{ "title": "The IOTA Field Map", "description": "Who else trains AI across compute they do not own or reserve, placed by relation to IOTA and compared by mechanism, largest run, and what is sold. Every cell sourced or marked unknown.", "subject": "IOTA, Bittensor subnet 9: who else is running, and how IOTA differs by mechanism", "compiled_by": "Mikyö Clark, from public sources. Not affiliated with Macrocosmos or any entity on this page.", "updated": "2026-09-29", "corrections": "connect@mikyo.one", "labels": { "verified": "checked against the code, a paper, the chain, or a live page, on the date given", "stated": "the project said it, in the source given", "secondhand": "reported by someone else; not said by the project in writing", "inferred": "reasoned from stated facts; the reasoning is given", "unknown": "the public record does not fill this cell" }, "frame": { "word_count": "under 150 words", "choice": [ "A model builder chooses between three things. A reserved cluster: matched hardware, held for months, the highest price. Spot: whatever is free, taken back without notice, the lowest price. Or a swarm: heterogeneous compute the builder does not own or reserve, joined and left mid-run, with the run surviving the machines. Every project on this page sells or studies the third.", "IOTA's mechanism, in one sentence: a model is cut into pipeline slices, each slice runs on a GPU Macrocosmos does not own, validators score each miner's work, and the chain pays miners in alpha by score share over a rolling day; the subnet has been on mainnet since June 2025 (relaunched August 2025). The difference from the field is the unit of payment (scored work on one slice, not a whole checkpoint or a trusted node) and the shape of the parallelism (pipeline across single cards, not data-parallel copies of a whole model)." ], "relations": { "rival-buyer": "Sells today, or on a stated date, a runtime, a managed run or capacity that a model builder without a reserved cluster would buy instead of IOTA's SDK or managed route. Pretraining, fine-tuning and RL all count; IOTA pitches all three. Raw GPU hours count only weakly.", "rival-supply": "Recruits the same class of card IOTA runs on (Apple Silicon Macs, 24 to 32 GB consumer GPUs, single A100, L40S or A6000 cards), and the owner must choose: the card earns from one or the other. Sharing TAO emission is not supply rivalry.", "complement": "Has a concrete path to buy from IOTA, sell compute to it, or distribute it. The card names the path. None of these paths is a public arrangement unless the card says so.", "reference": "Lineage, a benchmark or a competing claim: IOTA inherits from it, is measured against it, or is compared with it. It has no offer a builder could buy instead of IOTA and draws on no shared supply today." } }, "iota": { "paragraphs": [ "**Mechanism.** Pipeline-parallel across single GPUs, 16 stages and 3 replicas at 100B. An orchestrator assigns slices; validators score each miner's forward-pass work, and miners are paid in alpha by score share over a rolling window of about a day, with a cap and the excess burned. Control is central, the data path is peer to peer (Iroh since v3.0.0, 9 March 2026). The open way to join today is Train at Home on an Apple Silicon Mac with 16 GB; the GPU miner repository still lists 16 GB of VRAM. Runs survive nodes leaving; the second 16B run is the repeatability claim.", "**Largest run.** Orion-100B: 100B parameters, 1.1B tokens, 48 non-colocated A100 80GB across five US datacenters, \"provisioned from multiple providers\" (unnamed), about two days, 30.8% average MFU, 65% of co-located speed, $20 a replica-hour; a viability run, stopped for cost. Orion-16B: 256 GPUs at launch (27 July), c.180+ (5 August), 225 across three continents (20 August), then a mixed roster of 4090s, 5090s, A100s, L40S and A6000s, finished 23 September; GPUs from \"independently operated suppliers\", one named (Green Compute); 100B tokens by 14 August, about 20% MFU. How Orion suppliers were paid is not published. No checkpoint offered for download.", "**Sold, and to whom.** The iota SDK and Liquid Compute, announced 28 September 2026: early access by registration, the SDK page still \"coming soon\". Will Squires says the SDK covers pretraining, RL and fine-tuning; published results are on the legacy code base. Today the door is a managed route (\"You own the outcome. We run the training\"), by contact form, no pricing. The site also names GPU operators with idle capacity as SDK users. No named paying customer as of this page. Funded by emission and OTC alpha sales; Stillcore Capital invested in SN9 in May 2026 (secondhand, terms unknown). 164 companies in conversation was said from the Exploit stage and is secondhand.", "Every number above is on [the register](/iota-runs/) with its label and source." ], "voice": { "carries_label": "research-led in long form, company-led on X", "read": "2026-09-28", "carries": "Long form is Steffen's; on X the company account carries the numbers and the founders quote it. Counts as of 28 September are below.", "message": "Steffen: \"anyone can train models using globally distributed, heterogeneous and unreliable compute\" (28 September 2026); and on 30 May 2025: \"heterogeneous, unreliable, permissionless and token incentivized machines\". Will: \"Decentralised training can compete.\" (January 2025); the 2026 form: \"finally be leaving the research realm, and launch a product.\"", "accounts": [ { "channel": "@MacrocosmosAI", "url": "https://x.com/MacrocosmosAI", "note": "7,649; \"Building distributed intelligence on Bittensor. SN1, 9, 13\"; a weekday post at about 18:00 UTC; quotes founders by handle, never reposts them" }, { "channel": "@IOTA_SN9", "url": "https://x.com/IOTA_SN9", "note": "2,785; eight buyer-facing posts 10 to 25 September" }, { "channel": "@WSquires", "url": "https://x.com/WSquires", "note": "2,755; \"Co-Founder @MacrocosmosAI\"; four posts in thirty days" }, { "channel": "@macrocrux", "url": "https://x.com/macrocrux", "note": "2,712; \"CTO & Co-Founder @ Macrocosmos\"; four posts in thirty days; pinned: the 100B-token post, 13k views" }, { "channel": "Substack", "url": "https://macrocosmosai.substack.com", "note": "\"hundreds of subscribers\"; last post 5 August" }, { "channel": "YouTube", "url": "https://www.youtube.com/@MacrocosmosAI", "note": "5 subscribers, 0 videos; livestreams run on X" }, { "channel": "Discord", "url": "https://discord.gg/maYxtBzxAt", "note": "2,068 members" }, { "channel": "LinkedIn", "url": "https://www.linkedin.com/company/MacrocosmosAI", "note": "1,476; last post about July" }, { "channel": "GitHub", "url": "https://github.com/macrocosm-os", "note": "iota 34 stars, pushed 24 September; no SDK repository visible" } ], "founders": [ { "name": "Will Squires", "role": "CEO and co-founder", "handle": "WSquires", "bio": "\"Co-Founder @MacrocosmosAI\" (2,755). Proof of Talk lists him as CRO.", "statements": [ { "said": "\"We're very excited to finally be leaving the research realm, and launch a product.\" \"Pre training is too small of a market, so we've broadened the focus.\" \"We got told that the system must be: easy to use.\" \"We're ramping up slowly so we can serve customers super well.\"", "medium": "X", "venue": "quote of the Exploit Summit, 6h25m after the stage", "date": "2026-09-28", "url": "https://x.com/WSquires/status/2104681373533679638" }, { "said": "\"We have a huge list of improvements ready to go on the next run we launch.\"", "medium": "X", "venue": "quote of Steffen's 225-GPU post", "date": "2026-08-21", "url": "https://x.com/WSquires/status/2090737866317934930" }, { "said": "\"We always save the big bangs for Bittensor.\"", "medium": "X", "venue": "quote of the Exploit Summit announcement", "date": "2026-08-18", "url": "https://x.com/WSquires/status/2089757735202021743" }, { "said": "\"Liquid training changes the shape of the workload so it can occupy the gaps conventional training cannot use.\" \"The critical resource of the coming generation will be compute.\"", "medium": "Substack", "venue": "The Economics of Liquid Training; byline Will on the page, Macrocosmos in the archive", "date": "2026-07-14", "url": "https://macrocosmosai.substack.com/p/the-economics-of-liquid-training" }, { "said": "\"None of these teams have a live mechanism, or a live token.\" \"Our system is able to perform global cost arbitrage in order to train large models cheaply.\"", "medium": "interview", "venue": "taopill, Unsupervised Capital", "date": "2025-11-05", "url": "https://www.taopill.ai/p/an-interview-with-will-squires-from-macrocosmos" }, { "said": "\"There's a lot of industries that do not have data centers worth of compute.\" \"Cooperative training as a service.\"", "medium": "stage", "venue": "Novelty Search live from the Louvre, Proof of Talk (speaker inferred)", "date": "2025-06-11", "url": "https://www.youtube.com/watch?v=sCVtPIbIWHw" }, { "said": "\"Huge results. Decentralised training can compete.\"", "medium": "X", "venue": "", "date": "2025-01-21", "url": "https://x.com/WSquires/status/1881741832285216954" } ], "recurring_words": [ "onwards", "cracked", "cooking", "relentless", "constellation", "big bangs", "liquid compute", "customers", "pilot", "order of magnitude", "Advance." ] }, { "name": "Steffen Cruz", "role": "CTO and co-founder", "handle": "macrocrux", "bio": "\"CTO & Co-Founder @ Macrocosmos\" (2,712)", "statements": [ { "said": "\"We believe this fundamentally disrupts the economics of AI training.\" \"Anyone can train models using globally distributed, heterogeneous and unreliable compute with just a few lines changed from pure PyTorch.\"", "medium": "X", "venue": "the SDK, 3h45m after the stage", "date": "2026-09-28", "url": "https://x.com/macrocrux/status/2104641052565012805" }, { "said": "\"We were able to train it 3x cheaper than if we had simply reserved a node in a data center.\" The 1 June essay's own figure is 2.5x on replica entry cost.", "medium": "stage clip", "venue": "Exploit Summit, via The TAO Daily (secondhand)", "date": "2026-09-28", "url": "https://x.com/taodaily_io/status/2104601940294340972" }, { "said": "\"My current focus on Bittensor is making decentralized training both technically and commercially successful.\"", "medium": "written interview", "venue": "TAO.com, posted on X", "date": "2026-09-25", "url": "https://x.com/taodotcom/status/2103419292687421660" }, { "said": "\"If it can hold 10% of the model weights, we can train on it.\" \"When your system is built like @IOTA_SN9 there is no GPU scarcity.\"", "medium": "X", "venue": "", "date": "2026-08-20", "url": "https://x.com/macrocrux/status/2090544176274223467" }, { "said": "\"Utilisation is not the goal, useful work per dollar is, and the two are not always the same.\" \"Liquid training treats change in participation and available hardware and bandwidth not as a fault to be recovered from, but as the norm.\"", "medium": "Substack", "venue": "Deep Dive 2, with Eli Cohen", "date": "2026-08-05", "url": "https://macrocosmosai.substack.com/p/deep-dive-2-the-technical-requirements" }, { "said": "\"For the first time, an economically compelling case for training large models.\" \"An Orion replica (16 non-colocated A100s at $1.25 per hour) can be provisioned for $20 per hour.\"", "medium": "Substack", "venue": "the Orion-100B write-up, sole byline", "date": "2026-06-01", "url": "https://macrocosmosai.substack.com/p/orion-100b-distributed-pretraining" }, { "said": "\"Designed from day zero to work in an adversarial environment.\" Instead of ten data centers, \"10,000 MacBooks\" (close paraphrase).", "medium": "podcast", "venue": "Ventura Labs Ep. 50", "date": "2025-06-30", "url": "https://www.youtube.com/watch?v=zjRAyYRpImA" }, { "said": "\"A network of heterogeneous, unreliable, permissionless and token incentivized machines.\"", "medium": "X", "venue": "the primer launch", "date": "2025-05-30", "url": "https://x.com/macrocrux/status/1928491754736492662" } ], "recurring_words": [ "permissionless", "heterogeneous", "unreliable", "globally distributed", "incentivize", "orchestrated", "liquid", "frontier scale", "DPP", "commodity GPUs", "adversarial" ] } ], "mechanism_words": "Every word used in public for what IOTA does, with its latest sighting: decentralized (IOTA bio; Deep Dive 2, August), distributed (July), permissionless (August), heterogeneous and unreliable (today), liquid (June onward; \"The future is liquid\", 17 September), disaggregated (September site copy), interruptible (June essay; the stage today), scattered and mismatched (25 September), idle, stranded, underused, spare (June to today), \"compute nobody set aside for you\" (company, 22 September) and \"compute nobody reserved for us\" (IOTA, 23 September), orchestrated (the name), incentivised (the name), swarm (2025 only), trustless (2025 and the docs; in no 2026 founder or company post found)." } }, "assumptions": [ "An entity is in if it trains one model across machines that are not run as one cluster and exchange updates during the run, sells such a run or a runtime for one, or is a Bittensor subnet that pretrains a language model from scratch or sells compute or training. Pure inference marketplaces are out unless they have announced such training.", "Relation is judged against the buyer IOTA names (a model builder without a reserved cluster, for pretraining, fine-tuning or RL) and the cards IOTA runs on (Train at Home Macs, consumer GPUs, single datacenter cards), as of 29 September 2026. The tests are in the four relations above.", "Funding and headcount come from the 6 September 2026 landscape scan (`landscape/decentralized-ai.csv`) unless a card gives a newer source. They are aggregator or press figures and are labeled so." ], "rules": [ "An entity is added when it meets the inclusion rule and a primary source (its own docs, paper, repository, or announcement) can be opened for its mechanism. Every card carries the same eight fields. Every cell is filled or marked unknown. A company is never called a customer, partner, or pilot of IOTA unless Macrocosmos has said so in public.", "Mechanism and runs are taken from primary sources only. Funding and headcount may come from aggregators or press, labeled.", "The page is rendered from field.json by tools/render_field.py. Fork the data. Send corrections to connect@mikyo.one and they go in with a date." ], "entities": [ { "id": "covenant", "name": "Covenant AI", "kicker": "Templar, Basilica, GRAIL · left Bittensor April 2026", "url": "https://www.covenant.ai/", "relation": "rival-buyer", "secondary_relation": "reference", "last_verified": "2026-09-29", "people": [ { "name": "Sam Dare", "role": "founder", "handle": "DistStateAndMe", "said": "On leaving Bittensor, 9 April 2026: \"It is decentralization theatre.\"" }, { "name": "Joel Lidin", "role": "first author of the 72B paper" }, { "name": "Amir Sarfi", "role": "second author of the pipeline paper (first author Yazan Obeidi)" } ], "what_it_is": { "value": "\"Incentivized Internet-wide AI training\" (tplr.ai). Ran Bittensor subnets 3, 39 and 81 until 9 April 2026; SN3 is now Teutonic, SN39 is deprecated, SN81 is another team's. Now an independent lab with a GPU cloud, Basilica, paid in TAO.", "label": "verified" }, "mechanism": { "summary": { "value": "Data-parallel, full replica per peer (8 B200-class GPUs), SparseLoCo compresses updates 146x. Gauntlet scores each peer's update by loss before and after. Pipelining across peers tested at 1B; a 30B pipelined test ran 4 to 14 September 2026 and has finished.", "label": "verified" }, "detail": { "value": "**Data-parallel**: each peer holds a full replica sharded over its own 8 GPUs; SparseLoCo (30 local steps, top-k on 64x64 blocks, 2-bit) cuts communication more than 146x. **Verified** by Gauntlet: a validator scores every submitted pseudo-gradient by loss before and after, checks it beats random data, rates peers over time. Peers joined and left freely while it paid in alpha. A January 2026 paper adds **pipeline** parallelism with compressed activations (1B test). TEST_RUN_30B_001, \"with SparseLoCo and low-bandwidth pipelining\", ran 4 to 14 September 2026: 3 replicas of 4 stages on H100 PCIe 80 GB, about 18.4B tokens, about 37.5% training MFU (28% after overhead), about $0.82 per million tokens. The dashboard still reads \"Now Running\"; its API reports the run finished.", "label": "verified" } }, "largest_run": { "parameters_b": 72, "summary": { "value": "**Covenant-72B**, 1.1T tokens, 70+ permissionless peers at 8x B200 each, about 94.5% compute utilization, released 10 March 2026. Weights public.", "label": "verified" }, "detail": { "value": "**Covenant-72B**: 72B dense, about 1.1T tokens (1.09T DCLM plus 14.2B annealing), at least 70 unique peers each with 8 B200-class GPUs, an average of 16.9 contributing per step (cap 20), over commodity internet via Cloudflare R2. About 94.5% compute utilization; MMLU 67.1. Paper 9 March, weights 10 March 2026, on Hugging Face under Apache 2.0.", "label": "verified", "note": "MFU not published for the 72B; 94.5% is compute utilization. Crucible, an 8B run stated at 48 A100s, 48.3% MFU and $0.12 per million tokens, is on X and LinkedIn in September 2026 only; no paper, dashboard or weights found." } }, "sold": { "value": "Basilica: GPU rentals paid from TAO credits; a multi-node training deploy on rented A100 and H100 (since May 2026); a managed RL post-training API in SDK beta (since 27 August); RL rollout sessions with per-token cost. No public price list. Templar is dormant. No pretraining run can be bought.", "label": "verified" }, "paid": { "value": "None public.", "label": "unknown" }, "funding": { "value": "No round public. No token; ran three subnet alpha tokens until April 2026. 