
3. 24 Aug 2026 (273 likes), https://x.com/chutes_ai/status/2091888285748449320
   > We've been thinking a lot about this, lately... How do you build something that can last a thousand years?
   > Almost no company has done it. Most countries haven't either.
   > Chutes must become a network that outlasts any single company, team, or founder. Including us.
4. 28 Jul 2026, https://x.com/chutes_ai/status/2082142030210191660
   > Moonshot shipped Kimi K3's weights yesterday. 2.8T parameters, the largest open-weight model released to date. It is already live on Chutes.
   > Kimi K3 on Chutes: 1M context, native text + image + video input, $3.00 per million input tokens, $15.00 out. Runs inside a hardware-attested TEE with end-to-end encryption, like every model we serve.
5. 20 Jul 2026 (306 likes), https://x.com/chutes_ai/status/2079197210651001326
   > @jon_durbin pre-trained a 20B MoE for under $10 an hour of compute.
   > Not on a cluster. Eight rented single-L40S VMs scattered across two continents, plus a few 4090s and a 5090, all holding roughly 6 seconds per step.
   > Pre-training was supposed to be the part you couldn't do cheap.

Links: https://chutes.ai · https://api.chutes.ai · https://github.com/chutesai/chutes · paper https://arxiv.org/abs/2608.13573

Watch out: token volume fell from a 160B/day peak (Jun 2025) to ~31B/day (Sep 2026); the team says it deliberately cut free-tier traffic (tokens −45%, revenue −25%, Mar 2026). Independent analysts (Pine, ownyourmind) price only the OpenRouter-visible volume and get $1.3–2.4M/yr; the team dashboard says ~$4.6M. Emissions still cover 60–70% of costs by the team's own statement. Rayon also runs SN56 and controlled ~24% of all emissions in Apr 2026.

---

### Targon (SN4)

- **Netuid / name**: 4, "Targon". One of the first subnets (Oct 2023).
- **Team**: Manifold Labs Inc., Austin, Texas. CEO Rob Myers (@carromademethis); co-founder James Woodman (ex-COO Opentensor Foundation). $10.5M Series A (closed 28 Jul 2025) led by OSS Capital with DCG, Tobi Lütke, Ram Shriram, Jacob Steeves, Ala Shaabana. NVIDIA Inception member (Mar 2026).
- **What it produces**: A confidential compute cloud: GPU/CPU VMs, containers and inference that run inside Intel TDX / NVIDIA confidential-computing TEEs, so neither Manifold nor the machine owner can see the workload.
- **Customers**: Dippy AI (consumer companion app, ~50% of its text inference, "six-figure contract", Apr 2026); Ridges (SN62) and Score (SN44). GPU capacity "sells out on listing" (mid-2026). No named healthcare/finance/government customer despite the positioning. OpenRouter status in 2026 unverified.
- **Why it matters for the talk**: Verifiable privacy that hyperscalers only promise in contracts, with Intel engineers co-authoring the design, and Shopify's CEO as an investor. Manifold's mission page: "We do not believe Silicon Valley deserves a monopoly on AI talent and resources."

Numbers:

| Metric | Value | Source | Date |
|---|---|---|---|
| Index estimate | $5.5–6.0M/yr, "high confidence", #2 on index; "run rate approaching $6 million" | SubConnect | 26 Aug 2026 |
| Team claim (older) | ~$10.4M ARR, unaudited, no longer shown on targon.com | Manifold via analysts; flagged by Pine/ownyourmind | 2025–Apr 2026 |
| Growth claim | "Revenue also growing at 3.5x month on month" (no base given) | @carromademethis | 20 Aug 2026 |
| Revenue-funded buybacks | >550 TAO (>$115K) in the month to 20 Aug 2026; 2,000 TAO in the month to 29 May 2026 | @TargonCompute (wallet public) | May/Aug 2026 |
| GPUs | "1,500+ H200s" (team, 2025) vs a few dozen to ~230 cards visible on-chain CONFLICT | OSS Capital post; stats.targon.com/api/miners | Aug 2025 / 12 Sep 2026 |
| Intel whitepaper | "Decentralized Compute on Untrusted Hardware Using Intel TDX and Encrypted CVMs", co-authored with Intel | Intel community blog; arXiv 2607.21865 | 23 Mar 2026 |
| Market cap | 329,328 τ ≈ $77M; #3 | taostats snapshot | 12 Sep 2026 |
| Emission share | 3.6% (bittensor.ai 7d) / 7.5% (SubnetRadar) CONFLICT | trackers | 12 Sep 2026 |

Best recent posts (X: https://x.com/TargonCompute, 5,661 followers):

1. 20 Aug 2026, https://x.com/TargonCompute/status/2090532960005246982
   > Over the past month, we have repurchased over 550 TAO (equivalent to $115,000+) of SN4 alpha using organic revenues from the Targon Confidential Compute platform.
   > Our commitment to building a sustainable flywheel will continue with the growth of our permissionless compute market ☁️
2. 22 Jul 2026, https://x.com/TargonCompute/status/2079950828568293672
   > TargonOS is officially live. Confidential compute on Targon.com now enables:
   > → Rentable GPU/CPU VMs
   > → Root access from day one
   > → Hardware-attested isolation
   > → End-to-end encrypted workloads
   > Distributed compute, with confidentiality backed by industry leaders like @intel
3. 30 Jun 2026 (307 likes), https://x.com/TargonCompute/status/2072007227162980451
   > Exciting to see Chamath and J-Cal discussing the power of Targon distributed confidential compute on @theallinpod
   > "I do think this idea of distributed inference has a real place in the American ecosystem."
4. 6 May 2026 (350 likes), https://x.com/TargonCompute/status/2052063531093070071
   > Today we are excited to announce Targon Supply Portal. [...] → Permissionless: onboard directly to the Bittensor blockchain and start earning SN4 alpha, no conversations necessary. → Targon Managed: let us handle the blockchain, get paid out weekly in fiat or a currency of your choice.

Links: https://targon.com · https://docs.targon.com · https://github.com/manifold-inc/targon

Watch out: revenue is the least transparent of the top three (no dashboard; $10.4M claim vs $5.5–6M index). GPU count on-chain does not match marketing. Customers are mostly Bittensor-internal. Rob Myers said new Targon products launch at Exploit Conference; coordinate so the two talks do not collide.

---

### Engy (SN53)

Spelling confirmed by Jacob: "Engy". The team writes it lowercase "engy" on the site and on-chain. Website: https://engy.ai/.

- **Netuid / name**: 53, "engy", alpha symbol ب. The slot was a dead trading subnet (EfficientFrontier) with zero emissions from 21 Jan 2026; Engy took it over 4–6 Jul 2026, launched publicly 21 Jul 2026, emissions back on 23 Jul, permissionless mining open 25 Jul 2026.
- **Team**: Hanlin AI (hanlin.ai), San Francisco. Founder Ning Ren (@totheagi; ex-Google Brain, TensorFlow/TPU; GitHub `renning22`). Same team built SN11 TrajectoryRL.
- **What it produces**: A pay-per-token API (OpenAI- and Anthropic-compatible) that serves open frontier models (GLM-5.2/5.3, Kimi K3, DeepSeek V4 Flash, Qwen3.6/3.8) on mostly consumer GPUs, with a TOPLOC proof (a cryptographic fingerprint of the model's internal activations) attached to every response so a validator can check that the model you paid for is the model that answered.
- **Customers**: Developers and agent builders; a Discord test-key funnel for enterprises. Named Bittensor-internal users: Albedo SN97 (moved 100% of its DeepSeek and 40% of its GLM judge traffic to Engy), gm SN28 (resells Engy models), SN11 TrajectoryRL, Almanac. No named external enterprise yet; SubConnect says the team is "in discussions with enterprise clients representing potential contracts of up to $1 million in monthly revenue" (pipeline, not signed). OpenRouter listing: not confirmed.
- **Why it matters for the talk**: The sharpest "permissionless hardware beats the labs on price" example: a 2.8-trillion-parameter open model running on 80 gaming cards, sold at half the model maker's own price, with every answer verified. Jacob named it on This Week in Startups (17 Aug 2026).

Numbers:

| Metric | Value | Source | Date |
|---|---|---|---|
| Settled to miners, epoch 12 (2–9 Sep 2026, 7 days) | $14,318 (14,317,833,628 micro-USD); 77 miners; 3,566,420 requests; 56.0B tokens | provider.engy.ai/api/subnet/v1/epoch/latest (public, signed ledger) | fetched 12 Sep 2026 |
| Implied daily / annual | ~$2,045/day settled; ~$747K/yr if flat. Note: "settled" is the value of miner-served traffic at Engy's rate card; it excludes first-party-served models and is a proxy for billed traffic, not audited revenue | arithmetic on the ledger | 12 Sep 2026 |
| Index estimate | $400–600K/yr, "high confidence"; "$1,000–$2,000 in daily revenue, with 100% of revenue being used for Alpha buybacks" | SubConnect | 26 Aug 2026 |
| User deposits | ">$150,000" since 21 Jul launch | team via TAO Daily | 11 Aug 2026 |
| First 3 epochs (22–25 Jul) | $10.59, $69.98, $42.37 settled | SubnetRadar deep dive from the same ledger | 25 Jul 2026 |
| Model mix, epoch 12 | 64 Qwen3.6-35B miners, 5 Kimi K3, 4 GLM-5.2, 1 each DeepSeek V4 Flash / GLM-5.3 / GLM-5.3-flash / Qwen3.8 | ledger | 12 Sep 2026 |
| GPUs | 944 GPUs across 67 providers (560 RTX 5090, 222 RTX PRO 6000) | DeAI Dashboard (third-party) | 12 Sep 2026 |
| Prices (per M tokens in/out) | glm-5.3 $0.98/$3.08; glm-5.2 $0.68/$1.50; kimi-k3 $1.95/$9.75; deepseek-v4.1-flash $0.04/$0.08; qwen3.8-27b $0.045/$0.32 | engy.ai (live) | 12 Sep 2026 |
| Price vs makers | GLM-5.2 66% below Z.ai ($1.40/$4.40); Kimi K3 launched at $1.50/$7.50 vs Moonshot $3/$15 (raised to $1.95/$9.75 later) | SubnetRadar; @engyai | Jul–Aug 2026 |
| Market cap | 137,327 τ ≈ $32M; #8 | taostats snapshot | 12 Sep 2026 |
| Emission share | 1.6% (bittensor.ai 7d) / 3.3% (SubnetRadar) | trackers | 12 Sep 2026 |
| Gateway uptime | 99.72% over 90 days | engy.ai/status | 12 Sep 2026 |

CONFLICT: DeAI Dashboard shows "$1.73M/24h billed volume" for Engy. The team's own ledger shows ~$2K/day. Treat the dashboard number as wrong.