11 to 50 staff on LinkedIn.", "label": "unknown", "note": "Staff count from an aggregator, not re-read. A different company named Covenant (legal AI) raised in September 2026; unrelated." }, "relation_argument": "Rival-buyer: Basilica sells a model builder without a cluster a way to train on machines it does not own: GPU hours on A100 to H200 paid in TAO, a multi-node training deploy since May 2026, and a managed RL post-training API in beta since 27 August 2026. No pretraining run is for sale, and no outside GPU owner is paid to join a Covenant run today, so it does not draw on IOTA's miners. Reference: Covenant-72B (1.1T tokens, weights public) and the September 2026 30B pipeline test, which published MFU and cost per million tokens.", "sources": [ { "title": "Covenant-72B paper, arXiv 2603.08163, 10 March 2026", "url": "https://arxiv.org/abs/2603.08163" }, { "title": "Gauntlet, arXiv 2505.21684", "url": "https://arxiv.org/abs/2505.21684" }, { "title": "Heterogeneous low-bandwidth pretraining, arXiv 2601.02360", "url": "https://arxiv.org/abs/2601.02360" }, { "title": "Templar README and dashboard, read 29 September 2026", "url": "https://tplr.ai/dashboard" }, { "title": "Templar dashboard API, TEST_RUN_30B_001 (finished 14 September 2026)", "url": "https://tplr.ai/api/pipeline/overview" }, { "title": "Basilica SDK changelog (distributed deploy, RL API, rollout sessions)", "url": "https://github.com/one-covenant/basilica/blob/main/crates/basilica-sdk-python/CHANGELOG.md" }, { "title": "Weights on Hugging Face", "url": "https://huggingface.co/1Covenant/Covenant-72B" }, { "title": "The Block on the exit, 9 April 2026", "url": "https://www.theblock.co/amp/post/396959/covenant-ai-exits-bittensor-tao" } ], "signals": { "url": "https://www.covenant.ai/", "read": "2026-09-29", "js_rendered": true, "title": "Covenant", "h1": "ONE COVENANT, MANY ORDERS", "first_h2": "unknown", "meta_description": "One Covenant, Many Orders.", "og_description": "One Covenant, Many Orders.", "we_are_sentence": "unknown", "proof_numbers": [], "audience_named": [ "unknown" ], "cta_verb": "EXPLORE TEMPLAR", "contested_words_present": [ "decentralized", "permissionless", "distributed", "incentivized" ], "token_symbols": [], "dates_on_page": [ "unknown" ], "read_via": "browser pane", "product_page": { "url": "https://www.tplr.ai/", "title": "Templar - Incentivized Internet-wide AI training", "h1": "TEMPLAR INCENTIVIZED INTERNET-WIDE AI TRAINING", "meta_description": "Templar is a decentralized AI training network. Market-driven incentives for loss reduction, enabling permissionless distributed deep learning.", "cta_verb": "GET STARTED", "contested_words_present": [ "decentralized", "permissionless", "distributed", "incentivized" ] } }, "message": { "category": "a decentralized AI training network (tplr.ai)", "h1": "ONE COVENANT, MANY ORDERS", "promise": "Market-driven incentives for loss reduction, enabling permissionless distributed deep learning.", "proof": "unknown", "audience": "unknown", "cta": "EXPLORE TEMPLAR / GET STARTED", "label": "verified" }, "voice": { "carries_label": "mixed: founder voice, research substance", "read": "2026-09-29", "carries": "Before the April exit, Sam Dare's personal account carried the mission (Bittensor as prophecy, permissionless incentivised pretraining) and the results ran on the brand: the 72B post drew 6,099 likes on the Templar account against 267 on his quote of it. Since May the Templar account carries the substance (PULSE, Crucible) in a research register at 3 to 34 likes a post, the parent account has been silent since 24 June, and Dare quote-posts with one-line taglines. The exit letter was posted on the company account and signed by him: 1.6 million views.", "message": "\"The internet is the datacenter.\" Sam Dare, 3 September 2026, repeated 16 September, and by the company account on 12 June.", "accounts": [ { "channel": "@covenant_ai", "url": "https://x.com/covenant_ai", "note": "5,104 followers; bio \"One covenant, many orders.\"; no post since 24 June" }, { "channel": "@tplr_ai", "url": "https://x.com/tplr_ai", "note": "12,910; \"incentivised internet-wide training\"; six Crucible threads in September" }, { "channel": "YouTube", "url": "https://www.youtube.com/@CovenantLabs", "note": "74 subscribers; last upload 9 March (TGIF #28, with Jacob Steeves)" }, { "channel": "Substack", "url": "https://templarresearch.substack.com", "note": "five posts; last 20 May" }, { "channel": "Discord", "url": "https://discord.gg/N5xgygBJ9r", "note": "823 members" }, { "channel": "GitHub", "url": "https://github.com/one-covenant", "note": "templar 162 stars, last push 31 March; basilica pushed 28 September" } ], "founders": [ { "name": "Sam Dare", "role": "founder", "handle": "DistStateAndMe", "bio": "\"Founder / Member of Non Technical Staff @covenant_ai\" (4,584 followers)", "statements": [ { "said": "\"To our knowledge this is the first economic validation of decentralised pretraining.\" On Crucible: 8B on 48 A100s at $0.12 per million tokens.", "medium": "X", "venue": "quote-post of the Templar account", "date": "2026-09-03", "url": "https://x.com/DistStateAndMe/status/2095584616711110915" }, { "said": "\"It is decentralization theatre.\" \"Decentralized, permissionless AI training is not a Bittensor feature.\"", "medium": "X article", "venue": "the exit statement, on the company account, signed by him", "date": "2026-04-09", "url": "https://x.com/covenant_ai/status/2042380152831951300" }, { "said": "\"I sold my tokens because I couldn't see a reality where we remained a going concern.\"", "medium": "X", "venue": "reply", "date": "2026-04-28", "url": "https://x.com/DistStateAndMe/status/2049204831198544247" }, { "said": "\"Nobody had hope in crypto any more... But we posted about our 72B parameter run and it caught fire.\"", "medium": "podcast", "venue": "This Week in Startups E2268", "date": "2026-03-27", "url": "https://www.linkedin.com/videos/jasoncalacanis_nobody-had-hope-in-crypto-any-more-but-activity-7443452801449398272-VHJB" }, { "said": "\"Innovation happens at the edge. We innovate through scarcity.\"", "medium": "X", "venue": "the Covenant-72B post, pinned, 74k views", "date": "2026-03-10", "url": "https://x.com/DistStateAndMe/status/2031399702088814680" }, { "said": "\"The worlds most powerful foundation model will be built on Bittensor.\" Seven months before the exit.", "medium": "X", "venue": "", "date": "2025-09-04", "url": "https://x.com/DistStateAndMe/status/1963646903650062570" } ], "recurring_words": [ "the internet is the datacenter", "permissionless", "incentivised", "decentralised pretraining", "stranded compute", "commodity internet", "going concern", "mission bound", "LFG" ] } ], "reception": "Jacob Steeves, 10 April: \"I do not have the ability to suspend emissions.\" 11 April: \"He betrayed us all.\" (560k views). TAO fell 15 to 20 percent around the exit and the subnet tokens 50 to 68 percent, per press." }, "largest_model": { "value": "**Covenant-72B**, the same run: 72B dense, about 1.1T tokens, weights public.", "label": "verified", "parameters_b": 72 } }, { "id": "pluralis", "name": "Pluralis Research", "kicker": "Protocol Learning · Agora", "url": "https://pluralis.ai/", "relation": "reference", "secondary_relation": "rival-supply", "last_verified": "2026-09-29", "people": [ { "name": "Alexander Long", "role": "founder and CEO", "handle": "AlexanderLong", "said": "March 2025: low-bandwidth model parallelism \"remains an unsolved research challenge\"; \"the core thesis of Pluralis is that this is solvable.\"" }, { "name": "Thalaiyasingam Ajanthan", "role": "founding scientist", "handle": "tha_ajanthan" } ], "what_it_is": { "value": "\"A research lab focused on decentralized AI\" whose method, Protocol Learning, trains a model split across many participants so that \"no single party ever possesses the complete weights\" (Agora paper).", "label": "verified" }, "mechanism": { "summary": { "value": "Pipeline-parallel across 24 GB cards, one stage per worker, seven stages at 8B; activations compressed up to 100x. Verification is code integrity only. Head, tail and averaging ran on Pluralis nodes.", "label": "verified" }, "detail": { "value": "**Pipeline** stages across contributor GPUs plus data-parallel within a stage; Subspace Networks compress activations up to 100x, Async SPARTA averages 5% of parameters every 5 steps in place of per-step gradient all-reduce. Failed peers are banned for the round and batches reroute; fatal only if a whole stage empties. **Verification is weak by its own account**: \"a weak form of verification ... that checks code integrity\"; adversarial defense is roadmap. **Not fully decentralized**: in the 8B run Pluralis nodes held the head and tail and did every reduction (\"this run was not decentralized and required specific Pluralis nodes\"). Admission gates on hardware and RTT under 80 ms; the current 13B docs require North America, and other regions are \"routinely rejected\".", "label": "verified" } }, "largest_run": { "parameters_b": 8.6, "summary": { "value": "**Pluralis-8B**, 500B tokens, 330 contributor nodes over 40 days, about 20% MFU, \"63% of the efficiency of a centralized H100 baseline\" in tokens per pooled TFLOP. Trained 14 May to 23 June 2026.", "label": "verified" }, "detail": { "value": "**Pluralis-8B**: 8.6B, 500B tokens, 330 contributor nodes over the run (4090 and 5090 55%, L40S 22%, RTX 6000 Ada and PRO 6000 23%), at most 60 contributors at a time plus Pluralis-run head and tail nodes, 40 days, 170k tokens a second. MFU about 20% (X); 24.2% pooled in the peak window (paper). \"63% of the efficiency\" is 4.2 tokens per pooled spec TFLOP against Llama 3.1 8B on TorchTitan, H100. Trained 14 May to 23 June 2026; completion announced 6 July. A 13B systems test ran in early September: 1,215 nodes (278 Pluralis-run), peak 364, 62.4B tokens, 14.9% MFU, 61% protocol overhead.", "label": "verified", "note": "Whether the weights are public is not stated; the design says unextractable." } }, "sold": { "value": "Nothing. Contributors earn points; \"there is no financial reward tied to your score at this stage.\" No token. The homepage application is a community sign-up, not a sale.", "label": "verified" }, "paid": { "value": "None public. The only money out is USD 18,000 in ProtocolNanoGPT speedrun prizes.", "label": "verified" }, "funding": { "value": "$7.6M seed, March 2025, led by USV and CoinFund (company release). No later round found. About 19 staff.", "label": "stated", "note": "Headcount from Tracxn (secondhand)." }, "relation_argument": "**Reference**: the closest mechanism to IOTA's, one pipeline stage per 24 GB card over the internet with compressed activations between stages. **Rival for supply**, limited: its open runs admit the same 4090, 5090 and L40S class IOTA's miners use, and a card runs one job at a time; runs are capped and episodic and pay points with no financial reward. Not a rival for buyers: it sells nothing.", "sources": [ { "title": "Agora paper, arXiv 2607.13332, 14 July 2026", "url": "https://arxiv.org/abs/2607.13332" }, { "title": "Agora system docs, fault tolerance", "url": "https://pluralis.ai/docs/agora-system/fault-tolerance/" }, { "title": "Agora docs, 13B systems test (North America, 80 ms cap)", "url": "https://pluralis.ai/docs/quick-start/" }, { "title": "Agora dashboard (13B test numbers), read 29 September 2026", "url": "https://agora.pluralis.ai" }, { "title": "Multi-party training stack, January 2026 (verification quote)", "url": "https://pluralis.ai/blog/pluralis-multi-party-training-stack/" }, { "title": "Pluralis on X, 6 July 2026 (run announced done)", "url": "https://x.com/Pluralis/status/2074232225344995636" }, { "title": "Points page (no financial reward)", "url": "https://pluralis.ai/docs/quick-start/points/" }, { "title": "Seed round release, 19 March 2025", "url": "https://www.globenewswire.com/news-release/2025/03/19/3045635/0/en/Pluralis-Research-Pioneers-Protocol-Learning-to-Scale-Decentralized-AI-Announces-7-6M-Seed-Round-Led-by-USV-and-CoinFund.html" } ], "signals": { "url": "https://pluralis.ai/", "read": "2026-09-28", "js_rendered": false, "title": "Pluralis Research", "h1": "PLURALIS RESEARCH", "first_h2": "About", "meta_description": "Pluralis Research works on Protocol Learning: decentralized, communication-efficient model-parallel training for foundation models.", "og_description": "Protocol Learning: decentralized, communication-efficient model-parallel training for foundation models.", "we_are_sentence": "“To speak of justice requires questioning the global distribution of power that decides who in fact can train these models and who is merely subjected to them.”", "proof_numbers": [ "8.6B-parameter open pretraining run trained on 500B tokens across 330 contributor nodes" ], "audience_named": [ "unknown" ], "cta_verb": "Docs", "contested_words_present": [ "decentralized", "permissionless", "open", "collaborative" ], "token_symbols": [], "dates_on_page": [ "unknown" ] }, "message": { "category": "Pluralis Research works on Protocol Learning", "h1": "PLURALIS RESEARCH", "promise": "decentralized, communication-efficient model-parallel training for foundation models", "proof": "8.6B-parameter open pretraining run trained on 500B tokens across 330 contributor nodes", "audience": "unknown", "cta": "Docs", "label": "verified" }, "voice": { "carries_label": "founder-led argument, research-led brand", "read": "2026-09-28", "carries": "Alexander Long's account (3,635 followers) carries every ideological post; the company account (13,370) posts run metrics, paper acceptances and workshop logistics with almost no opinion, and the blog has had no Long byline since March 2025. Since April the policy framing is delivered at the ICLR and ICML workshops by Riccardo Patana, ex-Anthropic, hired as Head of Strategy, Product and Safety; Ajanthan fronts the academic community as lead organizer of the NeurIPS workshop.", "message": "\"Collaborative training and development of the models without anyone ever being able to see the complete weight set.\" Long's definition of protocol learning, 13 June 2026. The frame around it, repeated since March 2025: \"a third path\" between closed weights and open weights.", "accounts": [ { "channel": "@Pluralis", "url": "https://x.com/Pluralis", "note": "13,370; bio \"a research lab focused on collectively-owned AI\"; three posts in 30 days" }, { "channel": "Blog", "url": "https://pluralis.ai/blog/", "note": "eleven posts since July 2024, latest July 2026; research staff bylines" }, { "channel": "YouTube", "url": "https://www.youtube.com/@Pluralis_Research", "note": "112 subscribers; ICLR 2026 workshop recordings" }, { "channel": "LinkedIn", "url": "https://www.linkedin.com/company/pluralis-research", "note": "1,971; \"enables anyone to train and own frontier AI models\"" }, { "channel": "GitHub", "url": "https://github.com/PluralisResearch", "note": "node0 98 stars; ProtocolNanoGPT pushed today" } ], "founders": [ { "name": "Alexander Long", "role": "founder and CEO", "handle": "AlexanderLong", "bio": "\"Founder @Pluralis | ML PhD\" (3,635 followers; one original post in 30 days)", "statements": [ { "said": "\"If you cannot split a model over participants, I don't see how you keep the weight set private.\"", "medium": "X", "venue": "quote-posting Macrocosmos's ResBM post (128x activation compression, built for IOTA), without naming Bittensor", "date": "2026-04-13", "url": "https://x.com/AlexanderLong/status/2043766866825691380" }, { "said": "\"No-one aside from Pluralis is earnestly moving towards a realistic solution.\" \"I wanna own and impact the process not the output.\"", "medium": "X", "venue": "essay quoting Beff Jezos and Martin Casado", "date": "2026-09-13", "url": "https://x.com/AlexanderLong/status/2099264168805441806" }, { "said": "\"The only way out of this is to have an independent model supply chain via pooled compute.\"", "medium": "X", "venue": "", "date": "2026-06-09", "url": "https://x.com/AlexanderLong/status/2064443807522074924" }, { "said": "\"It lets you split the model up over participants, which gives you this path to sustainable economics.\"", "medium": "YouTube talk", "venue": "Protocol Learning Workshop, ICLR 2026", "date": "2026-04-26", "url": "https://www.youtube.com/watch?v=_2_S2qzsNXU" }, { "said": "\"The only two companies in decentralised AI that have main track papers this year are @PrimeIntellect and Pluralis.\"", "medium": "X", "venue": "", "date": "2025-07-14", "url": "https://x.com/AlexanderLong/status/1944553475184206241" }, { "said": "\"The holy grail of decentralized AI is Model-Parallel training over low-bandwidth interconnects. This is all Pluralis cares about.\"", "medium": "X", "venue": "", "date": "2025-04-18", "url": "https://x.com/AlexanderLong/status/1913346048124502297" } ], "recurring_words": [ "third path", "protocol learning", "unextractable", "weight set", "model supply chain", "pooled compute", "swarm", "consumer GPUs", "sovereign" ] }, { "name": "Thalaiyasingam Ajanthan", "role": "founding scientist", "handle": "tha_ajanthan", "bio": "\"Founding Scientist @Pluralis | Ex @AmazonScience, @OxfordTVG\" (451 followers; eight posts in 30 days)", "statements": [ { "said": "\"The TPS cost to use such liquid compute is only 1.5x than optimized centralised systems.\" The word \"liquid compute\" is theirs too.", "medium": "X", "venue": "the Agora release", "date": "2026-05-21", "url": "https://x.com/tha_ajanthan/status/2057347243217412333" }, { "said": "\"Frontier-scale training shouldn't require a hyperscaler's cluster.\"", "medium": "X", "venue": "call for papers, CODEC-FM at NeurIPS 2026", "date": "2026-07-27", "url": "https://x.com/tha_ajanthan/status/2081877294621364543" }, { "said": "\"We now want to see its scalability (in no of participants) and where it breaks.\"", "medium": "X", "venue": "the 13B stress test", "date": "2026-09-04", "url": "https://x.com/tha_ajanthan/status/2095737264659947943" } ], "recurring_words": [ "communication-efficient", "geo-distributed", "liquid compute", "economically sustainable", "collectively train (and own)" ] } ] }, "largest_model": { "value": "A **13B** systems test, about 1 to 8 September 2026: 1,215 nodes, peak 364, 62.4B tokens, 14.9% MFU. Not a completed model; the largest completed run is Pluralis-8B.", "label": "verified" } }, { "id": "prime-intellect", "name": "Prime Intellect", "kicker": "Hosted RL · GPU marketplace", "url": "https://www.primeintellect.ai/", "relation": "rival-buyer", "secondary_relation": "rival-supply", "last_verified": "2026-09-29", "people": [ { "name": "Vincent Weisser", "role": "CEO", "handle": "vincentweisser", "said": "July 2026: training should belong to \"every enterprise, every nation state\" (TechCrunch)." }, { "name": "Johannes Hagemann", "role": "CTO", "handle": "johannes_hage" }, { "name": "Sami Jaghouar", "role": "head of research", "handle": "samsja19" } ], "what_it_is": { "value": "\"The Open Superintelligence Stack\": compute, training, inference and sandboxes. The decentralized runs were 2024 and 2025 research, and the peer-to-peer protocol repository is archived; the 2026 business is hosted RL and a GPU marketplace.", "label": "verified" }, "mechanism": { "summary": { "value": "Data-parallel DiLoCo (each node a full replica, int8 pseudo-gradients, 400x less traffic) on H100 nodes from 30 compute providers. TOPLOC verifies RL rollouts from untrusted workers, not training. No decentralized run to join today.", "label": "verified" }, "detail": { "value": "**Data-parallel** DiLoCo with FSDP inside each node; int8 pseudo-gradient all-reduce, about 400x less traffic than DDP; ElasticDeviceMesh lets nodes join and leave mid-run. Pretraining ran on up to 14 nodes (112 H100s) from 30 independent providers, none verified; no open sign-up is on record. For RL, TOPLOC (a hash of activations) checks inference rollouts from a permissionless swarm; training stays on trusted nodes. The peer-to-peer protocol repository is archived (last code push 10 November 2025). No public decentralized run to join as of today.", "label": "verified" } }, "largest_run": { "parameters_b": 10, "summary": { "value": "**INTELLECT-1**, 10B, 1T tokens, up to 112 H100s across three continents, 36 to 41% MFU, November 2024. Weights public. Its largest run outside one reserved cluster; INTELLECT-3 ran on one cluster.", "label": "verified" }, "detail": { "value": "**INTELLECT-1**: 10B, 1T tokens, up to 112 H100 at once from 30 providers in five countries on three continents, 42 days, 83% global compute utilization, 36.2 to 41.4% MFU, November 2024; base model, checkpoints and data released. INTELLECT-2 (launched 15 April 2025, released May 2025) was 32B asynchronous RL with a permissionless rollout swarm. INTELLECT-3 (26 November 2025), a 106B mixture-of-experts with 12B active, ran on a single 512 H200 cluster over about two months. 2026 RL work, up to 1T parameters on 28 H200 nodes (June 2026), is single cluster. No decentralized run in 2026.", "label": "verified" } }, "sold": { "value": "A training run can be bought today, self-serve: Lab hosted RL post-training (LoRA on open models up to 35B-A3B, priced per million tokens; full fine-tuning in closed beta), GPUs on demand from 50+ providers (H100 $2.43 an hour, spot $0.94, as displayed on the home page), reserved clusters with idle-hour resale, inference, sandboxes. No pretraining offer.", "label": "verified" }, "paid": { "value": "**Ramp** (Series A post and case study); **Zapier** (TechCrunch and case study); **Flapping Airplanes** (TechCrunch); **Goodfire** (case study). A **Mixedbread** case study says it used prime-rl, the open library. \"Over 6k customers\" and \"over $100m in annualized revenue\", by the company.", "label": "stated" }, "funding": { "value": "$130M Series A, 8 July 2026, led by Radical Ventures with NVIDIA, Intel, Dell and Iconiq (own post and TechCrunch); over $150M total; $1B valuation (press). No token. San Francisco.", "label": "stated" }, "relation_argument": "Rival-buyer: Prime sells hosted RL post-training priced per token and GPU clusters of up to 256 on demand, self-serve, and names Ramp, Zapier and \"over 6k customers\". IOTA's SDK was pitched for RL and fine-tuning on 28 September 2026 (Will Squires), the same work. Rival-supply at the datacenter tier only: the marketplace takes A100-class and newer cards from 50+ providers and has no route for a single 4090 owner. No pretraining product and no decentralized run since 2025.", "sources": [ { "title": "INTELLECT-1 release, 29 November 2024", "url": "https://www.primeintellect.ai/blog/intellect-1-release" }, { "title": "INTELLECT-1 technical report, arXiv 2412.01152", "url": "https://arxiv.org/abs/2412.01152" }, { "title": "prime-diloco (ElasticDeviceMesh), last commit 10 April 2025", "url": "https://github.com/PrimeIntellect-ai/prime-diloco" }, { "title": "TOPLOC, arXiv 2501.16007", "url": "https://arxiv.org/abs/2501.16007" }, { "title": "Series A post, 8 July 2026", "url": "https://www.primeintellect.ai/blog/series-a" }, { "title": "TechCrunch, 8 July 2026", "url": "https://techcrunch.com/2026/07/08/prime-intellect-raises-130m-series-a-to-help-enterprises-build-their-own-ai-agents/" }, { "title": "INTELLECT-3: Technical Report (arXiv 2512.16144)", "url": "https://arxiv.org/abs/2512.16144" }, { "title": "INTELLECT-3 release, 26 November 2025", "url": "https://www.primeintellect.ai/blog/intellect-3" }, { "title": "Case studies (Ramp, Zapier, Mixedbread, Goodfire)", "url": "https://www.primeintellect.ai/case-study" }, { "title": "Hosted Training models and pricing", "url": "https://docs.primeintellect.ai/hosted-training/models-and-pricing.md" }, { "title": "protocol repository, archived", "url": "https://github.com/PrimeIntellect-ai/protocol" } ], "signals": { "url": "https://www.primeintellect.ai/", "read": "2026-09-29", "js_rendered": false, "title": "Prime Intellect - The Open Superintelligence Stack", "h1": "Own Your Intelligence", "first_h2": "Join Prime Intellect", "meta_description": "Train, deploy, and continuously improve your own models on an integrated compute, training, inference, and sandbox stack.", "og_description": "Train, deploy, and continuously improve your own models on an integrated compute, training, inference, and sandbox stack.", "we_are_sentence": "Train, deploy, and continuously improve your own models on an integrated compute, training, inference, and sandbox stack.", "proof_numbers": [ "FIND COMPUTE BOOK A DEMO On demand Instant access to 1-256 GPUs", "1 Get quotes from 50+ datacenters within 24 hours One request, parallel bids for options, fro", "INTELLECT-3: A 100B+ MoE trained with large-scale RL A 100B+ parameter Mixture-of-Experts model trained on our RL stack", "5 Single-Node Multi-Node H200 Available x2 · x1 $1.99/HR 80 GB VRAM · 184 GB RAM · 32 vCP H200 Availabl", "e x2 · x1 $1.80/HR 80 GB VRAM · 184 GB RAM · 32 vCP H200 Availabl", "e x2 · x1 $1.23/HR 80 GB VRAM · 184 GB RAM · 32 vCP H200 Availabl", "e x2 · x1 $0.47/HR 80 GB VRAM · 184 GB RAM · 32 vCP B300 Availabl", "e x2 · x1 $4.99/HR 288 GB VRAM · 480 GB RAM · 48 vCP B200 Availab", "le x2 · x1 $3.49/hr 192 GB VRAM · 384 GB RAM · 32 vCP H200 Availab", "le x2 · x1 $3.14/HR 141 GB VRAM · 182 GB RAM · 44 vCPUs H100 Avail", "able x2 · x1 $2.43/HR Spot 0", "M · 185 GB RAM · 32 vCPUs GH200 Available x2 · x1 $3.14/HR 96 GB VRAM · 480 GB RAM · 72 vCP RTX Pro 6000" ], "audience_named": [ "unknown" ], "cta_verb": "Start training; Book a call", "contested_words_present": [ "distributed", "open", "collaborative" ], "token_symbols": [], "dates_on_page": [ "unknown" ], "eyebrow": "The Open Superintelligence Stack", "read_via": "hero checked by hand against the live page on 2026-09-29; the scraper had joined the eyebrow to the headline and taken a call to action from lower on the page" }, "message": { "category": "The Open Superintelligence Stack", "h1": "Own Your Intelligence", "promise": "Train, deploy, and continuously improve your own models on an integrated compute, training, inference, and sandbox stack.", "proof": "INTELLECT-3: A 100B+ MoE trained with large-scale RL", "audience": "unknown", "cta": "Start training; Book a call", "label": "verified" }, "voice": { "carries_label": "brand-led, founder echoes", "read": "2026-09-28", "carries": "The company account (84,789 followers) out-reaches all three founders combined and carries the launches: the Environments Hub post drew 1.8 million views against 22.9 thousand for Weisser's protocol post the same season, and launches go out on the company account first. Every blog post, including the $130M Series A, is signed \"Prime Intellect Team\". Hagemann and Jaghouar are near-silent as public voices in what could be read.", "message": "\"Towards an open superintelligence future.\" Weisser, 14 February 2025; the same two words sit in their bio and the company's.", "accounts": [ { "channel": "@PrimeIntellect", "url": "https://x.com/PrimeIntellect", "note": "84,789; bio \"Open Superintelligence Stack\"" }, { "channel": "YouTube", "url": "https://www.youtube.com/@PrimeIntellect", "note": "five videos in 2026; last 14 August" }, { "channel": "Blog", "url": "https://www.primeintellect.ai/blog", "note": "two posts in 30 days; all signed Prime Intellect Team" }, { "channel": "Discord", "url": "https://discord.gg/primeintellect", "note": "8,388 members" }, { "channel": "LinkedIn", "url": "https://www.linkedin.com/company/primeintellect-ai", "note": "11,948; \"makes frontier AI training accessible to every company\"" }, { "channel": "GitHub", "url": "https://github.com/PrimeIntellect-ai", "note": "prime-rl 2,096 stars, pushed daily" } ], "founders": [ { "name": "Vincent Weisser", "role": "CEO", "handle": "vincentweisser", "bio": "\"ceo @primeintellect, open superintelligence\" (32,000 followers)", "statements": [ { "said": "\"The biggest risk is actually locking in a very narrow monoculture for superintelligence.\" \"One superintelligence is much less safe than infinite superintelligence.\"", "medium": "podcast", "venue": "The Generalist", "date": "2026-03-24", "url": "https://www.generalist.com/p/why-one-superintelligence-is-more" }, { "said": "\"You'll get the most compute per dollar using app.primeintellect.ai, at any scale, on demand.\"", "medium": "X", "venue": "selling the marketplace", "date": "2025-08-15", "url": "https://x.com/vincentweisser/status/1956381076080730544" }, { "said": "\"A peer-to-peer compute and intelligence network. Enabling collective creation, ownership, and access of sovereign open-source AI.\"", "medium": "X", "venue": "the protocol launch", "date": "2025-02-14", "url": "https://x.com/vincentweisser/status/1890468028287943036" }, { "said": "\"Advancing open source + decentralized AI with decentralized training and aggregating global compute.\"", "medium": "X", "venue": "quote-posting INTELLECT-1", "date": "2024-11-04", "url": "https://x.com/vincentweisser/status/1853264163663163698" } ], "recurring_words": [ "open superintelligence", "decentralized training", "sovereign open-source AI", "compute", "aggregating global compute", "peer-to-peer" ] }, { "name": "Johannes Hagemann", "role": "CTO", "handle": "johannes_hage", "bio": "\"co-founder/cto @PrimeIntellect | open superintelligence infra, longevity, techno-optimism\" (10,885)", "statements": [], "recurring_words": [] }, { "name": "Sami Jaghouar", "role": "head of research", "handle": "samsja19", "bio": "\"leading research at @PrimeIntellect\" (9,057)", "statements": [ { "said": "\"We trained the 10b model over 1T tokens across 3 continents.\"", "medium": "X", "venue": "INTELLECT-1", "date": "2024-11", "url": "https://x.com/samsja19/status/1859996275946737826" } ], "recurring_words": [] } ] }, "largest_model": { "value": "**INTELLECT-3**, a 106B mixture-of-experts with 12B active parameters, post-trained from GLM 4.5 Air with supervised fine-tuning and large-scale RL on 512 H200s across 64 nodes in one cluster, about two months; released 26 November 2025. Weights open.", "label": "verified", "parameters_b": 106, "active_b": 12, "note": "One cluster, not a distributed run. The largest run outside one reserved cluster remains INTELLECT-1 at 10B." } }, { "id": "nous", "name": "Nous Research", "kicker": "Psyche (NousNet) · DisTrO · DeMo", "url": "https://nousresearch.com/", "relation": "reference", "secondary_relation": null, "last_verified": "2026-09-29", "people": [ { "name": "Jeffrey Quesnelle", "role": "co-founder, CTO (own bio; press says CEO)", "handle": "theemozilla", "said": "February 2026: \"The smart contract's job is to assign work and ensure consensus on task completion.