Best recent posts (X: https://x.com/engyai, 2,280 followers; founder https://x.com/totheagi):

1. 7 Aug 2026 (218 likes), https://x.com/engyai/status/2085733487710228816
   > DeepSeek V4 Flash is live on engy.ai.
   > 50% cheaper than OpenRouter.
   > 68% cheaper than DeepSeek Official.
2. 27 Jul 2026, founder (1.0M views), https://x.com/totheagi/status/2081855316443205717
   > we got the full Kimi K3, 2.8T params, running on 80x RTX 5090s.
   > 20 tok/s single stream, day one, untuned. Last week we took GLM-5.2 from 30 to 110 tok/s on this same fleet. This number will climb.
   > A first for open weights: frontier intelligence served with zero HBM, the scarcest silicon in AI. Just GDDR7 gaming cards, plain ethernet, and the official MXFP4 weights, nothing requantized.
   > The most powerful open model on Earth, on the most abundant GPUs on Earth. Any lab, startup, or university can now own it, probe it, fine-tune it, run agents on it.
3. 4 Aug 2026 (201 likes), https://x.com/engyai/status/2084656485397885400
   > Kimi K3 is live on engy.ai.
   > next: DeepSeek V4 Flash and Qwen3.8.
4. 15 Aug 2026, https://x.com/engyai/status/2088597633606832467
   > Qwen3.8 27B is live on https://engy.ai now.
   > Image input is supported.
   > input (M): $0.045
   > output (M): $0.32
   > cached (M): $0.015
5. Third-party, Dendrite/Albedo, 11 Aug 2026, https://x.com/DendriteHQ/status/2087169458402189560
   > On Albedo/SN97, we've moved most of its LLM traffic from OpenRouter to @engyai - 100% of DeepSeek and 40% of GLM
   > Worst-case request stalls dropped from 75s to under 2s. Evaluations are now 10-30% faster

Links: https://engy.ai · https://engy.ai/pricing · https://provider.engy.ai · https://github.com/hanlinai/engy · https://hanlin.ai

Watch out: one "master validator" (the owner's hotkey) does all GPU verification and held ~49.5% of validator stake in July; other validators only check its signature. Self-mining accusations in Discord (Sep 2026) are unanswered beyond a denial. No slashing shipped yet; a cheating miner only loses the epoch. Verification is "sampled and probabilistic, not a full proof" (team's own one-pager). Revenue is real but small (~$2K/day). Alpha price ran +350% in 30 days before there was a product to price.

---

### Affine (SN120)

- **Netuid / name**: 120, "Affine", tagline "Reason Mining".
- **Team**: Founded by Jacob Steeves (Jul 2025). Ownership handed in Jun 2026 to a China-based team ("Tobias, Allan, bignickeye"); operates as the Affine Foundation. Jacob is listed as "chief scientist at Affine" (@affine_io, 13 Aug 2026).
- **What it produces**: A permanent post-training competition: miners fine-tune an open model (Qwen3.6-35B-A3B; from Jul 2026 also LoRA adapters on GLM-5.2 744B) and duel the reigning "king" on coding, terminal, tool-use and math environments; a challenger must beat the king by a margin to take the crown and the emissions; winning weights are published.
- **Customers**: None. No API is sold. Emissions-funded by design. Uses Lium/Targon for compute and Chutes for inference.
- **Why it matters for the talk**: Jacob's own answer to closed post-training pipelines: "a lab that's just a market mechanism". Also the best example of subnets composing: Affine miners buy Lium GPUs and Chutes inference with their emissions.

Numbers:

| Metric | Value | Source | Date |
|---|---|---|---|
| Model gains after base switch | Memory 11.7→70.2; NavWorld 33.8→48.3; SWE 33.1→58.2; Terminal 46.2→78.1 (AFFINE-35B-III vs base) | @affine_io | 24 Jun 2026 |
| Reward per slot | "each slot pays roughly $50,000 USD per week" (5 slots) | @affine_io | 23 Aug 2026 |
| Market cap | 205,817 τ ≈ $48M; #4 | taostats snapshot | 12 Sep 2026 |
| Emission share | 3.2% (bittensor.ai 7d) / 7.1% (SubnetRadar) | trackers | 12 Sep 2026 |
| Active miners | 3 validators / 5 miners paid (253 UIDs registered) | bittensor.ai; SubnetRadar | 12 Sep 2026 |
| External revenue | none | SubnetRadar; SubConnect (not listed) | 12 Sep 2026 |

Best recent posts (X: https://x.com/affine_io, 2,446 followers):

1. 23 Aug 2026, https://x.com/affine_io/status/2091498650119901215
   > We think the best Affine miner probably hasn't heard of Bittensor yet.
   > If you're an AI engineer or independent researcher who spends your time squeezing more capability out of open models, Affine gives that work a permissionless arena with real stakes. You don't need prior Bittensor experience, an application, or anyone's approval to compete.
   > Beat the current king under Affine's evaluation and you earn a place among the five models being rewarded. At current prices and emissions, each slot pays roughly $50,000 USD per week.
2. 27 Aug 2026, https://x.com/affine_io/status/2093008057261609231
   > One side effect of putting real money behind an AI evaluator is that the evaluator itself becomes an optimization target. [...] A live competition exposes the gap between what you intended a metric to reward and what it actually rewards under pressure.
3. 24 Jun 2026, https://x.com/affine_io/status/2069810780019491188
   > Qwen3.6-35B-A3B now forms the new substrate beneath Affine's open arena. By the very next day, the first champion had already been crowned. Following that, AFFINE-35B-III came along to break the frontier wide open:
   > Memory: 11.7 → 70.2 / NavWorld: 33.8 → 48.3 / SWE: 33.1 → 58.2 / Terminal: 46.2 → 78.1
   > [...] The base changed. The mechanism sharpened it.
4. 24 May 2026, https://x.com/affine_io/status/2058546037632159896
   > Honored to be invited by @Alibaba_Qwen and @alibaba_cloud to attend the very first global Qwen Conference in Singapore. [...] Affine will be joining the event on May 26, 2026 at the Sands Expo & Convention Centre, Singapore, where we'll give a special speech about Affine env design and post-training technics
5. Jacob, 29 Nov 2025, https://x.com/const_reborn/status/1994763568399749199
   > Affine miners train on Lium and sample on Chutes. Miner sell pressure is just buy pressure on Bittensor itself. It helps that those commodities are themselves the cheapest in the world. As competition increases the entire stack squeezes, there is no waste.

Links: https://affine.io · https://github.com/AffineFoundation/affine-cortex

Watch out: winner-take-all pays ~$86K/day of emissions to 1–5 anonymous wallets; the scorer was gamed at least twice (team admitted it, 27 Aug 2026); kings were found to be copies (Jul 2026); frequent rule changes. Jacob wrote on 20 Dec 2025 that he has "never bought Affine" and had considered turning off its emissions because of "self play" insinuations. Expect that question.

---

### Teutonic (SN3)

- **Netuid / name**: 3, "Teutonic". This was Templar (Covenant AI) until Covenant left Bittensor on 9–10 Apr 2026. Holders voted the slot to Jacob's team ~14 Aug 2026; renamed 15 Aug 2026.
- **Team**: Paper authors: Jacob Steeves, Krzysztof Warchoł, Hubert Korzeniewski, Antoni Bombała. Code under the `unarbos` GitHub org; models under Hugging Face `dendriteholdings`; Dendrite (@DendriteHQ) co-builds and mines. No official Teutonic X account (teutonic_ai is an unofficial fan account). Jacob's @const_reborn is the announcement channel.
- **What it produces**: Open-weight LLMs pre-trained through a "competition parallel" market: miners submit full checkpoints, validators run paired loss duels against the current king, a challenger is crowned only if a statistical lower bound shows a real loss drop, and the last five kings share emissions.
- **Customers**: None. Emissions-funded (about $50K/day of TAO to SN3, bittensor.ai, 12 Sep 2026).
- **Why it matters for the talk**: It is Jacob's direct test of "compute + talent coordinated by a digital market can out-train a lab", and it is the subnet he personally leads. Its 110B run is the biggest permissionless training attempt so far.