\" (Crypto Briefing)" }, { "name": "Teknium", "role": "co-founder, Hermes", "handle": "Teknium" }, { "name": "Bowen Peng", "role": "DeMo lead author", "handle": "bloc97_" } ], "what_it_is": { "value": "An open-model lab that now sells agents. Psyche, renamed NousNet in its repository, is \"distributed training of transformer-based AI models over the internet\" for \"collaboration between untrusted parties\"; its optimizer is DisTrO, from the DeMo paper.", "label": "verified" }, "mechanism": { "summary": { "value": "Data-parallel: every node holds a full replica, in practice a multi-GPU server; DeMo cuts traffic up to two orders of magnitude. Witness quorum on a Solana coordinator; recompute-and-punish described, not confirmed live. Joining needs a binary from the run administrator.", "label": "verified" }, "detail": { "value": "**Data-parallel**: every client trains a full replica on its shard, split across its GPUs by tensor parallel, and gossips a DeMo-compressed update (decoupled momentum, DCT plus top-k; up to 85x less data per GPU than AdamW-DDP in the paper). Elected **witnesses** vote on which results are applied; the docs list a recompute-and-punish task, but no source says it is live. Dropped clients are removed at the next round; below a minimum the run pauses. The authorizer allows permissionless runs, but the join guides assume a `run-manager` binary \"provided by the run administrator\". Rewards are points, with an optional token treasurer. The coordinator program's last transaction was 27 July 2026.", "label": "verified", "note": "psyche.network refused connections on 29 September 2026; docs read from the repository (PsycheFoundation/nousnet, formerly psyche), last pushed 24 March 2026. Coordinator activity read from Solana RPC: devnet last 27 July 2026, mainnet last 11 April 2026." } }, "largest_run": { "parameters_b": 40, "summary": { "value": "**Consilience 40B**, started 14 May 2025, last public checkpoint step 23,216 on 14 August 2025. Tokens, hardware, sites and MFU not published. Checkpoints public.", "label": "stated" }, "detail": { "value": "**Consilience 40B**: dense, 20T-token target, started 14 May 2025 on the Psyche testnet; Nous said the model fits \"a single H/DGX\". The last public checkpoint is step 23,216 on 14 August 2025. In September 2025 Nous called it \"the largest distributed pre-training run ever\", said the testnet run had proved training over internet bandwidth, and moved Psyche to ablations and Hermes 4.3. Checkpoints every 500 steps on Hugging Face.", "label": "stated", "note": "Tokens trained, node count, sites and MFU: unknown. Whether the 20T target was reached: not said; the model card still reads \"Training Duration: TBD\". Two checkpoint repositories exist (CqX3FUm4 ends at step 20,148 on 1 August 2025). Hermes-4.3 post-training ran on Psyche across 24 nodes in data centers at 144k tokens a second, December 2025." } }, "sold": { "value": "Hermes Agent (open source), Nous Portal subscriptions ($0 to $200 a month: inference on 300+ models, hosted agents), Hermes Business and Enterprise (14 September 2026). Psyche compute is not sold; no priced training offer.", "label": "verified" }, "paid": { "value": "None public for training.", "label": "verified" }, "funding": { "value": "$50M Series A, April 2025, Paradigm, at $1B on a future token (press); at least $75M at $1.5B led by Robot Ventures with USV, in talks July 2026 (TechCrunch; aggregators report it closed, Nous has not announced it). No official token.", "label": "secondhand" }, "relation_argument": "Reference: DeMo is a compression method in IOTA's lineage, and Consilience 40B is a published benchmark for pretraining outside one reserved cluster. Not a rival-buyer: Nous sells agents, inference and team accounts, with no training offer. Not a rival-supply: joining a Psyche run needs a binary from the run administrator, and the coordinator's last transaction was 27 July 2026.", "sources": [ { "title": "Psyche (NousNet) repository and docs", "url": "https://github.com/PsycheFoundation/nousnet" }, { "title": "Psyche architecture post, May 2025", "url": "https://nousresearch.com/nous-psyche" }, { "title": "DeMo, arXiv 2411.19870", "url": "https://arxiv.org/abs/2411.19870" }, { "title": "Consilience 40B checkpoints (to step 20,148)", "url": "https://huggingface.co/PsycheFoundation/consilience-40b-CqX3FUm4" }, { "title": "Hermes 4.3 on Psyche, December 2025", "url": "https://nousresearch.com/introducing-hermes-4-3" }, { "title": "The Block on the Series A, 25 April 2025", "url": "https://www.theblock.co/post/352000/paradigm-leads-50-million-usd-round-decentralized-ai-project-nous-research" }, { "title": "Hermes-4-405B model card", "url": "https://huggingface.co/NousResearch/Hermes-4-405B" }, { "title": "Consilience 40B checkpoints (to step 23,216, 14 August 2025)", "url": "https://huggingface.co/PsycheFoundation/consilience-40b-7Y9v38s5" }, { "title": "The Next Phase of Psyche, September 2025", "url": "https://nousresearch.com/the-next-phase-of-psyche" }, { "title": "TechCrunch on the $1.5B talks, 13 July 2026", "url": "https://techcrunch.com/2026/07/13/hermes-agent-maker-nous-research-in-talks-for-new-funding-at-1-5b-valuation/" } ], "signals": { "url": "https://nousresearch.com/", "read": "2026-09-29", "js_rendered": false, "title": "Nous Research", "h1": "Nous Research", "first_h2": "The Internet's Own AI", "meta_description": "Nous Research is a pioneer in open AI training and research. We created Hermes Agent, the most widely used open source agent harness in the world. We are on a mission to create and proliferate open access to intelligence.", "og_description": "Nous Research is a pioneer in open AI training and research. We created Hermes Agent, the most widely used open source agent harness in the world. We are on a mission to create and proliferate open access to intelligence.", "we_are_sentence": "Nous Research is a pioneer in open AI training and research. We created Hermes Agent, the most widely used open source agent harness in the world.", "proof_numbers": [ "4% smaller, saving us nearly $2m in engineering hours" ], "audience_named": [ "for agents" ], "cta_verb": "Read mission; The Internet's Own AI", "contested_words_present": [ "open", "sovereign" ], "token_symbols": [], "dates_on_page": [ "Sep 16, 2026", "Sep 17, 2026", "Sep 21, 2026", "Sep 23, 2026", "Sep 24, 2026" ], "read_via": "checked by hand against the 28 September capture on 2026-09-29" }, "message": { "category": "a pioneer in open AI training and research", "h1": "Nous Research", "promise": "We created Hermes Agent, the most widely used open source agent harness in the world.", "proof": "unknown", "audience": "for agents", "cta": "Read mission; The Internet's Own AI", "label": "verified" }, "voice": { "carries_label": "brand-led launches, founder-led narrative", "read": "2026-09-28", "carries": "The company account (275,027 followers) announces every launch in its own voice without quoting a founder, and blog posts are bylined Nous Research or pseudonyms, so the product story is brand-led. The people story rides on Teknium (130,042, the only founder in the brand's league), who posts first-person release notes on launch day and wrote the one founder-bylined post of 2026. Quesnelle and Malhotra carry the podcast and press circuit. A 141,148-member Discord is the largest community on this page.", "message": "\"Our mission is to be the open AI accelerator.\" Quesnelle, March 2025. Company copy today: \"We are on a mission to create and proliferate open access to intelligence.\"", "accounts": [ { "channel": "@NousResearch", "url": "https://x.com/NousResearch", "note": "275,027; bio \"A bunch of nerds making progress\"; links to the Hermes Agent site" }, { "channel": "Discord", "url": "https://discord.gg/nousresearch", "note": "141,148 members" }, { "channel": "YouTube", "url": "https://www.youtube.com/@Nousresearch", "note": "no video in 2026; last 19 December 2025" }, { "channel": "Blog", "url": "https://nousresearch.com/blog", "note": "bylines are Nous Research or pseudonyms; one founder byline (Teknium, September)" }, { "channel": "GitHub", "url": "https://github.com/NousResearch", "note": "hermes-agent 249,843 stars; DisTrO last push October 2025; Psyche (now nousnet) last push 24 March" } ], "founders": [ { "name": "Jeffrey Quesnelle", "role": "co-founder; bio says CTO", "handle": "theemozilla", "bio": "\"catholic, ai researcher, co-founder/cto of @NousResearch\" (13,504)", "statements": [ { "said": "\"At any moment, only about 50% of the GPUs in data centers are actually active.\" \"The entire game is intelligence per unit of energy.\"", "medium": "podcast", "venue": "Raoul Pal, The Journey Man, via Crypto Briefing", "date": "2026-02-18", "url": "https://cryptobriefing.com/jeffrey-quesnelle-centralization-in-ai-is-stifling-innovation-how-decentralization-can-democratize-access-and-the-critical-role-of-smart-contracts-in-ai-training-raoul-pal-the-journey-man/" }, { "said": "\"Using crypto rails allows for permissionless and disintermediated access to computing resources.\"", "medium": "podcast", "venue": "The Journey Man", "date": "2026-02-18", "url": "https://cryptobriefing.com/jeffrey-quesnelle-centralization-in-ai-is-stifling-innovation-how-decentralization-can-democratize-access-and-the-critical-role-of-smart-contracts-in-ai-training-raoul-pal-the-journey-man/" }, { "said": "\"Our mission is to be the open AI accelerator.\"", "medium": "podcast", "venue": "Sina Habibian, Into the Bytecode", "date": "2025-03-18", "url": "https://sinahab.com/jeffrey-quesnelle/" }, { "said": "\"Is there any real reason we can't make Llama 4 ourselves?\"", "medium": "podcast", "venue": "a16z Podcast, DisTrO", "date": "2024-09-27", "url": "https://a16z.com/podcast/distro-and-the-quest-for-community-trained-ai-models/" } ], "recurring_words": [ "open source", "crypto rails", "Llama 4", "centralizing force", "permissionless", "intelligence per unit of energy", "community" ] }, { "name": "Teknium", "role": "co-founder, Hermes Agent", "handle": "Teknium", "bio": "\"Cofounder and Lead Engineer - Hermes Agent @NousResearch\" (130,042)", "statements": [ { "said": "\"I use Hermes Agent everyday to develop Hermes Agent.\" \"All for 1% of the cost and 1% of the time we'd estimated.\"", "medium": "blog", "venue": "the one founder-bylined post", "date": "2026-09", "url": "https://nousresearch.com/refactoring-hermes-with-1393-agents" }, { "said": "\"The best shot to win against tech oligopolies is to embrace open source and open science.\" \"Stay tuned to Psyche's progress, as we have a lot coming out soon on that.\"", "medium": "interview", "venue": "Delphi Digital AMA", "date": "2025-10-09", "url": "https://www.delphiintelligence.io/research/ama-1-transcript-with-nous-research-co-founder-and-post-training-lead-teknium1" } ], "recurring_words": [ "Hermes Agent", "aligned to you", "open source and open science", "self-improving", "local and open models" ] }, { "name": "Karan Malhotra", "role": "co-founder", "handle": "karan4d", "bio": "\"flower-seeking\" (26,031)", "statements": [ { "said": "\"We believe in intelligence as a public good before everything else.\"", "medium": "podcast", "venue": "Peter Yang, Behind the Craft", "date": "2026-08-02", "url": "https://youtu.be/UWjh5Z4s8jY" }, { "said": "Compute contributions are \"less as a donation but more as a transaction.\" \"We see crypto as the method that allows us to perform this in a safe way.\"", "medium": "interview", "venue": "Fortune, the Paradigm round", "date": "2025-04-25", "url": "https://finance.yahoo.com/news/exclusive-crypto-vc-giant-paradigm-114000156.html" } ], "recurring_words": [ "public good", "water and air", "crypto ethos", "personal agent" ] } ] }, "largest_model": { "value": "**Hermes 4 405B**, post-trained from Meta's Llama 3.1 405B, released 26 August 2025, with a 70B and a 14B beside it. The base was Meta's; the post-training hardware is not published.", "label": "stated", "parameters_b": 405, "note": "Not a pretraining run. The largest run outside one reserved cluster is Consilience 40B on Psyche." } }, { "id": "flower", "name": "Flower Labs", "kicker": "Photon · SuperGrid · Endeavor", "url": "https://flower.ai/", "relation": "reference", "secondary_relation": "rival-buyer", "last_verified": "2026-09-29", "people": [ { "name": "Daniel J. Beutel", "role": "CEO", "handle": "daniel_janes" }, { "name": "Nic Lane", "role": "co-founder and chief science officer, Cambridge", "handle": "niclane", "said": "1 September 2026: \"Europe should not have to rent its intelligence indefinitely from a handful of US companies\" (Tech.eu)." }, { "name": "Lorenzo Sani", "role": "Photon first author" } ], "what_it_is": { "value": "The federated-learning framework company. Photon is its federated LLM pretraining system; SuperGrid the managed platform; Endeavor 1.0 (1 September 2026) its frontier model, built by continual pretraining on an open-weight base.", "label": "stated" }, "mechanism": { "summary": { "value": "Cross-silo federated: each invited institution holds a full replica and trains on its own data, 64 to 512x less traffic. No verification of updates; trusted consortium only.", "label": "verified" }, "detail": { "value": "**Data-parallel, federated**: each client runs 62 to 512 local steps (as printed) on data it never shares, then a FedAvg outer step through a server or all-reduce; 64x to 512x less communication than standard data-parallel. **No verification of updates**: clients are invited institutions and the aggregator is one of them or a trusted third party. Flower Hub signs app code (preview, May 2026), not training updates. Partial updates from survivors are accepted in two of three modes. Not permissionless.", "label": "verified" } }, "largest_run": { "parameters_b": 7, "summary": { "value": "**Photon 7B**, four clients at 8 H100 each in Utah, Texas, Quebec and Maharashtra, aggregator in England; 0.65x the wall time of centralized, modelled at 10 Gbps; per-device MFU 0.071. A 13B run is blog-only.", "label": "verified" }, "detail": { "value": "**Photon 7B** (paper, v2 14 June 2026): four clients each with 8 H100s in four regions, aggregator in England; wall time 95.6 hours against 147.9 centralized, modelled assuming Ring-AllReduce at 10 Gbps and equal throughput; per-device MFU 0.071 against 0.105 (corrected values); perplexity 16.9% lower than the centralized baseline at 7B, 13.4 to 16.9% from 1.3B to 7B. A 13B run on four clients with more than 2x throughput appears in the May 2025 blog only. Lizzy-7B's training method is not stated; Endeavor is continual pretraining on an open-weight base.", "label": "verified", "note": "Checkpoints for Photon runs: unknown." } }, "sold": { "value": "SuperGrid: Pro at EUR 20, 50 or 200 a month billed yearly, more with added credits; Max by sales with confidential compute and SOC 2 reports; free accounts get 3,000 sign-up credits. Endeavor as a managed service or private deployment, preview. Lizzy 7B open weights. Pilot Program and forward-deployed engineers. No decentralized pretraining run can be bought.", "label": "verified" }, "paid": { "value": "Payment not stated for anyone. Flower's Customer Outcomes page names docport (with AstraZeneca) and PharosAI (NHS biobanks), built with its engineers, and Banking Circle and JPMorgan AI Research, built independently. Tech.eu calls NHS and JP Morgan clients.", "label": "stated", "note": "Tech.eu wording is secondhand." }, "funding": { "value": "$20M Series A, February 2024, Felicis; $23.6M total (own post). No later priced round found. No token. London and Hamburg (Tech.eu).", "label": "stated" }, "relation_argument": "**Reference**: the federated line of training across sites, each institution holding a full model copy and its own data; Macrocosmos compared IOTA with federated learning in June 2025. **Rival for buyers** in one segment only: its platform buyer owns GPUs and data, unlike IOTA's, but for sovereign and public-sector programs, which IOTA's July essay names, Flower sells Lizzy 7B open weights, private Endeavor deployment and engineering support today.", "sources": [ { "title": "Photon, arXiv 2411.02908 (v2 14 June 2026)", "url": "https://arxiv.org/abs/2411.02908" }, { "title": "Photon blog, 9 May 2025", "url": "https://flower.ai/blog/2025-05-09-photon" }, { "title": "Pricing, read 29 September 2026", "url": "https://flower.ai/pricing" }, { "title": "Customer Outcomes, read 29 September 2026", "url": "https://flower.ai/work-with-us" }, { "title": "Endeavor 1.0 post, 1 September 2026", "url": "https://flower.ai/blog/2026-09-01-introducing-endeavor-1.0" }, { "title": "Series A post, 15 February 2024", "url": "https://flower.ai/blog/2024-02-15-announcing-series-a" }, { "title": "Tech.eu on Endeavor, 1 September 2026", "url": "https://tech.eu/2026/09/01/cambridge-university-spinout-launches-ai-model-competitive-with-openai-and-anthropic/" }, { "title": "Flower blog", "url": "https://flower.ai/blog/" } ], "signals": { "url": "https://flower.ai/", "read": "2026-09-28", "js_rendered": false, "title": "Flower: Advance Collaborative Superintelligence", "h1": "Collaborative Superintelligence", "first_h2": "Today, we use less than 1% of the world's data.", "meta_description": "Flower is a full-stack AI neolab with a mission to advance collaborative superintelligence by building open-source frontier models, agents, and infrastructure.", "og_description": "Flower is a full-stack AI neolab with a mission to advance collaborative superintelligence by building open-source frontier models, agents, and infrastructure.", "we_are_sentence": "Compute and data are fundamentally decentralized.", "proof_numbers": [ "15T Tokens Public Data Used today 2000T Tokens Non-Public Data Mostly unused How? Collaborative AI Unlocks" ], "audience_named": [ "unknown" ], "cta_verb": "Register now; Try Endeavor", "contested_words_present": [ "decentralized", "open", "sovereign", "federated", "collaborative" ], "token_symbols": [], "dates_on_page": [ "unknown" ], "read_via": "checked by hand against the 28 September capture on 2026-09-29" }, "message": { "category": "a full-stack AI neolab", "h1": "Collaborative Superintelligence", "promise": "advance collaborative superintelligence by building open-source frontier models, agents, and infrastructure", "proof": "Today, we use less than 1% of the world's data.", "audience": "unknown", "cta": "Register now; Try Endeavor", "label": "verified" }, "voice": { "carries_label": "research-led blog, brand-led X, no CEO byline found", "read": "2026-09-28", "carries": "The company account (3,107 followers) has more followers than Beutel (548) and Lane (2,711) and does the announcing; its 21 September post had 175 views. The flagship posts (Endeavor 1.0, FlowerBench) are bylined by Lane and research staff, not the CEO; release posts are signed The Flower Team. YouTube runs monthly. The one located Lane post on X quotes Chamath on an \"attempted oligopoly on intelligence.