Numbers:

| Metric | Value | Source | Date |
|---|---|---|---|
| Teutonic-I run | 2 Jun–10 Aug 2026; 2,163 duels; 203 coronations; 73 coldkeys; 8.6B params (marketed 10B) | teutonic.ai/paper.html | Aug 2026 |
| Benchmark claim | 62.28% mean on 11 tasks vs Covenant-72B 57.55%, INTELLECT-1 45.38% | paper; @DendriteHQ | 14 Aug 2026 |
| Counter-claim | Qwen3.5-9B-Base scores 62.48% on the same tasks; Teutonic-I weights correlate 0.96 with Qwen3.5-9B-Base; a 4 Jun king shared 56% exact weights with an existing model CONFLICT | Templar audit (tplr.ai/teutonic-1-audit.pdf; @tplr_ai) | 20 Aug 2026 |
| Teutonic-II | ~110B total / ~7.3B active MoE (MiMoV2-based), "50% larger than any prior decentralized run" | Hugging Face genesis card; @const_reborn | 24 Aug 2026 |
| Teutonic-II interim | MMLU 69%, GSM8K 72.3%, PIQA 82.9% | SubnetRadar digest | 6 Sep 2026 |
| Market cap | 163,404 τ ≈ $38M; #6 | taostats snapshot | 12 Sep 2026 |
| Emission share | 1.8% (bittensor.ai 7d) / 4.1% (SubnetRadar) | trackers | 12 Sep 2026 |
| Reach | 14 Aug tweet 2.6M views, 1,047 likes; 24 Aug tweet 529K views | fxtwitter | 12 Sep 2026 |

Best recent posts (announced from https://x.com/const_reborn):

1. 14 Aug 2026 (1,047 likes), https://x.com/const_reborn/status/2088349997645312205
   > Today, after 4 months, Subnet 3 (previously @tplr_ai) got retaken by holders through vote.
   > We're celebrating by releasing our first model.
   > Teutonic-I: the worlds most performant decentralized LLM.
   > https://teutonic.ai/paper.html
2. 24 Aug 2026 (852 likes), https://x.com/const_reborn/status/2091896286035235027
   > Announcing Teutonic-II-110B: a frontier permissionless training of a 110B LLM
   > We are scaling 10x beyond our previous model and 50% larger than any prior decentralized run.
3. 24 Aug 2026, https://x.com/const_reborn/status/2091896298400047207 (thread)
   > Teutonic does this by building a "loss based market" -- paying for model updates produced by any compute, any individual, any algorithm.
   > This makes the run's success dependent purely on your ingenuity.
4. 14 Aug 2026, Dendrite, https://x.com/DendriteHQ/status/2088353208267858385
   > The new king of decentralized AI has just been crowned.
   > A 10B model just beat 72B.
   > Teutonic-I 10B: 62.28%
   > Covenant-72B: 57.55%
   > Same 11-benchmark average.
   > SN3 has changed hands. @const_reborn is the owner.
5. Counter-post, Templar, 20 Aug 2026, https://x.com/tplr_ai/status/2090433758977778074 (links the audit PDF).

Links: https://teutonic.ai · https://teutonic.ai/paper.html · https://github.com/unarbos/teutonic · https://huggingface.co/dendriteholdings/Teutonic-I

Watch out (important for Jacob): the provenance dispute. Jacob conceded on Discord that miners used weights from Quasar, which took weights from Qwen. Sam Dare (Covenant) called it "somewhere on the spectrum between gross negligence and fraud" (synapz.org, 21 Aug 2026). Safe framing on stage: "the market found the cheapest path to lower loss, which included borrowing open weights; that is a verification problem we are fixing in Teutonic-II" rather than "the world's most performant decentralized LLM". The 80B run announced 11 May 2026 never produced a model.

---

### IOTA / Macrocosmos (SN9)

- **Netuid / name**: 9, "IOTA" (Incentivized Orchestrated Training Architecture).
- **Team**: Macrocosmos Ltd, London (company no. 15583737, inc. Mar 2024). CEO Will Squires (@WSquires), CTO Steffen Cruz (@macrocrux), AI lead Alan Aboudib. ~20–30 staff. Also runs SN1 Apex and SN13 Data Universe.
- **What it produces**: A permissionless pre-training system that splits a model into pipeline stages, gives each stage to a different miner, compresses activations 128× so the pieces can talk over the public internet, and lets anyone launch a training run (up to ~100B parameters) by paying TAO. Plus "Train at Home", an app that lets laptops and gaming GPUs join.
- **Customers**: None paying yet. CTO Cruz (May 2026): "we don't have any customers in yet... startups already committed... second half of this year." Target pricing 10–20% of centralized training cost (Substack, 14 Jul 2026).
- **Why it matters for the talk**: The strongest evidence that permissionless training reaches lab scale on rented and consumer hardware, and a live demo Jacob has already done: a 128-node run launched in 5 minutes for 5 TAO.

Numbers:

| Metric | Value | Source | Date |
|---|---|---|---|
| Orion-100B | 100B params; 48 A100s in 5 US datacenters; ~65% of colocated efficiency; trained only to 1.1B tokens (a systems proof, not a usable model) | Macrocosmos Substack; IOTA primer | 1–2 Jun 2026 |
| Orion-16B | 16B trained on RTX 4090/5090s (Green Compute + Targon), MFU >40%, USA to Hong Kong, >100B tokens by Aug | @mccrinbc; SubnetRadar | 27 Jul / Aug 2026 |
| Jacob's run | 128 nodes, 3B params, launched in 5 minutes for ~5 TAO (~$1,150) | @const_reborn | 3 Jun 2026 |
| Train at Home | 23K+ downloads; participants in 16 countries | @MacrocosmosAI; @IOTA_SN9 | 20 Apr / 30 Jun 2026 |
| Paper | IOTA, arXiv 2507.17766 | arXiv | Jul 2025 |
| Market cap | 131,633 τ ≈ $31M; #9 | taostats snapshot | 12 Sep 2026 |
| Emission share | 9.4% (bittensor.ai 7d) / 2.8% (SubnetRadar) CONFLICT | trackers | 12 Sep 2026 |
| Security | 12 wallets drained 9 Aug 2026; 950 TAO of excess emissions burned 8 Sep 2026 | SubnetRadar digests | Aug–Sep 2026 |

Best recent posts (X: https://x.com/IOTA_SN9, 2,733 followers; https://x.com/MacrocosmosAI, 7,610):

1. Macrocosmos, 1 Jun 2026 (418 likes), https://x.com/MacrocosmosAI/status/2061493162582118695
   > Today, we are launching the first stage of Project Orion.
   > Our early pre-training run of Orion-100B achieves upward of 65% of data-center training efficiency on hardware costing a fraction of the price.
   > Orion-100B is the first proof point for a simple idea: that underutilized compute around the world can be turned into frontier training capacity.
   > We believe that this work presents, for the first time, an economically compelling case for training large models using distributed approaches.
2. Jacob, 3 Jun 2026 (866 likes), https://x.com/const_reborn/status/2062258366739550618
   > Just orchestrated a 128 node permissionless decentralized training run, in 5 minutes, for 5 TAO, via @IOTA_SN9
   > They can do this up to 100B param models.
   > Unbelievable.
3. Macrocosmos, 27 Jul 2026, https://x.com/MacrocosmosAI/status/2081791889674772618
   > A 16-billion parameter model is now training live on @IOTA_SN9; globally distributed over three continents, heterogeneous, permissionless, owned by no single entity.
   > Today, we present the next stage of Project Orion - Orion-16B.
4. IOTA, 16 Jul 2026, https://x.com/IOTA_SN9/status/2077790675932680220
   > In June we launched Project Orion: our mission to build large, enterprise-scale models on iota's architecture.
   > We started with our 100B run - to our knowledge, the largest distributed LLM pretraining run over the open internet. It demonstrated iota's capabilities and proved the strength of our network.
5. IOTA, 30 Jun 2026, https://x.com/IOTA_SN9/status/2071985168966184998
   > Train at Home is being used around the globe.
   > At the moment, there are participants in 16 countries across three continents.

Links: https://iota.macrocosmos.ai · https://docs.macrocosmos.ai/subnets/subnet-9-iota · https://github.com/macrocosm-os/iota · https://arxiv.org/abs/2507.17766

Watch out: zero paying customers as of the last statement found; Orion-100B is a systems demo (1.1B tokens); a wallet-drain incident in Aug 2026; emission share swings widely between trackers.

---

### Oro (SN15)

- **Netuid / name**: 15, "ORO". Public launch 20 Apr 2026.
- **Team**: ORO AI, Vancouver, Canada. CEO Shardul Bansal (@shardiban), CTO Seth Schilbe; both ex-AWS distributed systems. Accepted to Y Combinator Fall 2026 (announced 8 Sep 2026), reported as the first crypto startup with a live token admitted to YC.
- **What it produces**: A daily competition where anyone submits a Python shopping agent; validators run it in sandboxed Docker containers against ShoppingBench (2.5M real products); the best agents earn emissions and their trajectories become training data for Oro's own open shopping model.
- **Customers**: None yet. Emissions-funded. Stated direction: "the best consumer agentic commerce experience" and a "trust layer" for agent shopping.
- **Why it matters for the talk**: A token-paid arena produced agents that, by Oro's claim, beat GPT-5.4 on a shopping benchmark within weeks, and mainstream VC (YC) now backs the model. A Canadian team for a Montreal audience.

Numbers:

| Metric | Value | Source | Date |
|---|---|---|---|
| Paid to miners | $200,000+ in 36 days; 1,206 unique miners | @oroagents | 20 May 2026 |
| Weekly payout | "$80,000+" per week; up to $20,000/day to a top agent | @oroagents | 14 May / 26 Jun 2026 |
| Trajectories | ~20,000 per day; post-training lifted Qwen3-4B from 18% to 42% | @oroagents | 1 Jun 2026 |
| Claim vs GPT-5.4 | "45 of our agents have beaten GPT 5.4"; best 63.0% (unverified by third parties) | Oro LinkedIn | 20 Apr 2026 |
| Hack | ~147,000 alpha (~$630K) drained from owner wallet; attributed to North Korea's Sapphire Sleet | tao.media; Protos; team post-mortem | 13 Jul 2026 |
| Market cap | 53,246 τ ≈ $12M; #22 | taostats snapshot | 12 Sep 2026 |
| Emission share | 1.8% (bittensor.ai 7d) / 0% (SubnetRadar, flagged "earning ~0%") CONFLICT | trackers | 12 Sep 2026 |

Best recent posts (X: https://x.com/oroagents, 2,476 followers):

1. 8 Sep 2026 (344 likes), https://x.com/oroagents/status/2097396633764032707
   > We're excited to announce that we're going to be joining Y Combinator in Fall 2026.
   > The team is super pumped to be working with @golda and co to continue on our journey of creating the best in class open source models for agentic commerce.
2. 1 Jun 2026, https://x.com/oroagents/status/2061441585535918535
   > We're excited to announce the next stage of our subnet: an automated post-training pipeline that will enable us to build the best product for AI shopping.
   > It's only been 50 days, but we're now getting 20k high-quality trajectories per day that are rich training signals for online shopping tasks.
   > Using the trajectories we have thus far, we saw an 18% → 42% climb on Qwen3-4B base using our post-training pipeline.
3. 14 May 2026, https://x.com/oroagents/status/2054985113612722559 (long post, excerpt)
   > We're building the trust layer for AI shopping, powered by open-source competition on Bittensor. Each week, we pay miners from around the world $80,000+ to compete to build the best shopping agent. And it's working – we're nearing 4000 agents submitted in just a few weeks, with hundreds added every day.
4. Jacob, 20 Apr 2026 (370 likes), https://x.com/const_reborn/status/2046344664681156770
   > Mine Oro: the unstoppable agent machine.
   (quoting Oro's launch post: "Introducing Oro, the biggest agent competition in the world. Powered by Bittensor, Live Now." https://x.com/oroagents/status/2046288314295373953, 1.04M views)

Links: https://oroagents.com · https://docs.oroagents.com · https://github.com/ORO-AI/oro · whitepaper https://app.oroagents.com/whitepaper

Watch out: no customers; the GPT-5.4 claim is self-reported; the Jul 2026 hack of a software wallet; ongoing crackdowns on miners hard-coding answers (12 Sep 2026).

---

### Albedo (SN97)

- **Netuid / name**: 97, "Albedo" ("Alchemical intelligence"). Registered Apr 2026 as "Constantinople", then "distil", then Albedo from late May 2026.
- **Team**: Dendrite (@DendriteHQ; dendrite.holdings; "Co-builders: SN3 Teutonic, SN66 Conjectures, SN97 Albedo"). GitHub `unarbos/albedo`; contributors include `unconst` (Jacob). The original mechanism was launched by "Arbos", an autonomous agent persona Jacob created (@arbos_born). Lead dev account: @weather_boss_.
- **What it produces**: A king-of-the-hill arena for coding agents: miners fine-tune Qwen3.6-35B-A3B on software-engineering trajectories and challenge the reigning king in 100–128-task matches drawn at random from a 12-million-trajectory dataset, judged by an ensemble of frontier LLMs; the winner's weights are published and served free.
- **Customers**: None; no revenue. Kings published on Hugging Face (`dendriteholdings`).
- **Why it matters for the talk**: A subnet launched by an AI agent, handed to an independent studio, where anyone can dethrone the incumbent coding model and the weights are public. It is also a lesson: the eval machine was compromised in July and the team fixed it in public.

Numbers:

| Metric | Value | Source | Date |
|---|---|---|---|
| Dataset | 12M trajectories, 122,000 PRs, 3,000 repos, 16 languages, 112B tokens | tao.media | 13 Jul 2026 |
| Genesis model | Qwen3.6-35B-A3B, 73.4% SWE-Bench Verified | tao.media | 13 Jul 2026 |
| Eval | 8-step conversations, 100 samples, three judge models, ~$1.50 per eval | SubnetRadar digest | 23 Jul 2026 |
| Market cap | 36,974 τ ≈ $8.7M; #29 | taostats snapshot | 12 Sep 2026 |
| Emission share | 1.5% (bittensor.ai 7d) / 3.1% (SubnetRadar) | trackers | 12 Sep 2026 |
| Stake flow | total staked −42% over 7 days | bittensor.ai | 12 Sep 2026 |
| Dev activity | 202 commits / 30 days | SubnetRadar | 12 Sep 2026 |

Best recent posts (X: https://x.com/DendriteHQ, 616 followers):

1. 11 Aug 2026 (118 likes), https://x.com/DendriteHQ/status/2087169458402189560
   > On Albedo/SN97, we've moved most of its LLM traffic from OpenRouter to @engyai - 100% of DeepSeek and 40% of GLM
   > Worst-case request stalls dropped from 75s to under 2s. Evaluations are now 10-30% faster
   > Great work by @AlgodTrading, @totheagi and the Engy team
2. 10 Jul 2026, https://x.com/DendriteHQ/status/2075615123935474135
   > GM
   > Dendrite officially started mining on SN3 Teutonic and SN97 Albedo.
   > One rule above all: max transparency.
   > One public key. Every method documented. Everything verifiable.
3. 18 Aug 2026, https://x.com/DendriteHQ/status/2089729165721551040
   > 70 days. 2,163 duels. 203 coronations. One 10B model left standing.
   > Teutonic-I beat Covenant-72B's 11-benchmark average with 7x fewer parameters, without synced gradients or a fixed recipe.
4. Arbos (origin account), 2 May 2026, https://x.com/arbos_born/status/2050596495485800658
   > A 4B-parameter model on SN97 is now scoring 0.94 on HumanEval — beating its 35B-parameter teacher (0.872) by nearly 7 points on code generation.
   > 8.75× smaller. Better at the actual task. On consumer hardware.

Links: https://github.com/unarbos/albedo · https://dendrite.holdings · https://huggingface.co/dendriteholdings

Watch out: eval machine on Lium was compromised (model substitution) in Jul 2026; five kings restored; Dendrite removed its own 20% miner allocation after self-mining criticism. Three names and three X accounts in five months. Small market cap, large stake outflow.

---

### Conjectures (SN66)

- **Netuid / name**: 66, "Conjectures" (conjectures.io). Launched 6 Aug 2026 on a slot the same team had used for a prior subnet ("Ninja").
- **Team**: Pseudonymous; GitHub org `conjectures-io` maintained under `unarbos`/`unconst`. Dendrite lists itself as co-builder. Note: the site links to Levent Alpöge and Dmitry Rybin only as examples of AI-assisted solves that did NOT come from the subnet ("Neither came from this subnet."). They are not the team.
- **What it produces**: Machine-checked Lean 4 proofs or counterexamples for open math problems (Erdős problems and Ben Green's open problems, drawn from Google DeepMind's formal-conjectures repository), with TAO bounties paid to whoever submits a proof that passes the deterministic kernel check.
- **Customers**: None. 100% of emissions go to a treasury that funds bounties. Roadmap mentions sponsored problems (companies pay to post) and USDC payment; not live.
- **Why it matters for the talk**: A market where the "customer" is a proof checker. Anyone, human or AI, anywhere, gets paid for settling a question that stood open for years. In its first five weeks it paid for several Erdős/Green problems.

Numbers:

| Metric | Value | Source | Date |
|---|---|---|---|
| Solve announced | "near-half-density case of Green's Problem 51, open for 16 years", verified in Lean | @conjectures_io | 9 Sep 2026 |
| Payouts | 4 solves approved 12 Sep (Erdős 96, Green 15, Erdős 196, Erdős 272); 2 approved 10 Sep; first payout 4,130 alpha + two 750-alpha partials 7 Aug | SubnetRadar digests | Aug–Sep 2026 |
| Catalog | 162 audited targets (142 Erdős + 20 Green); prize pool "over $16,000"; per-problem bounty ~$3.1–3.6K; 0.5 TAO submission fee | subnetalpha.ai; conjectures.io | Sep 2026 |
| Sponsorship | Gold sponsor of the Caltech Mathathon (100 teams; alongside Anthropic and Cognition) | @conjectures_io | 2 Sep 2026 |
| Market cap | 12,989 τ ≈ $3.0M; #76 | taostats snapshot | 12 Sep 2026 |
| Emission share | 0.17% (bittensor.ai) / 0.01% (SubnetRadar); 19% above the deregistration floor | trackers | 12 Sep 2026 |
| Price | −76% from 7-day high | SubnetRadar | 12 Sep 2026 |

Best recent posts (X: https://x.com/conjectures_io, 689 followers):

1. 9 Sep 2026, https://x.com/conjectures_io/status/2097747177606967570
   > Today, we're announcing a solution found by our miners to the near-half-density case of Green's Problem 51, open for 16 years.
   > The result establishes how much structure must exist in sums of dense sets. Verified in Lean through Conjectures.
   > Full proof below.
2. 11 Sep 2026 (164 likes), https://x.com/conjectures_io/status/2098411910249009375
   > Conjectures, the incentive mechanism for mathematical discovery.
3. 2 Sep 2026, https://x.com/conjectures_io/status/2095174640234586353
   > We're excited to join the Mathathon Challenge at @Caltech as a Gold Sponsor.
   > The event will bring together 100 teams from around the world, from IMO gold medalists to mathematics professors and researchers at frontier labs, to work on some of the hardest open problems in mathematics.
   > The challenge will include problems from Conjectures, each with an additional bounty for the team that solves it.
   > We'll also have a table at the event alongside sponsors including @AnthropicAI and @cognition.
4. 3 Sep 2026, https://x.com/conjectures_io/status/2095528659721257353
   > You no longer need a complete solution to get paid.
   > Useful partial work in Lean can now earn a share of the bounty: lemmas, definitions, APIs, special cases, tactics, partial proofs, and checked counterexample bounds.

Links: https://conjectures.io · https://conjectures.io/faq · https://github.com/conjectures-io/conjectures-validator

Watch out: tiny and near deregistration; payouts are manual via a 2-of-3 multisig (verification is trustless, payment is not); bounties are small; the two famous July solves were not from the subnet. Good for a "what a market can pay for" anecdote, not for scale claims.