\"", "message": "\"Advance Collaborative Superintelligence\" is identical across the X bio, the LinkedIn tagline and the GitHub description. No founder sentence was found that says it in the first person.", "accounts": [ { "channel": "@flwrlabs", "url": "https://x.com/flwrlabs", "note": "3,107; 21 September post had 175 views" }, { "channel": "YouTube", "url": "https://www.youtube.com/@flowerlabs", "note": "monthly; Flower AI Summit 2026 talks uploaded May" }, { "channel": "Blog", "url": "https://flower.ai/blog/", "note": "five posts in 30 days, four of them release notes signed The Flower Team" }, { "channel": "LinkedIn", "url": "https://www.linkedin.com/company/flwrlabs", "note": "11,782; \"full-stack AI neolab\"; names Owkin, Red Hat, J.P. Morgan as collaborators" }, { "channel": "GitHub", "url": "https://github.com/flwrlabs/flower", "note": "7,150 stars" } ], "founders": [ { "name": "Daniel J. Beutel", "role": "CEO", "handle": "daniel_janes", "bio": "\"Co-Founder & CEO @flwrlabs, AI Research @Cambridge_Uni\" (548)", "statements": [ { "said": "\"Public, centralized data is only a tiny fraction of all the data in the world.\"", "medium": "interview", "venue": "TechCrunch", "date": "2023-08-08", "url": "https://techcrunch.com/2023/08/08/flower-lands-3-6m-to-grow-its-platform-for-federated-learning/" } ], "recurring_words": [] }, { "name": "Nic Lane", "role": "co-founder and chief science officer; Cambridge", "handle": "niclane", "bio": "\"co-founder & CSO @flwrlabs (YCW23)\" (2,711)", "statements": [ { "said": "\"Competitive with leading models from OpenAI and Anthropic, while remaining available as a service or for private deployment.\"", "medium": "blog", "venue": "Endeavor 1.0, co-bylined", "date": "2026-09-01", "url": "https://flower.ai/blog/2026-09-01-introducing-endeavor-1.0" }, { "said": "\"Lucky the world has @flwrlabs to balance against this.\" Quoting Chamath on an \"attempted oligopoly on intelligence.\"", "medium": "X", "venue": "", "date": "2026-04-11", "url": "https://x.com/niclane/status/2042907020085338117" }, { "said": "\"Europe should not have to rent its intelligence indefinitely from a handful of US companies.\"", "medium": "interview", "venue": "Tech.eu", "date": "2026-09-01", "url": "https://tech.eu/2026/09/01/cambridge-university-spinout-launches-ai-model-competitive-with-openai-and-anthropic/" } ], "recurring_words": [ "collaborative superintelligence", "frontier-class", "private deployment", "their own", "sovereign" ] } ] }, "largest_model": { "value": "**Endeavor 1.0**, 1 September 2026, 'a frontier-class generalist for reasoning, coding, and long-horizon agent work.' Built on the open-weight ecosystem by continual pretraining and post-training. Parameter count and base model not published; benchmarks self-reported.", "label": "stated", "note": "The largest run outside one reserved cluster is Photon at 7B." } }, { "id": "gensyn", "name": "Gensyn", "kicker": "Verde · open-1b · Delphi", "url": "https://www.gensyn.ai/", "relation": "reference", "secondary_relation": null, "last_verified": "2026-09-29", "people": [ { "name": "Ben Fielding", "role": "co-founder and CEO", "handle": "benfielding", "said": "April 2026, on Delphi: settlement \"performed by verifiable intelligent oracles, not insiders.\" (Gensyn's launch release credits the line to them.)" }, { "name": "Harry Grieve", "role": "co-founder and CTO", "handle": "harrygrieve" }, { "name": "Rafael Frongillo", "role": "head of research, announced 23 September 2026; prediction markets and information elicitation" } ], "what_it_is": { "value": "Once \"HTTP for machine learning compute\" (co-founders, Epicenter interview); today the homepage says \"machines that predict the future\" and the live product is Delphi, an AI-settled information market. Four pivots on the public record: ML compute protocol (2022 to 2024), RL Swarm testnet (March 2025), personal-model apps BlockAssist and CodeAssist (2025, now sunset), Delphi (testnet December 2025, mainnet April 2026).", "label": "inferred" }, "mechanism": { "summary": { "value": "No live distributed training. RL Swarm (data-parallel, each node its own small model) has run no official swarm since January 2026 and is marked paused. Verde and open-1b make every training step bitwise auditable; open-1b ran on 48 H100s across six Google Cloud nodes.", "label": "verified" }, "detail": { "value": "**Verification is the work**: Verde (refereed delegation on bitwise-reproducible operators, RepOps) and open-1b (a chained state hash for each of 80,957 steps over weights, optimizer state, gradients and batch, replayable on one machine by auditors). The reproducible runtime cost about 5% MFU. RL Swarm, its permissionless testnet, held a full small model per node; the docs said \"no official swarms running\" by 19 January 2026 and mark it **paused**. No Gensyn-run swarm can be joined and no GPU work is paid; the open-1b audit is unpaid.", "label": "verified" } }, "largest_run": { "parameters_b": 1.61, "summary": { "value": "**open-1b**, 1.61B, 400B tokens, 48 H100s on six Google Cloud nodes, 27.8 days of training, about 5% MFU, released 15 September 2026. Weights and every step's hash public. Run centrally, not a distributed run.", "label": "verified" }, "detail": { "value": "**open-1b**: 1.61B, 400B tokens, 48 H100 on six a3-megagpu-8g nodes on Google Cloud (one cluster, inferred; the paper names no site), 27.8 active days over a 29.5-day span, about 5% MFU (\"roughly 5x slower than an optimized PyTorch stack\"), every step hashed, weights Apache 2.0. A verifiability demonstration, run \"in a centralized way.\" RL Swarm (March 2025 to about January 2026) trained 0.5B to 1.5B models per node.", "label": "verified" }, "shared_run": false }, "sold": { "value": "No training run can be bought; the open-1b harness is \"not a general-purpose training framework.\" Delphi sells market creation and trading in USDC: creators take 1.5% of volume, and a 0.5% protocol fee buys $AI on Uniswap, of which 70% is burned, 29% goes to the Community Treasury and 1% to the caller. Buybacks are live on chain (latest 17 September 2026). Nothing on Delphi requires $AI. REE (SDK MIT, compiler and operator binaries proprietary) and AXL are free tooling.", "label": "verified" }, "paid": { "value": "No training customer public. Delphi, 2 September 2026: $250,426 cumulative volume, $76,980 verifiably settled, 26,843 trades by 2,809 traders, 600-plus markets.", "label": "stated" }, "funding": { "value": "About $78M raised since 2020 (The Block, April 2026): seed and pre-seed of $7M-plus, $43M Series A led by a16z (June 2023), $16.7M led by a16z crypto at a $1B fully diluted valuation (October 2025), and an $11.7M $AI token sale (December 2025). $AI distributed April 2026. London.", "label": "secondhand" }, "relation_argument": "The reference for verification in this field. Verde and RepOps made an ML step reproduce bit for bit across hardware, and open-1b published a chained hash for each of its 80,957 steps so anyone can replay one. Gensyn has run no official training swarm since January 2026; its live product is Delphi, a prediction market settled by AI, and its new head of research works on prediction markets. open-1b trained on six Google Cloud H100 nodes at about 5% MFU, not on compute outside a single reserved cluster. Its decentralized-training repositories have had no commits since April 2026, and the docs offer only \"check back later,\" with no date. Nothing it sells today competes for IOTA's buyer or its miners, so it carries no secondary relation until a training network returns with a date.", "sources": [ { "title": "Delphi launch post, 22 April 2026", "url": "https://www.gensyn.ai/news/delphi" }, { "title": "Prediction-market research hires, 23 September 2026", "url": "https://www.gensyn.ai/news/prediction-market-research-frongillo-waggoner" }, { "title": "Testnet docs with the pause notice, read 29 September 2026", "url": "https://docs.gensyn.ai/testnet" }, { "title": "RL Swarm docs, archived 19 January 2026: no official swarms running", "url": "https://web.archive.org/web/20260119091246/https://docs.gensyn.ai/testnet/rl-swarm" }, { "title": "open-1b announcement, 15 September 2026", "url": "https://www.gensyn.ai/news/introducing-open-1b-auditable-training" }, { "title": "open-1b paper, arXiv 2609.17380 (section 5: six GCP nodes, about 5% MFU)", "url": "https://arxiv.org/abs/2609.17380" }, { "title": "open-transformers: open-1b pretraining and audit-replay harness", "url": "https://github.com/gensyn-ai/open-transformers" }, { "title": "open-1b weights on Hugging Face (Apache 2.0)", "url": "https://huggingface.co/collections/Gensyn/open-1b" }, { "title": "Verde, arXiv 2502.19405", "url": "https://arxiv.org/abs/2502.19405" }, { "title": "REE repository and licenses", "url": "https://github.com/gensyn-ai/ree" }, { "title": "$AI token docs: buyback and 70/29/1 split", "url": "https://docs.gensyn.network/ai-token" }, { "title": "$AI BuyBack Vault on the Gensyn explorer", "url": "https://gensyn-mainnet.explorer.alchemy.com/address/0x2CBEE00F91A2BC50a7D5C53DFfa6BAB79d7E0243" }, { "title": "Delphi volume post, 2 September 2026", "url": "https://x.com/gensynai/status/2095196623877849458" }, { "title": "The Block on Delphi, fees and funding, 22 April 2026", "url": "https://www.theblock.co/news/business/2026-04-22-a16z-crypto-gensyn-delphi-ai-settled-information-markets-platform-398419" }, { "title": "Gensyn mainnet explorer stats, read 29 September 2026", "url": "https://gensyn-mainnet.explorer.alchemy.com/api/v2/stats" }, { "title": "gensyn-ai repositories (GitHub API), read 29 September 2026", "url": "https://github.com/orgs/gensyn-ai/repositories?sort=updated" } ], "signals": { "url": "https://www.gensyn.ai/", "read": "2026-09-29", "js_rendered": false, "title": "Gensyn | machines that predict the future", "h1": "machines that predict the future", "first_h2": "Recent news", "meta_description": "Gensyn builds AI that forecasts – and keeps improving – by verifiably closing the loop between prediction and reality.", "og_description": "Gensyn builds AI that forecasts – and keeps improving – by verifiably closing the loop between prediction and reality.", "we_are_sentence": "Gensyn builds AI that forecasts – and keeps improving – by verifiably closing the loop between prediction and reality.", "proof_numbers": [], "audience_named": [ "unknown" ], "cta_verb": "Delphi app", "contested_words_present": [ "decentralised", "open", "swarm" ], "token_symbols": [], "dates_on_page": [ "23 Sept 2026", "15 Sept 2026", "8 Sept 2026", "8 Jul 2026", "12 Jun 2026", "10 Jun 2026", "5 Jun 2026", "7 May 2026", "22 Apr 2026", "15 Apr 2026", "17 Mar 2026", "11 Dec 2025", "8 Dec 2025", "12 Nov 2025" ], "read_via": "re-read live on 2026-09-29; proof_numbers and dates corrected by hand" }, "message": { "category": "Gensyn builds AI that forecasts", "h1": "machines that predict the future", "promise": "AI that forecasts, and keeps improving, by verifiably closing the loop between prediction and reality", "proof": "unknown", "audience": "unknown", "cta": "Delphi app", "label": "verified" }, "voice": { "carries_label": "brand-led, founder amplification", "read": "2026-09-28", "carries": "The company account (93,021 followers, at least twenty posts in the two weeks to today) carries launches, event logistics and daily audit updates, and retweets the founders rather than the reverse: eight of its last twenty posts are retweets of Fielding, Grieve and staff. The founders post on launch days at comparable engagement (Grieve's open-1b post 500 likes against 271 for the company's), but the cadence is the brand's. Every 2026 press quote is Fielding's; Grieve writes the long-form X articles and his launch posts out-reach Fielding's on the same day. The research talks on YouTube are staff, not founders.", "message": "\"We can log and audit every single step of AI training and inference.\" Fielding, 15 September 2026. The company account's bio: \"machines that predict the future.\"", "accounts": [ { "channel": "@gensynai", "url": "https://x.com/gensynai", "note": "93,021 on 29 September; twenty-plus posts in two weeks" }, { "channel": "@Delphi_fyi", "url": "https://x.com/Delphi_fyi", "note": "1,232; \"Permissionless, verifiable, AI-settled.\"" }, { "channel": "@GensynFND", "url": "https://x.com/GensynFND", "note": "3,548" }, { "channel": "News", "url": "https://www.gensyn.ai/news", "note": "open-1b 15 September; research hires 23 September" }, { "channel": "Discord", "url": "https://discord.gg/gensyn", "note": "97,475 members on 29 September" }, { "channel": "YouTube", "url": "https://www.youtube.com/@gensynai", "note": "357 subscribers, 30 videos; mostly researcher talks; open-1b video 173 views" }, { "channel": "LinkedIn", "url": "https://www.linkedin.com/company/gensynai", "note": "8,972 (28 September, not rechecked)" } ], "founders": [ { "name": "Ben Fielding", "role": "CEO", "handle": "benfielding", "bio": "\"Co-founder & CEO @GensynAI - the network for machine intelligence. I like modular, composable, decentralised, and evolutionary machine learning\" (7,303)", "statements": [ { "said": "\"The crowd are auditing open-1b - proof that we don't need ordained auditors, we need tech that let's the public audit.\"", "medium": "X", "venue": "retweeted by the company", "date": "2026-09-18", "url": "https://x.com/benfielding/status/2100991976560881898" }, { "said": "\"We can log and audit every single step of AI training and inference, providing proof of its training data, recipe, biases, and weights.\"", "medium": "X", "venue": "open-1b", "date": "2026-09-15", "url": "https://x.com/benfielding/status/2099880497845592276" }, { "said": "\"We just hit one million models trained over decentralised infrastructure ... decentralised AI is getting pretty hard to deny at this point.\"", "medium": "X", "venue": "the testnet, about two months before official swarms stopped", "date": "2025-11-14", "url": "https://x.com/benfielding/status/1989129208270926083" }, { "said": "\"We're officially moving from research and infrastructure to live deployments and adoption.\"", "medium": "X", "venue": "on the a16z State of Crypto report", "date": "2025-10-22", "url": "https://x.com/benfielding/status/1981070366517973344" }, { "said": "\"In the future, humanity will rely on AI to settle disputes, contracts, markets, etc.\"", "medium": "X", "venue": "the Judge demo", "date": "2025-08-27", "url": "https://x.com/benfielding/status/1960722847124611177" }, { "said": "\"There are no whitelists, no centralised aggregation servers, no singular models in the middle.\"", "medium": "X", "venue": "a 72B RL run on the testnet", "date": "2025-04-30", "url": "https://x.com/benfielding/status/1917657159615029598" }, { "said": "On the Nous and Prime Intellect runs: \"next we need a way to establish trust over the execution to make it truly decentralised.