---

### Extra 1. gm / SayGM (SN28)

Added because Jacob praised it twice in September and because it is the only subnet reselling the labs' own frontier models below list price.

- **Netuid / name**: 28, on-chain "SayGM"; brand "gm" / "Good Morning". Beta live 28 May 2026.
- **Team**: Built by the Taostats team (GitHub org `taostat`; miners log in via auth.taostats.io; Marcus Graichen). London.
- **What it produces**: "OpenRouter inside a TEE": your prompt goes into an Intel TDX confidential VM, then to miners who bring their own Anthropic/OpenAI/Google/Chutes/Engy keys and bid a discount below list price. Users get Claude Fable 5.1 and GPT-6 Astra at 8–15% off, and open models at up to 65% off, with emissions subsidising the miners' thin margins.
- **Customers**: Developers and privacy-sensitive users buying credits; no enterprise names. Not on OpenRouter; it competes with OpenRouter.
- **Why it matters for the talk**: A Bittensor market undercut the labs on their own models and listed GPT-6 Astra a week before OpenRouter did. Jacob: "GM had astra before openrouter."

Numbers:

| Metric | Value | Source | Date |
|---|---|---|---|
| Cumulative tokens | 96.48B network tokens processed, >30B output | @say_gm_ | 7 Sep 2026 |
| Index estimate | $2.5–3M/yr, "medium confidence"; ~$60K beta revenue earlier; buybacks not disclosed | SubConnect | 26 Aug 2026 |
| Traffic value | $636.66 served / 11,726 requests / 40 miners in one 72-minute epoch (≈$12.7K/day gross if steady) | saygm.com/miners ledger | 12 Sep 2026 |
| What Jacob linked ("Insane numbers") | Claude Fable 5 at $8.43 in / $42.15 out per M tokens, 15.7% below Anthropic; today 15.0% off, 3 providers, 61.8 tok/s, 1,685 requests / 18.0M tokens billed in the last 12 windows | saygm.com/models/claude-fable-5 | 1 / 12 Sep 2026 |
| Catalog | 33 models, 36% average discount | @say_gm_ | 12 Aug 2026 |
| Market cap | 97,946 τ ≈ $23M; #12 | taostats snapshot | 12 Sep 2026 |
| Emission share | 1.3% (bittensor.ai 7d) / 2.4% (SubnetRadar) | trackers | 12 Sep 2026 |

Best recent posts (X: https://x.com/say_gm_, 1,017 followers):

1. 4 Sep 2026, https://x.com/say_gm_/status/2095786804146569348
   > GPT-6 Astra is live on Gm.
   > OpenAI's smartest model yet. Tell it what you need done. It opens the browser, writes the code, runs the terminal, ships the result.
   > Routed through a verifiable TEE, at 8.3% discount on SayGm.
2. Jacob, 4 Sep 2026 (306 likes), https://x.com/const_reborn/status/2095850305774641562
   > GM had astra before openrouter
3. Jacob, 1 Sep 2026 (300 likes), https://x.com/const_reborn/status/2094904943853367696
   > https://saygm.com/models/claude-fable-5
   > Insane numbers from gm. Congrats to the team.
4. 7 Sep 2026, https://x.com/say_gm_/status/2096886191081615619
   > GM is approaching 100 billion recorded network tokens processed.
   > We're at 96.48B.
   > Including more than 30B output tokens generated.
5. 26 Aug 2026, https://x.com/say_gm_/status/2092613038700658942
   > DeepSeek V4 Flash is 75% cheaper on saygm than list price.
   > Not a promo code. Not a limited window. Live provider bidding, updated in real time.
   > Openrouter shows you the sticker price. We show you what it actually costs.

Links: https://saygm.com · https://saygm.com/miners · https://docs.saygm.com · https://github.com/taostat/gm-miner

Watch out: the discounts are paid for by emissions; the team says margins are "tight" and is redesigning incentives (Sep 2026). It resells lab APIs under their terms of service. Gross traffic value is not gm's revenue; nobody publishes gm's net take.

---

### Extra 2. Score (SN44)

Added for market cap (#5) and the only named Big-Four customer on the network.

- **Netuid / name**: 44, "Score".
- **Team**: Score Technologies (UK); Manako (@manakoai) is the enterprise go-to-market arm. Team of 13.
- **What it produces**: Small, fast computer-vision models that detect and track every player, car or object in ordinary camera feeds, trained and validated by competing miners, sold as real-time analytics.
- **Customers**: PwC France (alliance, Apr 2026), Reading FC, Eyeball Football, AVIA fuel stations (~400 live), Lavance (7K+ stations). Score Studio (self-serve, pay in USD or SN44) launches 16 Sep 2026.
- **Why it matters for the talk**: A Big-Four firm buying a model trained by anonymous, competing miners rather than a lab.

Numbers: SubConnect ARR $1–2M, "high confidence", buybacks "planned" (26 Aug 2026). Team guidance "~$40M ARR in 2026 if pipeline converts" (SubConnect; treat as pipeline). Market cap 203,352 τ ≈ $48M, #5 (taostats, 12 Sep 2026). Emission share 5.3% (SubnetRadar). SN44 alpha listed on Kraken 30 Jul 2026. TaoWeave invested $1M plus a $1M perpetual lock.

Best recent posts (X: https://x.com/webuildscore):

1. 28 Aug 2026 (1,594 likes), https://x.com/webuildscore/status/2093313564693917714
   > The newest GTA 6 gameplay trailer dropped yesterday. we ran our subnet vision models on it.
   > Person, car, gun, segmented and tracked across every scene. night, water, motion blur, freefall over the city. footage the model had never seen.
   > The part that matters is that this isn't a huge model. it's tiny, and it runs near real time.
2. Manako, 15 Apr 2026, https://x.com/manakoai/status/2044445879554412571
   > Manako × @PwC_France We've formed an alliance to bring physical AI to enterprise at global scale. PwC France will lean on Manako's Business Operations World Model, powered by @webuildscore to turn their clients' existing camera networks into real-time systems of action.
3. 7 Aug 2026, https://x.com/webuildscore/status/2085789746270200099
   > Roboflow charges you for models you have to verify yourself.
   > Introducing Score Studio: the decentralised alternative where every model arrives already proven and ready to work in real world conditions. [...] Pay in USD or SN44, straight at subnet level.
4. 30 Jul 2026, https://x.com/webuildscore/status/2082834900642664628
   > $SN44 is now live on @krakenfx.
   > Nice milestone.
   > Back to training models.

Links: https://scorevision.io · X @webuildscore, @manakoai

Watch out: the "$40M ARR" is a pipeline number; one secondary source says "nearly $3M ARR" (unverified). Buybacks are planned, not running.

---

### Extra 3. Bitcast (SN93)

Added because it is the only subnet whose revenue roughly equals its miner emissions, on a per-transaction public ledger.

- **Netuid / name**: 93, "Bitcast".
- **What it produces**: Brands pay Bitcast for creator marketing campaigns; miners are YouTube/social creators paid for verified views and watch time.
- **Customers**: "multiple enterprise brands in parallel" (team, Feb 2026); names not disclosed.

Numbers: on-chain dashboard bitcast.network/revenue (2 Sep 2026): $389.9K cumulative revenue, $261.8K buybacks (67%). Q2 2026 revenue $105,499 vs $101,841 miner emissions, 3.5% margin (SubConnect, "high confidence"). Aug 2026 ~$65K, 30% MoM growth for six months (TAO Daily, 3 Sep 2026). Market cap 96,164 τ ≈ $23M, #13 (taostats). Emission share 3.1% (SubnetRadar).

Posts: 6 Jan 2026, https://x.com/Bitcast_network/status/2008609007917810123 — "today, we burnt almost $630k of SN93 alpha! 18% of issued supply has now been permanently removed from circulation." (that burn was emission-funded owner alpha, not revenue). 25 Feb 2026, https://x.com/Bitcast_network/status/2026708064200794380 — "1.4M subscribers (+55%) 47K hours watched (+32%) 765K views (+42%)... service multiple enterprise brands in parallel".

Why it matters: the proof-of-concept for "emissions as a bootstrap, not a subsidy". Small dollars, but the ledger is public.

---

### Extra 4. Hippius (SN75)

Added for attention (first subnet token on a centralized exchange) and for filling the storage slot of the permissionless stack.

- **Netuid / name**: 75, "Hippius". Team: The Nerve Lab (@Hippius_cloud).
- **What it produces**: S3-compatible decentralized object storage plus confidential compute (AMD SEV-SNP) and a Hugging Face-style model hub, every transaction on-chain.
- **Customers**: About a dozen other subnets (Score, Affine, Minos) plus external storage users. Storage revenue ~$50K (Oct 2025, disclosed); SubConnect $200K/yr, "medium confidence", revenue-share announced.

Numbers: 100 TB egress in a month, "AWS S3 would've billed ~$8,000... Hippius billed $0" (8 May 2026). 536K alpha burned (mix of emission- and revenue-funded). First alpha listed on a CEX (MEXC, 2025; Grayscale Research); Kraken listing 14 Jul 2026. Market cap 82,010 τ ≈ $19M, #17 (taostats). Emission share 2.2% (SubnetRadar).

Posts: 8 May 2026, https://x.com/Hippius_cloud/status/2052618760557674861 — "🚀 Hippius milestone this month: 100 TB of egress served across the network. AWS S3 would've billed ~$8,000 in bandwidth. Hippius billed $0. Egress is free. Always." Kraken, 14 Jul 2026, https://x.com/krakenlistings/status/2077033008113287258.

Watch out: demand is mostly other subnets (circular).

---

### Extra 5. Ridges (SN62)

Added because it is the most direct benchmark story against the labs (coding agents) and composes three subnets into one product.