\"", "medium": "X", "venue": "", "date": "2024-12-02", "url": "https://x.com/benfielding/status/1863642571299274754" }, { "said": "\"Reminder: Gensyn does not have a token, if you buy one you are being scammed.\" The sale opened eight months later.", "medium": "X", "venue": "", "date": "2025-04-14", "url": "https://x.com/benfielding/status/1911822199528534063" } ], "recurring_words": [ "decentralised AI", "no whitelists", "sovereign", "self-owned", "auditable", "the network for machine intelligence", "hard to deny" ] }, { "name": "Harry Grieve", "role": "co-founder and CTO", "handle": "harrygrieve", "bio": "\"Co-founder & CTO @gensynai\" (3,505; the older handle _grieve returns 404)", "statements": [ { "said": "\"Unlike models by other AI labs, which require that you trust their training process, open-1b is fully auditable and replayable.\"", "medium": "X", "venue": "open-1b", "date": "2026-09-15", "url": "https://x.com/harrygrieve/status/2099888587655299381" }, { "said": "\"Just 2 years ago, advocating for decentralised training strategies was non-consensus!\"", "medium": "X", "venue": "", "date": "2025-09-12", "url": "https://x.com/harrygrieve/status/1966504198926684279" }, { "said": "\"Our core technical primitives: Communication, Consistent Execution, and Verification.\"", "medium": "X", "venue": "", "date": "2025-02-26", "url": "https://x.com/harrygrieve/status/1894846257207878063" }, { "said": "\"Great work from @PluralisHQ - one transformer block per node, connected via a novel Hivemind fork.\"", "medium": "X", "venue": "on Pluralis", "date": "2025-09-17", "url": "https://x.com/harrygrieve/status/1968432520506011940" }, { "said": "\"I hope Gensyn's greatest contribution will be to become the economic foundation of a parallel machine civilization.\"", "medium": "interview", "venue": "ChainCatcher, via Bitget (translated)", "date": "2025-11-20", "url": "https://www.bitget.com/news/detail/12560605072983" } ], "recurring_words": [ "auditable", "replayable", "hard to build and hard to stop", "Communication, Consistent Execution, and Verification" ] } ] }, "largest_model": { "value": "**open-1b**, 1.61B, the same run: 48 H100s on six Google Cloud nodes, run centrally so that every step could be hashed and audited.", "label": "verified", "parameters_b": 1.61, "note": "No run outside one reserved cluster at model scale; RL Swarm trained 0.5B to 1.5B per node." } }, { "id": "lium", "name": "Lium", "kicker": "Datura · GPU rental", "netuid": 51, "url": "https://lium.io", "relation": "rival-supply", "secondary_relation": "complement", "last_verified": "2026-09-29", "people": [ { "name": "Fish", "role": "founder, Datura", "handle": "fish_datura", "said": "Launch post: \"Anyone can add compute. Anyone can rent it.\"" } ], "what_it_is": { "value": "\"A GPU rental marketplace\": providers list machines by the hour, renters get pods billed by the second, and multi-node clusters on one fabric (in the CLI and SDK since 15 September; web-only before).", "label": "verified" }, "mechanism": { "summary": { "value": "Compute subnet. Rented machines are paid from one pool, eligible idle machines from another, while Lium runs its own job on them; idle B300 still $6.40 a GPU-hour (a cut to $1.25 was closed unmerged). 8.1% of TAO-in (chain, 29 September), rank 2.", "label": "verified" }, "detail": { "value": "Not a training system: a marketplace whose validator pays three pools: **rented** machines 13% by revenue share, **idle** eligible 1- and 8-GPU machines by hourly anchor under a per-model cap, the rest burned. Idle nodes run a Lium Default Job (Dolphin inference or the Pearl fallback); a provider who runs their own job forfeits idle pay. The idle B300 rate is $6.40 (PR 1386); PR 1491's cut to $1.25 was closed unmerged on 28 September. 8.1% of TAO-in (chain, 29 September), rank 2.", "label": "verified" } }, "largest_run": { "parameters_b": null, "summary": { "value": "Not a training subnet. Fleet: 1,329 GPUs across 470 nodes (11 September), 4090 through B300; 124 pods and 370 GPUs free on 29 September.", "label": "secondhand" }, "detail": { "value": "No run. **Fleet**: 1,329 GPUs across 470 nodes as of 11 September 2026 and 21 GPU models rented in August (PC Tech, secondhand); roughly 500 H100s in May (third-party review, secondhand). 124 pods with 370 free GPUs on 29 September (public node API, verified). Utilization: unknown.", "label": "secondhand" } }, "sold": { "value": "Hourly pods, asks on 29 September: H200 $3.00, B200 $5.35 to $5.60, B300 $8.50 to $12.95, A100 $1.06, RTX 4090 $0.35 to $0.85, RTX 5090 $0.50 to $1.00 a GPU-hour; spot; clusters. Providers keep 95%.", "label": "verified" }, "paid": { "value": "900 paying accounts and 7,549 rentals in August 2026 (PC Tech); about $432k a month in mid-May (third-party review). Both secondhand; Lium publishes no revenue figure.", "label": "secondhand" }, "funding": { "value": "Emission-funded subnet; no round public. Datura AI Corp.", "label": "unknown" }, "relation_argument": "Rival-supply: Lium pays idle 4090s, 5090s, A100s, L40S and A6000s from emission, the card classes in IOTA's fleets, and a provider who runs their own job on an idle node forfeits that pay (rental income is kept). Complement: Lium already runs outside paid work (Dolphin inference) on idle nodes, choosing the job that earns most per GPU-hour, and its pods can be rented by IOTA miners.", "sources": [ { "title": "lium.io pricing, live asks, read 29 September 2026", "url": "https://lium.io/pricing" }, { "title": "Public node list, read 29 September 2026", "url": "https://lium.io/api/public/v1/nodes" }, { "title": "Default Jobs: own job forfeits idle pay; Dolphin and Pearl", "url": "https://docs.lium.io/providers/portal/default-jobs.md" }, { "title": "Emission pools and eligibility", "url": "https://docs.lium.io/providers/rewards/emission.md" }, { "title": "Idle B300 at $6.40, PR 1386, merged 16 September 2026", "url": "https://github.com/Datura-ai/lium-io/pull/1386" }, { "title": "Idle B300 cut to $1.25, PR 1491, closed unmerged 28 September 2026", "url": "https://github.com/Datura-ai/lium-io/pull/1491" }, { "title": "Idle-pay GPU list and caps, validator config", "url": "https://github.com/Datura-ai/lium-io/blob/main/neurons/validators/src/incentive/config.py" }, { "title": "Clusters in the CLI and SDK, PR 152, 15 September 2026", "url": "https://github.com/Datura-ai/lium/pull/152" }, { "title": "Fleet and rental counts, PC Tech, 14 September 2026 (secondhand)", "url": "https://pctechmag.com/2026/09/lium-is-changing-the-gpu-rental-market-by-raising-the-bar-on-quality/" }, { "title": "May revenue estimate, Own Your Mind (secondhand)", "url": "https://ownyourmind.ai/tokenomics/lium-bittensor-subsidy-ratio/" }, { "title": "legible.network record, SN51", "url": "https://legible.network/sn/51/" }, { "title": "taostats SN51", "url": "https://taostats.io/subnets/51" } ], "signals": { "url": "https://lium.io", "read": "2026-09-28", "js_rendered": false, "title": "Rent High-Performance GPUs at Unmatched Prices | Agent-First Compute | lium.io", "h1": "HIGH-PERFORMANCE GPU PODS", "first_h2": "unknown", "meta_description": "Agent-first compute on Lium: rent H100, B200, RTX 4090 and more GPU pods at marketplace prices. Deploy in seconds via browser or CLI. Billed per second.", "og_description": "Agent-first compute on Lium: rent H100, B200, RTX 4090 and more GPU pods at marketplace prices. Deploy in seconds via browser or CLI. Billed per second.", "we_are_sentence": "unknown", "proof_numbers": [ "g Toggle Filters Showing 120 / Available Pods Sort Download 1 X NVIDIA RTX 6000 Ada Generation $0", "7 /HOUR RENT NOW CPU 8 X INTEL(R) XEON(R) GOLD 6548Y+ Memory 63 GB VRAM 48 GB Hard d", "D (hotkey) 5D4o4KuXymtiPbD7fxPyRGaHm4TS7QEL2thx8vnN257jL5Me 1 X NVIDIA RTX 6000 Ada Generation $0", "D (hotkey) 5D4o4KuXymtiPbD7fxPyRGaHm4TS7QEL2thx8vnN257jL5Me 8 X NVIDIA B200 SPOT $44", "55 /GPU RENT NOW CPU 160 X INTEL(R) XEON(R) PLATINUM 8570 Memory 1,764 GB VRAM 1,432 G", "D (hotkey) 5FqRyjFhdBzPAARZTdgi1WyqPBEGJyPJRsrXZN9CStRdN3s8 8 X NVIDIA GeForce RTX 4090 $6", "85 /GPU RENT NOW CPU 256 X AMD EPYC 7B13 64-CORE PROCESSOR Memory 504 GB VRAM 192 GB H", "D (hotkey) 5CMUxZ9EQpRsxtuFE3f8vcUoXheZMWnHex4WQ2dJCzxwMHaM 8 X NVIDIA H100 80GB HBM3 $19", "45 /GPU RENT NOW CPU 208 X INTEL(R) XEON(R) PLATINUM 8480+ Memory 1,772 GB VRAM 637 GB", "D (hotkey) 5HKg6WXycb5RVfTJXhYPZA1aChSMNZSVboheozWdUfmvuz19 8 X NVIDIA GeForce RTX 5090 $6", "85 /GPU RENT NOW CPU 256 X AMD EPYC 7B13 64-CORE PROCESSOR Memory 504 GB VRAM 255 GB H", "43 /GPU RENT NOW CPU 208 X INTEL(R) XEON(R) PLATINUM 8480+ Memory 1,772 GB VRAM 637 GB" ], "audience_named": [ "unknown" ], "cta_verb": "Rent now; Get started for free", "contested_words_present": [ "open" ], "token_symbols": [], "dates_on_page": [ "unknown" ], "read_via": "checked by hand against the 28 September capture on 2026-09-29", "site_kind": "app", "what_it_shows": "a list of GPU pods with the card, the price per hour and per GPU, the location, uptime and reliability, a pip command to spin one up, and filters by GPU type; no landing page in front of it", "spoken": [ { "text": "HIGH-PERFORMANCE GPU PODS", "source": "the app's header, read 28 September 2026", "url": "https://lium.io", "label": "verified" }, { "text": "pip install lium.io && lium up --gpu A100 # spin up a pod from your terminal", "source": "the app's header, read 28 September 2026", "url": "https://lium.io", "label": "verified" } ] }, "message": { "category": "Agent-first compute", "h1": "HIGH-PERFORMANCE GPU PODS", "promise": "rent H100, B200, RTX 4090 and more GPU pods at marketplace prices. Deploy in seconds via browser or CLI. Billed per second.", "proof": "Showing 120 Available Pods", "audience": "unknown", "cta": "Rent now; Get started for free", "label": "verified" }, "voice": { "carries_label": "founder-led in private, brand-led in public", "read": "2026-09-29", "carries": "Fish sets pricing and product policy by Discord directive: seven dated directives in September alone, each turned into a pull request within a day, though several (PR 1401, 1453, 1491) closed unmerged. The public voice is the product account and third-party explainers; the founder's own account has 4,097 followers and a two-word bio. The clearest positioning statement of the year, the order on 22 September to strip Bittensor from renter-facing docs, was issued in Discord, not on stage. No podcast or conference appearance by Fish was found.", "message": "\"Anyone can add compute. Anyone can rent it. No KYC. Pay in crypto.\" Fish, the launch post, 31 July 2025.", "accounts": [ { "channel": "@lium_io", "url": "https://x.com/lium_io", "note": "5,149; bio \"High-Performance GPUs | Unmatched Pricing\"" }, { "channel": "GitHub", "url": "https://github.com/Datura-ai/lium-io", "note": "several pull requests a day; founder directives quoted in the bodies" }, { "channel": "Discord", "url": "https://discord.gg/lium", "note": "where the directives are issued" }, { "channel": "Site", "url": "https://lium.io", "note": "no blog" } ], "founders": [ { "name": "Fish", "role": "founder, Datura", "handle": "fish_datura", "bio": "\"For Bittensor.\" (4,097)", "statements": [ { "said": "Directive: \"remove Bittensor from the RENTER and GENERAL docs.\"", "medium": "Discord, quoted in a pull request", "venue": "Datura-ai/lium PR 281", "date": "2026-09-22", "url": "https://github.com/Datura-ai/lium/pull/281" }, { "said": "Directive: \"We can't pay them and not use the gpu.\"", "medium": "Discord, quoted in a pull request", "venue": "lium-io PR 1453", "date": "2026-09-18", "url": "https://github.com/Datura-ai/lium-io/pull/1453" }, { "said": "Directive: \"ensure idle incentive is not more than rental rates.\"", "medium": "Discord, quoted in a pull request", "venue": "lium-io PR 1401", "date": "2026-09-17", "url": "https://github.com/Datura-ai/lium-io/pull/1401" }, { "said": "\"Anyone can add compute. Anyone can rent it. No KYC. Pay in crypto. Spin up machines in 10 seconds.\"", "medium": "X", "venue": "the Lium launch", "date": "2025-07-31", "url": "https://x.com/fish_datura/status/1950868697372823804" } ], "recurring_words": [ "compute", "rent", "idle", "incentive", "no KYC", "provider", "slots" ] } ] }, "largest_model": { "value": "No model of its own. Sells GPU hours and multi-node clusters; see the run field.", "label": "verified" } }, { "id": "targon", "name": "Targon", "kicker": "Manifold Labs · confidential compute", "netuid": 4, "url": "https://targon.com", "relation": "complement", "secondary_relation": "rival-buyer", "last_verified": "2026-09-29", "people": [ { "name": "Robert Myers", "role": "founder and CEO, Manifold Labs" }, { "name": "James Woodman", "role": "co-founder; formerly COO, Opentensor" } ], "what_it_is": { "value": "Attested GPU compute inside Intel TDX confidential VMs, plus bare metal (14 September) and sandboxes (23 September), sold \"for AI Training and Deployments.\" Inference now sells as Sybil; the old Targon API returns 410.", "label": "verified" }, "mechanism": { "summary": { "value": "Compute subnet. Miners install a signed OS on 8-GPU Intel servers, attest every boot, keep cards live; paid per verified card in an auction class whether or not a customer uses it. 4.5% of TAO-in (chain, 29 September), rank 7.", "label": "verified" }, "detail": { "value": "Not a training system. A miner installs TargonOS, which wipes every disk and attests the machine on every boot; install profiles are 8x H100, H200, B200, B300 or RTX PRO 6000 Blackwell on Intel. Each auction names target cards and a price cap, and every verified card earns its share **whether or not a customer is using it**; the validator code has no rental term. Weight no auction covers goes to burn. 4.5% of TAO-in (chain, 29 September), rank 7.", "label": "verified" } }, "largest_run": { "parameters_b": null, "summary": { "value": "Not a training subnet. Fleet: 224 GPUs live in the emission auctions on 29 September (112 H200); \"1,500+ H200s\" was the Series A claim (July 2025). Utilization unknown.", "label": "verified" }, "detail": { "value": "No run. **Fleet**: 224 GPUs live in the auctions on 29 September 2026 (72 B300, 112 H200, 24 H100, 16 RTX PRO 6000; verified). \"1,500+ H200s\" stated at the Series A, 28 July 2025; the \"1,000+\" figure appears only in third-party profiles. Machines outside the auctions: unknown. The homepage utilization tile renders 0. On 5 June 2026 Manifold wrote that \"Macrocosmos trained a 100-billion-parameter model across 48 distributed GPUs over regular commodity internet.