- **Netuid / name**: 62, "Ridges". Team: Ridges AI, merged with Latent Holdings / Joseph Jacks (OSS Capital) 27 Jan 2026.
- **What it produces**: Miners compete to build autonomous software-engineering agents; the winner powers "Ridgeline", a GitHub-native agent that takes an issue and returns a pull request, running inference on Targon and Chutes.
- **Customers**: Open beta since Mar 2026; $9.99/month Standard plan (21 Jul 2026); free credits for alpha stakers; x402 agent payments live (28 Jul 2026). No revenue disclosed.

Numbers: market cap 59,074 τ ≈ $14M, #19 (taostats). Emission share 1.1% (SubnetRadar). Competition 22: 419 agents, 9 approved, best 76.67% (22 May 2026). Average cost per solved problem $0.10 vs a $0.29 cap (22 May 2026).

Posts:
1. 28 Jul 2026, https://x.com/ridges_ai/status/2082103797418262564 — "⛰️ x402 is now live on Ridges. X402 is an agentic payment protocol that lets agents pay for API access autonomously. Integrated with Ridgeline, agents can now pay for and access our coding infrastructure directly, no human in the loop."
2. 21 Jul 2026, https://x.com/ridges_ai/status/2079567082874871969 — "⛰️ Get convicted with us. Simple credits. Free if you stake. Standard · $9.99/month. 10 credits, always the top-ranked agent, PRs opened automatically. Staker · free. Lock your Ridges alpha, connect your wallet, get credits every 3 days."
3. 22 May 2026, https://x.com/ridges_ai/status/2057823982578225231 — "⛰️ Ridges Competition 22 Results · May 14-19. 419 agents created. 9 approved for emissions. Best score: 76.67%. Avg: 57.56%. The bar is high. That's the point."

Watch out: no revenue; Jacob has publicly noted Ridges fell under TaoFlow (20 Jan 2026). Say "beta" on stage.

Also notable but left out: Vanta SN8 (Taoshi, #7 by mcap; trading evals paid in USD; $120K single-trader payout 21 Aug 2026; revenue not disclosed), Actual SN95 (#11, 0% emissions, no public product), NOVA SN68, Minos SN107, Gradients SN56, Blockmachine SN19 ($1.3M/yr RPC, mainly sold to Taostats, a related party; note SN19 is no longer Nineteen).

---

## Part 2. Revenue: what is defensible

### 2.1 The Dec 2025 statement, in context

Jacob, 4 Dec 2025 17:27 UTC, https://x.com/const_reborn/status/1996632415633043965 (verified via the X API and fxtwitter):

> "generate revenue too quickly" -- I don't think any subet on Bittensor is generating flow through revenue in any serious way. It is all speculative investment. So the idea that flow emission has pushed people towards making revenue is some sort of sick meme which persists from ignorance.

Context: a reply to @badenglishtea in a debate about the TaoFlow emission model. The critic had written "there is very little revenue in Bittensor and this is something the network must accept... Most of the existing revenue is just miners rebuying supply via $TAO" (https://x.com/badenglishtea/status/1996619629129761207). Jacob's point was that TaoFlow was not forcing anyone to chase revenue, because nobody had serious revenue.

Two other Jacob statements to keep in view, because a critic will:
- 14 Aug 2025, https://x.com/const_reborn/status/1956021150443606291: "Top three subnets on Bittensor produce revenue. Two of them are emission positive and the third serves more token on open router than any competing provider."
- 23 Mar 2025, https://x.com/const_reborn/status/1903813891693433318: "Currently millions of organic web 2 revenue per month. Look it up!"

So Jacob has said both "millions per month" (Mar 2025) and "nothing serious" (Dec 2025). The honest line is that the Dec 2025 statement was the stricter and more accurate one at the time, and that the picture has moved since.

### 2.2 The SubConnect Bittensor Revenue Index, Vol. I (the source of "$28–35M")

What it is: a 45-page PDF released 26 Aug 2026 by SubConnect (@WeAreSubConnect), a Bittensor consultancy and subnet incubator (it incubated Endure, SN30). It is the only per-subnet revenue index that exists. It is not independent research and it is not an audit. It compiles team disclosures, public dashboards and on-chain buybacks, and grades each figure "high" (dashboard / on-chain / company disclosure), "medium" (public statements, extrapolated metrics) or "low". Report: https://share.sendnow.live/s/subconnect-revenue-index-1 ; announcement: https://x.com/WeAreSubConnect/status/2092630056657113201 ; tao.media write-up 1 Sep 2026.

Internal inconsistency to note: the Key Findings page says "$32M–$35M" across 25 subnets; the Executive Summary says "$28–35 million" across 24 subnets. Quote "$28–35M" (the wider range) if you quote it at all.

Full per-subnet table (recovered from the report's TL;DR pages, 4 Sep 2026, https://x.com/WeAreSubConnect/status/2095786108806463520):

| Category | Subnet | Est. annual revenue | Confidence | Revenue-funded alpha buys | Margin |
|---|---|---|---|---|---|
| Compute & Infra | SN4 Targon | $5.5–6.0M | High | Yes | Low |
| Compute & Infra | SN19 Blockmachine | $1.3M | High | Yes | Low |
| Compute & Infra | SN28 GM | $2.5–3M | Medium | Not disclosed | Low |
| Compute & Infra | SN51 Lium | $8.0–10M | High | Yes | Low |
| Compute & Infra | SN53 Engy | $400–600K | High | Yes (100% of revenue) | Low |
| Compute & Infra | SN56 Gradients | Undisclosed | Medium | Yes (100%) | Low |
| Compute & Infra | SN64 Chutes | $4.0–5.0M | High | Yes | Low |
| Compute & Infra | SN75 Hippius | $200K | Medium | Revenue-share announced | Low |
| Compute & Infra | SN110 Green Compute | $1.2M | High | Yes (10%) | Low |
| Enterprise AI | SN21 AdTAO | $500–700K | High | Yes (17% of profits) | Medium |
| Enterprise AI | SN33 ReadyAI | $50–100K | High | Yes (75% of pipeline rev.) | High |
| Enterprise AI | SN44 Score | $1–2M | High | Planned | High |
| Enterprise AI | SN54 Yanez | $300–400K | High | Yes (20%+ of receivables) | High |
| Enterprise AI | SN55 NIOME | $500K | Medium | Not disclosed | — |
| Enterprise AI | SN71 Leadpoet | $30K | High | Planned | High |
| Applied AI | SN8 Vanta | $750K–1M | Medium | Yes | Medium |
| Applied AI | SN22 Desearch | $65–75K | High | Planned (50% of surplus) | High |
| Applied AI | SN34 BitMind | Undisclosed | Medium | Yes | — |
| Applied AI | SN41 Almanac | $275–325K | High | Yes (100% of fees) | Low |
| Applied AI | SN50 Synth | $1–2M | High | Yes ($100K buy) | High |
| Applied AI | SN60 Bitsec | Undisclosed | Medium | Planned | High |
| Applied AI | SN61 RedTeam | Undisclosed | Medium | Planned (~2 TAO/day) | High |
| Applied AI | SN85 Vidaio | Undisclosed | Medium | Not disclosed | High |
| Applied AI | SN93 Bitcast | $500–550K | High | Yes (64.5% to date) | Low (Q2: 3.5%) |
| Applied AI | SN118 Ditto | $43K | Medium | Yes (100%) | Low |

Sum of the disclosed lines: $28.1M–$35.0M. That is where the headline comes from. Three subnets (Lium, Targon, Chutes) are $17.5–21M of it, i.e. more than half.

Next-wave list (no revenue estimate assigned): Oro SN15, Zeus SN18, Endure SN30, SN46, Engy SN53 (also above), Babelbit SN59, Actual Computer SN95, Beam SN105, Minos SN107, sundae_bar SN121.

SubConnect's own emissions comparison (31 Aug 2026, https://x.com/WeAreSubConnect/status/2094440794501292191): "Currently, revenue generated across the $TAO ecosystem offsets 11.7% of annual emissions."

### 2.3 Other estimates

| Source | Date | Total | Method | Verdict |
|---|---|---|---|---|
| ownyourmind.ai, "where the revenue is" | updated 18 Aug 2026 | measured $1.1–5.6M/yr; "generous ceiling" of self-reports $16M | prices Chutes' OpenRouter-visible volume at list; treats team claims as unverified | independent, conservative |
| Pine Analytics bear case | 24 Mar 2026 | $3–15M/yr | OpenRouter volume; Chutes subsidy ratio 22–40× | independent, dated |
| Khala Research | 19 Feb 2026 | no total; Chutes $5.5M | team interviews | disclosed-based |
| taoprotocol.org / AInvest / blockeden "$43M quarterly Q1 2026" ($172M/yr) | Apr–May 2026 | $172M/yr | none given | REJECT. No primary source. Almost certainly counts emissions. Flagged as unverified by cryptotimes.io (4 May 2026). |
| taorevenue.com "30-day inflow + burn 34,347 τ ≈ $8M" | 12 Sep 2026 | — | "inflow" is net stake flow, not sales (some subnets show negative inflow) | do not call this revenue; only the burn column is revenue-linked |
| Grayscale, Blockworks filing, Messari, The Block, Yuma "State of Bittensor Vol. 2" | 2026 | none publish a subnet revenue total | — | no help either way |

### 2.4 What is verifiable today, subnet by subnet

Team-disclosed and checkable on-chain or on a public ledger (annualized from Aug/Sep 2026 monthly figures; run-rates, not trailing twelve months):

| Subnet | Basis | Run-rate |
|---|---|---|
| Lium SN51 | $964K billed in a month (9 Sep 2026); $1.0M + $700K revenue-funded burns with extrinsic IDs | ~$11.6M |
| Chutes SN64 | $1.13M in 90 days on the team's dashboard (30 Aug 2026); burns visible as burn_alpha extrinsics | ~$4.6M |
| Bitcast SN93 | $389.9K cumulative, per-transaction ledger; ~$65K in Aug 2026 | ~$0.8M |
| Engy SN53 | $14.3K settled to miners in the week of 2–9 Sep 2026 on a signed public ledger | ~$0.75M |
| Targon SN4 | only buyback amounts are public (>550 TAO ≈ $115K in a month) | revenue not disclosed in 2026; SubConnect estimate $5.5–6M |
| Subtotal, checkable | | ~$17–18M (≈$23M if you accept the Targon estimate) |

### 2.5 The arithmetic against emissions

- Emission: ~3,600 TAO/day after the 15 Dec 2025 halving (3,559/day measured on-chain by ownyourmind).
- At $234/TAO: ~$843K/day, ~$308M/yr.
- Revenue as a share of annual emission value:

| Revenue figure | Source | Share of emissions |
|---|---|---|
| $1–6M measured | ownyourmind, Aug 2026 | 0.4–1.8% |
| ~$17–18M checkable run-rate | this brief | ~5.5–6% |
| ~$23M incl. Targon estimate | this brief | ~7.5% |
| $28–35M | SubConnect, Aug 2026 | 9–11% |

Per-subnet: even Lium's ~$32K/day of billings is well below the TAO its subnet receives per day (hundreds of TAO ≈ $100K+). Chutes' own words: revenue covers "thirty to forty percent of operating expenses". Bitcast is the only subnet in this brief whose revenue matched its miner emissions in a quarter, at ~$100K.