\"", "label": "verified" } }, "sold": { "value": "Confidential VMs and sandboxes by consumption: H200 $3.59, B200 $5.35, H100 $3.09, RTX PRO 6000 $1.69 a GPU-hour; B300 $6.50 on manifold.inc, $8.50 on targon.com; H200, B200, B300 out of stock on 29 September. Bare metal: 4x RTX 4090, Austin, $2.00 an hour. Inference via Sybil.", "label": "verified" }, "paid": { "value": "\"Roughly $10 million in annualized external revenue\" (Manifold, 5 June 2026). No named enterprise customer; Bittensor-internal users Dippy, Ridges, Score (secondhand).", "label": "stated" }, "funding": { "value": "$10.5M Series A, 28 July 2025, led by OSS Capital with DCG, Tobi Lutke and others. Austin (secondhand).", "label": "stated" }, "relation_argument": "Complement: Targon sells capacity IOTA or its miners could rent (H100 VM $3.09, RTX PRO 6000 $1.69 a GPU-hour; 4x 4090 bare metal $2.00 an hour). Not rival-supply: its miners run 8-GPU Intel servers under a locked OS, none of the card classes in IOTA's fleets, though it pays per verified card, used or not. Rival-buyer, weak: a builder could rent Targon hours for training, which is raw GPU time with no training runtime.", "sources": [ { "title": "Manifold, \"Capital Finds the Machine\", 5 June 2026", "url": "https://manifoldlabs.substack.com/p/capital-finds-the-machine" }, { "title": "Manifold releases (bare metal, sandboxes, TargonOS)", "url": "https://www.manifold.inc/releases" }, { "title": "TargonOS announcement, consumer tier planned, about March 2026", "url": "https://www.manifold.inc/releases/targon-os" }, { "title": "Targon inventory and prices, read 29 September 2026", "url": "https://targon.com/inventory" }, { "title": "Auction targets and live cards, read 29 September 2026", "url": "https://stats.targon.com/targets" }, { "title": "Miner docs: install profiles and per-card pay", "url": "https://github.com/manifold-inc/targon/blob/main/docs/miner/miner.md" }, { "title": "Validator weights: pay per verified card", "url": "https://github.com/manifold-inc/targon/blob/main/internal/validator/callbacks/weights.go" }, { "title": "OSS Capital on the Series A, 28 July 2025", "url": "https://oss.capital/oss-capital-leads-10-5m-series-a-in-manifold-labs-alongside-industry-legends/" }, { "title": "legible.network record, SN4", "url": "https://legible.network/sn/4/" }, { "title": "taostats SN4", "url": "https://taostats.io/subnets/4" } ], "signals": { "url": "https://targon.com", "read": "2026-09-28", "js_rendered": false, "title": "Targon", "h1": "Confidential Compute Without Compromise", "first_h2": "Connect With Developers Powering Targon", "meta_description": "Scale with Secure GPU & CPU Rentals on a Lightning-Fast Cloud for Training and Deployment", "og_description": "Scale with Secure GPU & CPU Rentals on a Lightning-Fast Cloud for Training and Deployment", "we_are_sentence": "Targon is a decentralized compute network of trusted execution environments.", "proof_numbers": [ "0 + GPUs & CPUs 0 % Uptime < 0 ms Latency 0 % Utilization Targon Vir" ], "audience_named": [ "unknown" ], "cta_verb": "Docs", "contested_words_present": [ "decentralized", "distributed" ], "token_symbols": [], "dates_on_page": [ "unknown" ], "read_via": "checked by hand against the 28 September capture on 2026-09-29; tiles still 0" }, "message": { "category": "a decentralized compute network of trusted execution environments", "h1": "Confidential Compute Without Compromise", "promise": "Scale with Secure GPU & CPU Rentals on a Lightning-Fast Cloud for Training and Deployment", "proof": "unknown", "audience": "unknown", "cta": "Docs", "label": "verified", "note": "The stat tiles for GPUs, uptime and utilization still render 0 (29 September)." }, "voice": { "carries_label": "brand-led, two named founders behind it", "read": "2026-09-29", "carries": "Every written channel is in the company's voice with no byline: a weekly Substack from February to June (sixteen posts, none since 19 June), the releases page, both X accounts, the Intel whitepaper. Myers carries the message in person at conferences (Proof of Talk in June, the Exploit Summit keynote on 28 September); Woodman carried it on podcasts and X in 2025. No 2026 verbatim quote from either founder was found.", "message": "\"A subnet's price is a payroll, not a scoreboard.\" Company voice, 5 June 2026. Woodman, 2025: \"Targon is playing a different game than Chutes.\"", "accounts": [ { "channel": "@manifoldlabs", "url": "https://x.com/manifoldlabs", "note": "5,292; \"A Decentralized Frontier AI Lab building @TargonCompute & @SybilChat\"" }, { "channel": "@TargonCompute", "url": "https://x.com/TargonCompute", "note": "5,700; \"The Secure Compute Cloud powering AI's best builders\"" }, { "channel": "Substack", "url": "https://manifoldlabs.substack.com", "note": "The Subnet Signal: weekly 13 February to 19 June, then stopped" }, { "channel": "Releases", "url": "https://www.manifold.inc/releases", "note": "Bare Metal and Sandboxes in September; unbylined" } ], "founders": [ { "name": "Robert Myers", "role": "founder and CEO; Bittensor's first miner", "handle": null, "bio": "No X handle found. Exploit Summit keynote, 28 September: \"Targon: From Confidential Compute to Full-Stack Cloud.\"", "statements": [ { "said": "Company voice: \"Roughly $10 million in annualized external revenue.\" And on IOTA: \"Macrocosmos trained a 100-billion-parameter model across 48 distributed GPUs over regular commodity internet.\"", "medium": "Substack", "venue": "The Subnet Signal, unbylined", "date": "2026-06-05", "url": "https://manifoldlabs.substack.com/p/capital-finds-the-machine" } ], "recurring_words": [ "confidential compute", "TEE", "Intel TDX", "enterprise", "full-stack cloud", "attestation" ] }, { "name": "James Woodman", "role": "co-founder; formerly COO, Opentensor", "handle": "jameswoodmanv", "bio": "\"co-founder @manifoldlabs | prev coo @opentensor\" (2,955)", "statements": [ { "said": "\"I think we have six months, maybe 12 where if you don't get organic demand, it's done.\"", "medium": "podcast", "venue": "Ventura Labs Ep. 53", "date": "2025-07-28", "url": "https://www.youtube.com/watch?v=6I6AlTuV1rA" }, { "said": "\"Do you believe that all compute hours are equal?\" \"The moat that we really believe we can have is through the exchange, the liquidity.\"", "medium": "podcast", "venue": "Ventura Labs Ep. 53", "date": "2025-07-28", "url": "https://www.youtube.com/watch?v=6I6AlTuV1rA" }, { "said": "\"Targon is playing a different game than Chutes, and it is certainly playing a different game than Nineteen.\"", "medium": "X", "venue": "", "date": "2025-03-23", "url": "https://x.com/jameswoodmanv/status/1903844853299163436" } ], "recurring_words": [ "liquidity", "exchange", "organic demand", "compute marketplace", "digital commodities" ] } ] }, "largest_model": { "value": "No model of its own. Sells GPU hours; its inference product, Sybil, serves other labs' open models.", "label": "verified" } }, { "id": "chutes", "name": "Chutes", "kicker": "Rayon Labs · inference, training research", "netuid": 64, "url": "https://chutes.ai", "relation": "rival-supply", "secondary_relation": "reference", "last_verified": "2026-09-29", "people": [ { "name": "Jon Durbin", "role": "core contributor and backend lead, Rayon Labs", "handle": "jon_durbin", "said": "28 April 2026: \"< $10/hr to pre-train a 20b MoE. Not too bad.\" Eight single-L40S VMs in Poland, France and the USA, \"with a few extra 4090s and a 5090\"." }, { "name": "Timon Agar", "role": "credited on the July training post, engineering and product team" } ], "what_it_is": { "value": "Serverless inference for open models, paid per token, on miner hosts inside Intel TDX confidential VMs: H200, B200 and RTX Pro 6000; TEE is required for every deployment since 12 May 2026. Separately, Parallax: its own pretraining runs on rented RTX 5090 hosts.", "label": "verified" }, "mechanism": { "summary": { "value": "Inference subnet: a miner's score is the seconds its instances were deployed, busy or idle, times a compute multiplier, over one day. Separately, Parallax trains an 8B MoE on rented 8x RTX 5090 hosts, not SN64 miners. Emission 6.0% (taomarketcap, block 9171075, 29 September 2026).", "label": "verified" }, "detail": { "value": "Not a training subnet: score is instance lifetime times a multiplier for GPU class, bounty, urgency and TEE, with startup paid at 0.3x, over a one-day window; an idle deployed instance still earns. Every chute runs in an attested TDX VM since 12 May 2026. **Training** lives outside the subnet in **Parallax**: a July post trained a recurrent model \"within 0.6% of the centralized baseline at matched steps\" and called it \"an in-progress research direction rather than a finished system\"; nine production runs followed in September on rented RTX 5090 hosts. Emission 6.0% (taomarketcap, block 9171075, 29 September 2026).", "label": "verified" } }, "largest_run": { "parameters_b": 7.8, "summary": { "value": "**Parallax 8B alpha**, 7.8B MoE (1.2B active), 574B tokens on 240 rented RTX 5090s in 30 hosts across 13 countries, 10 to 12 September 2026. Weights public for a later run.", "label": "verified" }, "detail": { "value": "**Parallax**: a 7.8B ternary MoE, 1.2B active. The largest run by tokens, 8B alpha P1, trained 574B tokens on 30 hosts of 8x RTX 5090 (240 GPUs) in 13 countries from 10 to 12 September 2026, mean MFU 51%. The largest by GPUs, 8B delta, used 32 hosts (256 GPUs) on 18 to 19 September. Hosts are rented (\"$0.65 per GPU-hour, relays included\") and sync through CPU relays over the public internet. The tech report, written by AI from logs, says it scores below a 1.8B dense model on six benchmarks. **Inference fleet**: count unpublished; active H200s cut in March 2026, revenue per GPU $4.05 to $5.89.", "label": "verified", "note": "Run figures read from the live register API on 29 September 2026; the rental price and benchmark comparison are stated in the tech report. Revenue per GPU is stated, unit not given." } }, "sold": { "value": "Per-token inference (Kimi K2.6 $0.50 in, $2.85 out per million), Plus $10 and Pro $20 a month, private TEE deployments on RTX Pro 6000 at $1.80 an hour plus a $5.40 setup fee. Parallax is not sold.", "label": "verified" }, "paid": { "value": "OpenRouter about 8.1B tokens a day, May to August 2026 (Own Your Mind tracking). Chutes states revenue per trillion tokens near $300,000 on 7-day averages (May 2026); no absolute revenue.", "label": "secondhand" }, "funding": { "value": "Emission-funded; revenue stated to go to alpha buybacks. Rayon also runs Gradients. Pseudonymous; the founders post names twelve contributors and says it is hiring.", "label": "stated" }, "relation_argument": "Rival for supply through Parallax: its runs rent hosts of 8x RTX 5090, up to 256 GPUs in one September run, the card class in IOTA's Orion-16B fleet, and a rented box cannot mine SN9 at the same time. Chutes' inference fleet does not overlap: every deployment runs in a confidential VM on H200, B200 or RTX Pro 6000 hosts, paid while the instance is up, busy or idle. Parallax is also the nearest published comparison to IOTA's register, and it is not sold.", "sources": [ { "title": "Parallax live register, read 29 September 2026", "url": "https://parallax.chutes.ai/" }, { "title": "Parallax tech report, version of 28 September 2026", "url": "https://parallax.chutes.ai/tech-report.pdf" }, { "title": "Parallax 8B theta weights on Hugging Face", "url": "https://huggingface.co/chutesai/parallax-8b-theta" }, { "title": "Non-blocking recurrent training, 8 July 2026", "url": "https://chutes.ai/news/non-blocking-recurrent-training" }, { "title": "Jon Durbin on X, 28 April 2026", "url": "https://x.com/jon_durbin/status/2049260278244675914" }, { "title": "From Volume to Value, 20 March 2026 (fleet and revenue per GPU)", "url": "https://chutes.ai/news/from-volume-to-value-building-a-sustainable-ai-inference-platform-2" }, { "title": "chutes-api (scoring, TEE requirement), commit 3b5609f, 3 September 2026", "url": "https://github.com/chutesai/chutes-api" }, { "title": "Pricing, read 29 September 2026", "url": "https://chutes.ai/pricing" }, { "title": "Own Your Mind on SN64 (OpenRouter volume), updated 24 September 2026", "url": "https://ownyourmind.ai/tokenomics/chutes-bittensor-revenue-machine/" }, { "title": "legible.network record, SN64", "url": "https://legible.network/sn/64/" }, { "title": "taomarketcap SN64 (emission), block 9171075", "url": "https://api.taomarketcap.com/public/v1/subnets/64/" } ], "signals": { "url": "https://chutes.ai", "read": "2026-09-29", "js_rendered": false, "title": "Chutes | Serverless AI Compute", "h1": "Breakthrough Serverless Compute for AI, at Scale.", "first_h2": "SOTA Open-Source LLMs, Available here first.", "meta_description": "Deploy, run and scale any AI model in seconds. Try directly through our platform, or use our easy-to-use API in seconds.", "og_description": "Deploy, run and scale any AI model in seconds. Try directly through our platform, or use our easy-to-use API in seconds.", "we_are_sentence": "Powering Trillions of Tokens per Month, Chutes is the leading open-source, decentralized compute provider for deploying, scaling and running open-source models in production.", "proof_numbers": [ "Plus $10 per month 5X the value of pay-as-you-go 6% off PAYG pricing PAYG request", "limit View limits Get Started Best Value Pro $20 per month 5X the value of pay-as-you-go 10% off PAYG pricing PAYG reques", "Plus $10 per month 5X the value of pay-as-you-go 6% off PAY", "yond limit View limits Get Started Best Value Pro $20 per month 5X the value of pay-as-you-go 10% off PA", "chutes TEE Hot LLM Pricing $0.12 in / $0", "ntelligence per dollar chutes TEE Hot LLM Pricing $0.30 in / $2", "chutes TEE Hot LLM Pricing $1.00 in / $1", "chutes TEE Hot LLM Pricing $0.24 in / $2", "chutes TEE Hot LLM Pricing $0.98 in / $3", "chutes TEE Hot LLM Pricing $0.50 in / $2", "chutes TEE Hot LLM Pricing $0.45 in / $3", "chutes TEE Hot LLM Pricing $1.25 in / $3" ], "audience_named": [ "for Everyone" ], "cta_verb": "Explore Models; Create Account", "contested_words_present": [ "decentralized", "open" ], "token_symbols": [], "dates_on_page": [ "unknown" ], "read_via": "checked by hand against the 28 September capture on 2026-09-29" }, "message": { "category": "the leading open-source, decentralized compute provider", "h1": "Breakthrough Serverless Compute for AI, at Scale.", "promise": "Deploy, run and scale any AI model in seconds.", "proof": "Powering Trillions of Tokens per Month", "audience": "for Everyone", "cta": "Explore Models; Create Account", "label": "verified" }, "voice": { "carries_label": "pseudonymous collective, one named face", "read": "2026-09-29", "carries": "The company's own post says there is no CEO and lists twelve contributors. Jon Durbin appears on podcasts and gave the Exploit Summit keynote on 28 September, and is the only contributor most outlets can name, yet on air they say they are not the CEO. Most blog posts are unbylined and the X account speaks as the product. Parallax reports through a live register and a tech report the company says was written by AI from its logs. The company skipped Proof of Talk in June citing infrastructure work.", "message": "\"I'm not the CEO, I'm not the president, anything like that. I can't control what these other groups are doing.\" Durbin, 12 April 2026.", "accounts": [ { "channel": "@chutes_ai", "url": "https://x.com/chutes_ai", "note": "11,733; \"The most secure AI inference on earth.\"" }, { "channel": "News", "url": "https://chutes.ai/news", "note": "fifteen posts in a year; at least two bylines" }, { "channel": "Founders post", "url": "https://chutes.ai/news/who-are-the-founders-of-chutes-ai", "note": "\"There is no CEO or central decision maker\", 12 November 2025" }, { "channel": "Parallax register", "url": "https://parallax.chutes.ai/", "note": "live per-host training numbers; tech report and weights linked" } ], "founders": [ { "name": "Jon Durbin", "role": "core contributor, backend lead", "handle": "jon_durbin", "bio": "\"Human. Backend dev http://chutes.ai\" (7,221)", "statements": [ { "said": "Keynote title: \"DeAI Is Dead, Long Live DeAI.\"", "medium": "conference", "venue": "Exploit Summit, Montreal", "date": "2026-09-28", "url": "https://exploitsummit.com/" }, { "said": "\"< $10/hr to pre-train a 20b MoE. Not too bad.\"", "medium": "X", "venue": "a run on eight single-L40S VMs in Poland, France and the USA, plus a few 4090s and a 5090", "date": "2026-04-28", "url": "https://x.com/jon_durbin/status/2049260278244675914" }, { "said": "\"All of the revenue that we generate can go immediately into buying back the token.\" \"We only serve TEE models on OpenRouter now.\"", "medium": "podcast", "venue": "On Chain Ep. 7", "date": "2026-04-12", "url": "https://www.youtube.com/watch?v=hbdrRbtMOS0" }, { "said": "\"It was all permissionless miners, decentralized network, you know, everything just worked and it was magical.\"", "medium": "YouTube", "venue": "Novelty Search SN64", "date": "2026-02-06", "url": "https://www.youtube.com/watch?v=Cri8C-sfItg" }, { "said": "\"Targon is great, I've been a fan of the team and the work they are doing.