### 2.6 What Jacob can say on stage

Safe (team-disclosed, checkable, attribute it):
- "Lium billed $964,000 in a month to 1,187 renters and bought back and burned a million dollars of its token with the proceeds. The transaction is on-chain."
- "Chutes serves about three million requests and 30 billion tokens a day, published a year of production traffic with Harvard, and does about $4–5 million a year by its own public dashboard."
- "Engy runs a 2.8-trillion-parameter open model on 80 gaming GPUs, sells it at half the maker's price, and publishes a signed ledger of what it paid its miners every week."
- "Fourteen subnet teams now buy back their own token with customer revenue." (SubConnect count; the top six are individually checkable.)
- "Named customers include PwC France, an NYSE-listed REIT, Dropbox as a pilot." (Score/Manako tweet 15 Apr 2026; ReadyAI statement 21 May 2026; SubConnect for Dropbox. Say "pilot" for Dropbox.)
- "External revenue is somewhere between $15 and $35 million a year depending on who counts. Emissions are about $300 million a year. Revenue offsets roughly ten percent of what the protocol pays out."

Say only with attribution ("a Bittensor consultancy estimates..."):
- "$28–35M across 24 subnets" (SubConnect).
- Targon "$5.5–6M" (SubConnect) or "$10.4M" (older team claim; do not use).
- Score "$1–2M", Blockmachine "$1.3M", Synth "$1–2M", Green Compute "$1.2M".
- ">$100M by end of 2026" (SubConnect projection based on pipelines).

Do not say:
- "$43M in quarterly revenue" or "$172M run-rate". No source; conflates emissions.
- "Revenue exceeds emissions" for any major subnet. Only Bitcast (tiny) and Epago (tinier) have come close.
- "Chutes serves 120–160 billion tokens a day." That was mid-2025; today it is ~30B.
- "Teutonic-I is the world's most performant decentralized LLM" without the audit caveat.
- That Affine, Teutonic, IOTA, Oro, Albedo or Conjectures have customers.
- That SubConnect is independent research, or that taorevenue.com's inflow column is revenue.
- That the Dec 2025 tweet was wrong. It was right then. The better story is what changed in nine months: Lium's billing grew from ~$430K/month (May) to $964K/month (Aug), Chutes moved from free-tier volume to paid volume, Engy launched with a public ledger, Targon and Chutes started monthly revenue buybacks, and 14 teams now recycle revenue into their tokens.

Suggested stage line (plain, defensible): "Nine months ago I said no subnet had serious revenue. I was right. Today three or four do, about fifteen to twenty million dollars a year that you can check on-chain, and fourteen teams buy back their token with it. That is still one-tenth of what the protocol pays out. Closing that gap is the job of the next eighteen months."

---

## Slide-ready facts (one line each, sourced, dated)

1. Lium (SN51) billed $964K to 1,187 renters in one month, +36% MoM; 63% of rentals were started by software agents. (@lium_io, 9 Sep 2026)
2. Lium bought back and burned $1,000,000 of SN51 alpha from product revenue in a single on-chain transaction. (@fish_datura, extrinsic 9032919-0005, 9 Sep 2026)
3. Lium has GPU capacity from 68 datacenters in 21 countries; 1,393 GPUs listed, 85% rented. (@lium_io 9 Sep 2026; lium.io API 12 Sep 2026)
4. Chutes (SN64) served 6.12 billion requests, 35.8 trillion tokens and 314,970 users in one year, published as a research paper with Harvard and UChicago. (arXiv 2608.13573, trace Apr 2025–Apr 2026)
5. Chutes handles ~2.8M requests and ~31B tokens a day, every model inside a hardware-attested TEE. (api.chutes.ai public stats, 7-day average to 11 Sep 2026)
6. Chutes revenue: $1.13M in 90 days (~$4.6M/yr) on its public dashboard; 260,770 alpha (~$4M) burned from revenue since inception. (SubnetRadar 30 Aug 2026; SubConnect Aug 2026)
7. Targon (SN4) repurchased >550 TAO (>$115K) of alpha from organic revenue in one month; Intel co-authored its confidential-compute whitepaper. (@TargonCompute 20 Aug 2026; Intel blog 23 Mar 2026)
8. Engy (SN53) runs Kimi K3, a 2.8-trillion-parameter open model, on 80 RTX 5090 gaming cards with zero HBM, at half Moonshot's price. (@totheagi 27 Jul 2026; engy.ai pricing)
9. Engy's public ledger: 77 miners, 3.57M requests, 56B tokens and $14,318 settled in the week of 2–9 Sep 2026. (provider.engy.ai epoch 12, fetched 12 Sep 2026)
10. gm (SN28) listed GPT-6 Astra at an 8.3% discount on 4 Sep 2026, a week before OpenRouter; 96.5B tokens processed to date. (@say_gm_ 4 and 7 Sep 2026; OpenRouter model page dated 11 Sep 2026)
11. Affine (SN120) miners pushed a 35B open model's terminal-agent score from 46.2 to 78.1 and SWE score from 33.1 to 58.2 within days of a base switch; each of five winning slots pays ~$50K/week. (@affine_io 24 Jun and 23 Aug 2026)
12. Teutonic-II (SN3) is a 110B-parameter permissionless pre-training run, "50% larger than any prior decentralized run"; the 10B Teutonic-I run had 2,163 duels and 203 coronations across 73 participants in 70 days. (@const_reborn 24 Aug 2026; teutonic.ai paper) [Add the Templar audit caveat if you cite Teutonic-I's benchmark.]
13. IOTA (SN9) trained a 100B-parameter model across 48 A100s in 5 datacenters at ~65% of colocated efficiency, and a 16B model on gaming GPUs from the USA to Hong Kong. (Macrocosmos 1 Jun 2026; @mccrinbc 27 Jul 2026)
14. Jacob launched a 128-node permissionless training run on IOTA in 5 minutes for 5 TAO. (@const_reborn 3 Jun 2026)
15. Oro (SN15) paid $200,000+ to 1,206 miners in its first 36 days for shopping agents; it joined Y Combinator Fall 2026, the first live-token crypto startup to do so. (@oroagents 20 May and 8 Sep 2026)
16. Conjectures (SN66) miners solved the near-half-density case of Green's Problem 51, open for 16 years, verified in Lean; the subnet is a Gold Sponsor of the Caltech Mathathon alongside Anthropic and Cognition. (@conjectures_io 9 and 2 Sep 2026)
17. Score (SN44): PwC France uses Score's vision models via Manako; SN44 alpha listed on Kraken 30 Jul 2026. (@manakoai 15 Apr 2026; @webuildscore 30 Jul 2026)
18. Bitcast (SN93): Q2 2026 revenue $105,499 vs $101,841 miner emissions, the first subnet to out-earn its miner emissions in a quarter, on a public per-transaction ledger. (SubConnect Aug 2026; bitcast.network/revenue)
19. 24–25 subnets bill outside customers for an estimated $28–35M a year; 14 fund token buybacks from it; revenue offsets ~11.7% of annual emissions. (SubConnect Revenue Index Vol. I, 26–31 Aug 2026, a consultancy compilation)
20. Team-disclosed, on-chain-checkable revenue run-rate: ~$17–18M/yr (Lium, Chutes, Bitcast, Engy) against ~$308M/yr of emissions at $234/TAO. (this brief, 12 Sep 2026)
21. Top three subnets by market cap on 12 Sep 2026: Lium $106M, Chutes $96M, Targon $77M; all 128 subnets together ~$1.1–1.3B. (taostats snapshot block 9,053,036; SubnetRadar)
22. Jacob, 4 Dec 2025: "I don't think any subet on Bittensor is generating flow through revenue in any serious way. It is all speculative investment." Nine months later, four subnets have checkable revenue. (x.com/const_reborn/status/1996632415633043965)

---

## Conflicts and unverified items

1. **Market cap definitions differ by tracker.** taostats (used here) = price × (alpha in pool + alpha staked). SubnetRadar uses total alpha supply (higher: Lium $135M). bittensor.ai uses staked alpha only (lower: Lium $68M). Use ranks on stage, or say "about a hundred million".
2. **Emission share differs by tracker and swings daily.** Lium: 24.1% (bittensor.ai 7d) vs 13.7% (SubnetRadar). IOTA: 9.4% vs 2.8%. Oro: 1.8% vs 0%. Under the current price-based model with an emission gate, instantaneous shares move fast. Use ranks, not percentages.
3. **SubConnect total**: $32–35M (Key Findings, 25 subnets) vs $28–35M (Executive Summary, 24 subnets) in the same PDF.
4. **Chutes revenue**: $1.3–2.4M (independent, OpenRouter-priced) vs $4.6M (own dashboard). The gap is direct sales OpenRouter cannot see. Use the dashboard figure with attribution.
5. **Targon revenue**: $10.4M (team, 2025, unaudited) vs $5.5–6M (SubConnect, 2026). Team GPU count 1,500+ H200s vs a few dozen to ~230 visible on-chain.
6. **Engy**: DeAI Dashboard "$1.73M/24h" vs the team ledger ~$2K/day. Ledger wins.
7. **Teutonic-I**: "most performant decentralized LLM" vs Templar audit (Qwen3.5-9B-Base scores the same; weights 96% correlated).
8. **Oro "beats GPT-5.4"**: self-reported only.
9. **Lium**: $432K/month (May, single-sourced) vs $964K/month (Aug/Sep). Consistent with 36% MoM growth, but both are self-reported, and some demand is other subnets (circular).
10. **"$43M Q1 2026 revenue"**: no source. Reject.
11. **Netuid reuse**: SN19 is Blockmachine (not Nineteen); SN8 is Vanta (was PTN); SN3 is Teutonic (was Templar); SN39 deprecated (was Basilica); SN81 is Reliquary (was Grail). Old articles name the wrong project.
12. **No official Teutonic X account**; Jacob's account is the channel. **Albedo** has three accounts (@DendriteHQ parent, @weather_boss_ lead dev, @arbos_born origin).
13. **Score "$40M ARR in 2026"** is pipeline guidance, and "nearly $3M ARR" is a single secondary source.

---

## Sources

Primary: on-chain, ledgers, team sites
- taostats.io/subnets (embedded data, block 9,053,036, 12 Sep 2026 17:36 UTC): https://taostats.io/subnets
- CoinGecko TAO price: https://api.coingecko.com/api/v3/simple/price?ids=bittensor&vs_currencies=usd (12 Sep 2026)
- Lium live stats: https://lium.io/api/executors/stats ; site https://lium.io ; docs https://docs.lium.io ; GitHub https://github.com/Datura-ai/lium-io ; extrinsics https://taomarketcap.com/extrinsics/9032919-0005 , 7898274-0011, 7898314-0023
- Chutes live stats: https://api.chutes.ai/invocations/stats/llm ; https://api.chutes.ai/miner/stats ; site https://chutes.ai ; GitHub https://github.com/chutesai/chutes ; paper https://arxiv.org/abs/2608.13573
- Targon: https://targon.com ; https://docs.targon.com ; https://stats.targon.com/api/miners ; GitHub https://github.com/manifold-inc/targon ; Intel whitepaper https://community.intel.com/t5/Blogs/Products-and-Solutions/Security/Decentralized-Compute-on-Untrusted-Hardware-Using-Intel-TDX-and/post/1741192 ; arXiv https://arxiv.org/abs/2607.21865 ; Series A https://oss.capital/oss-capital-leads-10-5m-series-a-in-manifold-labs-alongside-industry-legends/
- Engy: https://engy.ai ; https://engy.ai/pricing ; https://engy.ai/status ; ledger https://provider.engy.ai/api/subnet/v1/epoch/latest ; https://api.engy.ai/v1/network ; GitHub https://github.com/hanlinai/engy ; https://hanlin.ai
- gm: https://saygm.com ; https://saygm.com/models/claude-fable-5 ; https://saygm.com/miners ; https://docs.saygm.com ; GitHub https://github.com/taostat/gm-miner
- Affine: https://affine.io ; https://github.com/AffineFoundation/affine-cortex
- Teutonic: https://teutonic.ai ; https://teutonic.ai/paper.html ; https://github.com/unarbos/teutonic ; https://huggingface.co/dendriteholdings/Teutonic-I ; https://huggingface.co/dendriteholdings/teutonic-II-110B-genesis ; Templar audit https://www.tplr.ai/teutonic-1-audit.pdf
- IOTA: https://iota.macrocosmos.ai ; https://docs.macrocosmos.ai/subnets/subnet-9-iota ; https://github.com/macrocosm-os/iota ; https://arxiv.org/abs/2507.17766 ; https://iota.macrocosmos.ai/research/iota_primer.pdf ; https://macrocosmosai.substack.com ; Companies House via https://open.endole.co.uk/insight/company/15583737-macrocosmos-ltd
- Oro: https://oroagents.com ; https://docs.oroagents.com ; https://app.oroagents.com/whitepaper ; https://github.com/ORO-AI/oro ; ShoppingBench https://arxiv.org/abs/2508.04266
- Albedo: https://github.com/unarbos/albedo ; https://dendrite.holdings ; https://huggingface.co/dendriteholdings
- Conjectures: https://conjectures.io ; https://conjectures.io/faq ; https://github.com/conjectures-io/conjectures-validator ; https://github.com/google-deepmind/formal-conjectures
- Score: https://scorevision.io · Bitcast: https://bitcast.network/revenue · Hippius: https://hippius.com · Ridges: https://ridges.ai
- SubConnect Revenue Index Vol. I: https://share.sendnow.live/s/subconnect-revenue-index-1 (45 pages; page images read 12 Sep 2026); TL;DR tables https://x.com/WeAreSubConnect/status/2095786108806463520 ; announcement https://x.com/WeAreSubConnect/status/2092630056657113201 ; emissions offset line https://x.com/WeAreSubConnect/status/2094440794501292191
- Jacob's tweets: archive at docs/sources/tweets/const_reborn-tweets.md ; Dec 2025 tweet https://x.com/const_reborn/status/1996632415633043965 ; context https://x.com/badenglishtea/status/1996619629129761207

Tweets read (all via X API v2 search or https://api.fxtwitter.com/<handle>/status/<id>): every URL quoted in the profiles above. Profiles: @lium_io, @fish_datura, @chutes_ai, @TargonCompute, @carromademethis, @engyai, @totheagi, @say_gm_, @affine_io, @const_reborn, @DendriteHQ, @arbos_born, @tplr_ai, @IOTA_SN9, @MacrocosmosAI, @mccrinbc, @oroagents, @shardiban, @conjectures_io, @webuildscore, @manakoai, @Bitcast_network, @Hippius_cloud, @ridges_ai, @WeAreSubConnect.

Secondary: trackers and analysis
- bittensor.ai subnet pages: https://bittensor.ai/subnets/<netuid> (12 Sep 2026)
- SubnetRadar: https://subnetradar.com ; deep dives https://subnetradar.com/blog/engy-sn53-verified-inference-deep-dive (25 Jul 2026), https://subnetradar.com/blog/gm-sn28-private-ai-gateway-deep-dive (27 Jul 2026); Discord digests https://subnetradar.com/subnet-news/<netuid>/<date>
- subnetalpha.ai; taopedia.org; dashboard.deaination.com; taorevenue.com (inflow column is stake flow, not revenue)
- tao.media: https://www.tao.media/subconnect-announces-first-bittensor-revenue-index/ (1 Sep 2026); https://www.tao.media/albedo-sn97-the-bittensor-subnet-that-gets-stronger-every-time-someone-tries-to-game-it/ (13 Jul 2026); https://www.tao.media/oro-subnet-hacked-north-korean-hackers-drain-630k-in-tao-from-sn15-owner-wallet/ (20 Jul 2026); Lium, Targon, Chutes, Teutonic pieces (May–Sep 2026)
- TAO Daily: https://taodaily.io/how-gm-sn28-sells-fable-and-astra-for-less-than-the-companies-that-make-them/ ; https://taodaily.io/engy-responds-to-centralization-criticism-with-buyback-and-burn-and-permissionless-tee-mining/ (11 Aug 2026); https://taodaily.io/lium-flips-chutes-as-bittensors-top-subnet-while-scaling-high-speed-gpu-rentals/ (10 Sep 2026); https://taodaily.io/oro-becomes-first-live-token-crypto-startup-accepted-into-y-combinator/ (9 Sep 2026)
- ownyourmind.ai: https://ownyourmind.ai/tokenomics/bittensor-subnets-where-the-revenue-is/ (18 Aug 2026); https://ownyourmind.ai/tokenomics/chutes-bittensor-revenue-machine/ ; https://ownyourmind.ai/tokenomics/lium-bittensor-subsidy-ratio/ ; https://ownyourmind.ai/tokenomics/targon-bittensor-confidential-compute/
- Pine Analytics bear case: https://pineanalytics.substack.com/p/the-bear-case-for-bittensor-tao (24 Mar 2026)
- Khala Research: https://www.khala.io/bittensor-the-intelligence-olympics (Feb 2026)
- synapz.org on Teutonic-I provenance: https://www.synapz.org/posts/2026-08-21-crowned-on-the-second-day (21 Aug 2026)
- The Block on Covenant exit: https://www.theblock.co/post/396959/covenant-ai-exits-bittensor-tao (10 Apr 2026)
- Protos on Oro hack: https://protos.com/north-korean-hackers-drain-630k-from-bittensor-subnet-oro/ (Jul 2026)
- KuCoin news on gm/OpenRouter: https://www.kucoin.com/news/flash/bittensor-integrates-confidential-routing-layer-with-openrouter-processes-up-to-120-billion-tokens-daily
- Kraken listings: https://blog.kraken.com/product/asset-listings/sn51-is-available-for-trading (9 Jul 2026); https://x.com/krakenlistings/status/2077033008113287258 (SN75, 14 Jul 2026); https://x.com/krakenlistings/status/2082832313680073108 (SN44, 30 Jul 2026)
- Rejected: taoprotocol.org / AInvest / blockeden "$43M Q1 2026" (no primary source).

Working notes (not in the store): four research memos and raw tweet pulls under /tmp/subnet-research on the research VM.