\"", "medium": "X", "venue": "reply to Woodman", "date": "2025-03-23", "url": "https://x.com/jon_durbin/status/1903858704606277910" } ], "recurring_words": [ "TEE", "confidential", "OpenRouter", "permissionless", "miners", "revenue", "free" ] } ] }, "largest_model": { "value": "**Parallax 8B theta**: 7.8B MoE, 1.2B active, trained on FineWeb-Edu; weights public, MIT, tagged research checkpoint, 28 September 2026.", "label": "verified", "parameters_b": 7.8 } }, { "id": "gradients", "name": "Gradients", "kicker": "Rayon Labs · fine-tuning", "netuid": 56, "url": "https://gradients.io", "relation": "rival-buyer", "secondary_relation": "complement", "last_verified": "2026-09-29", "people": [ { "name": "WanderingWeights (Chris)", "role": "co-lead" }, { "name": "Besim", "role": "co-lead" } ], "what_it_is": { "value": "Fine-tuning as a product: text (instruct, DPO, GRPO) and image jobs by UI and API, \"pay for what you use\"; miners compete with training code in tournaments of three types (text, image, environment), each opening on Mondays.", "label": "verified" }, "mechanism": { "summary": { "value": "Training subnet, post-training only. Miners submit code; validator trainer nodes run it on 1 A100 up to 8 H100. Customer jobs run on Runpod by default since April 2026. Emission 1.33% (taomarketcap, block 9171075, 29 September 2026).", "label": "verified" }, "detail": { "value": "Miners supply **code, not compute**: the validator pulls a repository at an exact commit and runs it on validator-controlled trainer nodes, sized per task from 1 A100 to 8 H100. **Customer jobs** default to Runpod through dstack since 8 April 2026; on 22 September 2026 Runpod was added as an evaluation backend beside Basilica (SN39). Entry fees 0.4 to 0.7 TAO per tournament, burned; winners' code is republished. Emission 1.33% (taomarketcap, block 9171075, 29 September 2026).", "label": "verified" } }, "largest_run": { "parameters_b": null, "summary": { "value": "Fine-tunes up to 70B-class models per job; no pretraining. About $100 small, $250 at 20B, $500 at 70B (secondhand); another record cites $10 to $50 an hour.", "label": "secondhand" }, "detail": { "value": "No pretraining run. Jobs fine-tune open models up to the 70B class on single nodes of up to 8 H100. Price points conflict: $100, $250 and $500 by size (SimplyTao, undated) against $10 to $50 an hour for text and $5 for image (legible.network, from docs); the live price table does not load. \"Over 3,000 paying users, primarily hobbyists and smaller organizations\" (SimplyTao).", "label": "secondhand" } }, "sold": { "value": "Fine-tuning jobs by UI and API, pay per use; the price table is rendered by script and still did not load on 29 September.", "label": "verified" }, "paid": { "value": "None named. 3,000+ paying users, secondhand.", "label": "secondhand" }, "funding": { "value": "Emission-funded plus job fees; Rayon Labs per SimplyTao and the paper's affiliation; site footer reads Grads.", "label": "secondhand" }, "relation_argument": "Rival for the buyer: Gradients sells fine-tuning today, by UI and API, to people who do not own GPUs, and on 28 September Macrocosmos said the iota SDK now covers fine-tuning and RL as well as pretraining; the SDK page still reads \"coming soon\". Complement as a possible workload: Gradients' paid jobs run on Runpod as single-node jobs of up to 8 H100, which could run on IOTA only after a port to the SDK; no such run has been shown.", "sources": [ { "title": "G.O.D repository (trainers, Runpod default 8 April, evaluation backend 22 September, fees), commits to 28 September 2026", "url": "https://github.com/gradients-ai/G.O.D" }, { "title": "Tournament fees API, read 29 September 2026", "url": "https://api.gradients.io/tournament/fees" }, { "title": "gradients.io pricing, read 29 September 2026", "url": "https://www.gradients.io/pricing" }, { "title": "SimplyTao guide to SN56 (prices, users, leads)", "url": "https://simplytao.ai/blog/your-simple-guide-to-gradients-sn56" }, { "title": "Will Squires on the SDK's scope, 28 September 2026", "url": "https://x.com/WSquires/status/2104681373533679638" }, { "title": "iota SDK page, read 29 September 2026", "url": "https://iota.macrocosmos.ai/sdk" }, { "title": "legible.network record, SN56", "url": "https://legible.network/sn/56/" }, { "title": "taomarketcap SN56 (emission), block 9171075", "url": "https://api.taomarketcap.com/public/v1/subnets/56/" } ], "signals": { "url": "https://www.gradients.io/", "read": "2026-09-29", "js_rendered": false, "title": "Gradients | Anyone Can Train AI", "h1": "INTELLIGENCE, SIMPLIFIED.", "first_h2": "On-demand", "meta_description": "Anyone Can Train AI on Bittensor. AI Training, Decentralized.", "og_description": "Anyone Can Train AI on Bittensor. AI Training, Decentralized.", "we_are_sentence": "Gradients allows anyone in the world to train image & text models - intelligence, simplified.", "proof_numbers": [], "audience_named": [ "Anyone Can Train AI Platform Pricing News Team Sign In INTELLIGENCE, SI", "Anyone can Train on Bittensor" ], "cta_verb": "Train with Us", "contested_words_present": [ "open" ], "token_symbols": [], "dates_on_page": [ "unknown" ], "read_via": "checked by hand against the 28 September capture on 2026-09-29" }, "message": { "category": "Gradients allows anyone in the world to train image & text models", "h1": "INTELLIGENCE, SIMPLIFIED.", "promise": "Anyone Can Train AI on Bittensor. AI Training, Decentralized.", "proof": "unknown", "audience": "Anyone", "cta": "Train with Us", "label": "verified" }, "voice": { "carries_label": "mixed, quiet since 2025", "read": "2026-09-29", "carries": "The founder carried it in person in 2025 through a sole-author paper and two long podcasts, in first person, with a clear buyer thesis. In 2026 the only channels observed are the product account, a blog with no post since November 2025, and third-party guides that name Besim as co-lead without a quote from them. No 2026 first-person statement from either lead was found.", "message": "\"Real customers want one model. They don't want three models.\" Subia-Waud, Ventura Labs, 11 August 2025.", "accounts": [ { "channel": "@gradients_ai", "url": "https://x.com/gradients_ai", "note": "2,412; \"The world's best AutoML platform, powered by Subnet 56\"" }, { "channel": "News", "url": "https://www.gradients.io/news", "note": "six posts, November 2024 to November 2025; none in 2026" }, { "channel": "GitHub", "url": "https://github.com/gradients-ai/G.O.D", "note": "commits to 28 September" } ], "founders": [ { "name": "Christopher Subia-Waud (WanderingWeights)", "role": "founder and co-lead", "handle": null, "bio": "No X handle resolves. PhD in AI; the paper's sole author.", "statements": [ { "said": "\"Real customers want one model. They don't want three models.\" \"Chutes on compute, Gradients on training, Affine on RL.\"", "medium": "podcast", "venue": "Ventura Labs Ep. 57", "date": "2025-08-11", "url": "https://www.youtube.com/watch?v=qJyL3koTZqs" }, { "said": "\"I'd pick working in Bittensor and actually working on decentralized AI any day of the week because it's real.\"", "medium": "podcast", "venue": "Ventura Labs Ep. 57", "date": "2025-08-11", "url": "https://www.youtube.com/watch?v=qJyL3koTZqs" }, { "said": "\"When miners compete for rewards, they develop optimization strategies that centralized approaches overlook.\"", "medium": "paper", "venue": "arXiv 2506.07940, Gradients: When Markets Meet Fine-tuning", "date": "2025-06-09", "url": "https://arxiv.org/abs/2506.07940" } ], "recurring_words": [ "miners compete", "real customers", "post-training", "AutoML", "tournament", "enterprise", "it's real" ] }, { "name": "Besim", "role": "co-lead", "handle": null, "bio": "Named as co-lead by SimplyTao, March 2026. No quote found.", "statements": [], "recurring_words": [] } ] }, "largest_model": { "value": "No model of its own. Sells compute or fine-tuning; see the run field.", "label": "verified" } }, { "id": "teutonic", "name": "Teutonic", "kicker": "king of the hill · the old Templar slot", "netuid": 3, "url": "https://teutonic.ai/", "relation": "reference", "secondary_relation": null, "last_verified": "2026-09-29", "people": [ { "name": "Jacob Steeves (Const)", "role": "Bittensor co-founder; first author of the Teutonic-I report; also runs Affine", "handle": "const_reborn", "said": "11 May 2026: \"We construct the loss landscape as a market.\"" }, { "name": "Arbos", "role": "maintainer; also Albedo SN97" } ], "what_it_is": { "value": "\"A king-of-the-hill pretraining system for Bittensor subnet 3\": miners upload checkpoints, a validator duels challenger against king, the winner is paid a reign.", "label": "verified" }, "mechanism": { "summary": { "value": "\"Competition parallel\": every miner pretrains the king further on its own rented multi-GPU node and uploads a whole checkpoint; nothing is shared. Paired loss duel on held-out text. Emission 2.4% today (aggregator).", "label": "verified" }, "detail": { "value": "**No distributed training.** Each miner continues pretraining the current king with its own data on an 8x H200 or B200 node or larger, per the project's guide, and uploads a full checkpoint; \"Miners do not jointly compute an optimization step, exchange gradients, or average parameters.\" The validator runs a paired cross-entropy duel on held-out sequences chosen by block hash and crowns a challenger only if the confidence bound clears the threshold. The eval set began as FineWeb-Edu; since 9 September it is a five-source mix including math, code and reasoning text. Every king is public. The current king and up to four prior kings share the emission, 20% each. The acceptance threshold has moved: llms.txt (24 August) says 0.5 nats over 2,000 sequences; 0.0095 over 20,000 on 19 September; 0.003 over 30,000 since 21 September.", "label": "verified" } }, "largest_run": { "parameters_b": 110, "summary": { "value": "**Teutonic-II-110B**: 110B MoE (about 7.3B active), trained by rival miners one checkpoint at a time; not one shared run. Duels since 24 August 2026; 34 coronations by 29 September, lately about one a day. Tokens, hardware and MFU unknown by design.", "label": "verified" }, "detail": { "value": "**Teutonic-II-110B**: 256-expert MoE, about 7.3B active, randomly initialized genesis. The dashboard's first challenge record is 24 August 2026. On 29 September the king was UID 179, crowned that day at 00:06 UTC, reign 34; the reign number counts coronations. Since genesis: 32 distinct UIDs and 14 distinct coldkeys have held the crown; 666 duels, 33 accepted (5.0%). Teutonic-I (Quasar-10B architecture, a 10B hybrid gated-linear-attention model, random init) ran 2 June to 10 August 2026: 2,163 duels, 203 coronations, a 9.3% acceptance rate, about 640 hours of one eight-GPU evaluator. Tokens trained and miner hardware are not logged; the report gives no centralized baseline.", "label": "verified", "note": "The report calls the evaluator time 640 GPU-hours; at 17.7 minutes per duel on one eight-GPU server it is about 640 server-hours." }, "shared_run": false }, "sold": { "value": "Nothing. Every king is free on Hugging Face; miners are paid in SN3 alpha.", "label": "verified" }, "paid": { "value": "None public.", "label": "verified" }, "funding": { "value": "Emission-funded. Dendrite Holdings lists Teutonic as \"developing\" and says it co-created SN3 in 2026.", "label": "inferred", "note": "Dendrite listing verified on dendrite.holdings, 29 September 2026; emission-funded is inferred, no other revenue found." }, "relation_argument": "Reference: the other pretraining subnet on Bittensor, first-authored by the network's co-founder, with a 110B mixture-of-experts model (about 7.3B active). It sells nothing, so it does not compete for IOTA's buyer. Its guide tells miners to train on an 8x H200 or B200 node, rented on Lium, so it does not draw on the single cards IOTA's miners run; the overlap is TAO emission and alpha buyers, which is not supply.", "sources": [ { "title": "Teutonic validator repository", "url": "https://github.com/unarbos/teutonic" }, { "title": "Teutonic-II protocol (llms.txt), 24 August 2026", "url": "https://teutonic.ai/llms.txt" }, { "title": "Teutonic-I 10B: Competition Parallel Decentralized Training (Steeves, Warchoł, Korzeniewski, Bombała), August 2026", "url": "https://teutonic.ai/paper.html" }, { "title": "Teutonic-II dashboard data (reigns, thresholds, dataset mix)", "url": "https://teutonic.ai/dashboard.json" }, { "title": "Dendrite Holdings, subnet list", "url": "https://dendrite.holdings" }, { "title": "Genesis checkpoint", "url": "https://huggingface.co/dendriteholdings/teutonic-II-110B-genesis" }, { "title": "legible.network record, SN3 (read 9 to 19 September 2026)", "url": "https://legible.network/sn/3/" }, { "title": "tao.media on the May 2026 announcement of an 80B run (secondhand; superseded by Teutonic-II)", "url": "https://www.tao.media/teutonic-subnet-begins-training-80b-ai-model-on-bittensor-marking-largest-decentralized-training-run-yet/" }, { "title": "Welcome to Train at Home, Macrocosmos Substack (unveiled 9 December 2025 with Const running it live)", "url": "https://macrocosmosai.substack.com/p/welcome-to-train-at-home" } ], "signals": { "url": "https://teutonic.ai/", "read": "2026-09-29", "js_rendered": false, "title": "Teutonic Dashboard", "h1": "Teutonic", "first_h2": "DATASET MIX", "meta_description": "unknown", "og_description": "unknown", "we_are_sentence": "THE VALIDATOR IS READY FOR THE NEXT MODEL", "proof_numbers": [ "ED LCB -- WAITING FOR FIRST 10% CHECKPOINT WAITING 0% QUEUE 0 MODELS POSITION UID MODEL DIGEST HOTKEY COLDKEY BLOCK STATE SUBMIT" ], "audience_named": [ "unknown" ], "cta_verb": "unknown", "contested_words_present": [], "token_symbols": [], "dates_on_page": [ "unknown" ], "site_kind": "dashboard", "what_it_shows": "the TAO price, the subnet's alpha, the reigning king's hash, the genesis checkpoint, the validator repository, the report, the dataset mix and a loss curve; refreshed every few seconds from dashboard.json", "spoken": [ { "text": "Teutonic is a king-of-the-hill LLM pretraining competition. At any moment one checkpoint is the \"king\".", "source": "teutonic.ai/llms.txt, section 0, last verified by them 24 August 2026", "url": "https://teutonic.ai/llms.txt", "label": "verified" }, { "text": "You sell alpha for TAO, buy more GPU time, train again. That is the whole loop.", "source": "teutonic.ai/llms.txt, section 0", "url": "https://teutonic.ai/llms.txt", "label": "verified" }, { "text": "Everything here is derived from the live validator source code and chain contract, not from marketing copy.", "source": "teutonic.ai/llms.txt, preamble", "url": "https://teutonic.ai/llms.txt", "label": "verified" }, { "text": "Each hotkey gets exactly ONE submission, permanently. Plan accordingly.", "source": "teutonic.ai/llms.txt, section 0", "url": "https://teutonic.ai/llms.txt", "label": "verified" } ], "read_via": "the front door is a live dashboard, so the words come from llms.txt, the guide the project wrote for agents" }, "message": { "category": "Teutonic Dashboard", "h1": "Teutonic", "promise": "unknown", "proof": "THE VALIDATOR IS READY FOR THE NEXT MODEL", "audience": "unknown", "cta": "unknown", "label": "verified" }, "voice": { "carries_label": "founder-led, no brand channel", "read": "2026-09-29", "carries": "Every dated Teutonic statement traces to Jacob Steeves's account (30,759 followers), his interviews, or the report he first-authored. The only Teutonic X account is unofficial, with seven followers; the maintainer's own bio advertises a different subnet. Nothing about Macrocosmos or IOTA by name was found in his 2026 record; the closest is Macrocosmos's own account that he ran Train at Home live at its unveiling on 9 December 2025.", "message": "\"Teutonic 80 B is the largest decentralized training ever.\" Steeves, 11 May 2026. The run that shipped in June was 10B; the 110B began in August.", "accounts": [ { "channel": "@const_reborn", "url": "https://x.com/const_reborn", "note": "30,759; the only channel that matters" }, { "channel": "teutonic.ai", "url": "https://teutonic.ai/", "note": "dashboard, report, protocol file; no posts" }, { "channel": "GitHub", "url": "https://github.com/unarbos/teutonic", "note": "12 stars; pushed 24 September" }, { "channel": "Novelty Search", "url": "https://www.youtube.com/@OpentensorFoundation", "note": "Steeves hosts the community call on the Opentensor channel" } ], "founders": [ { "name": "Jacob Steeves (Const)", "role": "Bittensor co-founder; Teutonic-I first author; CEO, Affine", "handle": "const_reborn", "bio": "\"Building neurons, turning the web into a brain. Missionary from the Church of RAO @opentensor\" (30,759)", "statements": [ { "said": "\"Teutonic 80 B is the largest decentralized training ever.\"", "medium": "X", "venue": "", "date": "2026-05-11", "url": "https://x.com/const_reborn/status/2053802447324188766" }, { "said": "\"We construct the loss landscape as a market.\"", "medium": "announcement", "venue": "quoted by tao.media", "date": "2026-05-11", "url": "https://www.tao.media/teutonic-subnet-begins-training-80b-ai-model-on-bittensor-marking-largest-decentralized-training-run-yet/" }, { "said": "\"I wrote the original version of Templar. I bought the subnet for Sam.\" \"I don't have unilateral decision-making over the emissions.\"", "medium": "YouTube", "venue": "clip on the Covenant exit", "date": "2026-04-10", "url": "https://www.youtube.com/shorts/RI_HN-a9e4c" }, { "said": "\"Templar is, you know, I think by far the most impressive one so far I've seen.\" Two weeks before the exit.", "medium": "YouTube", "venue": "VirtualBacon interview", "date": "2026-03-26", "url": "https://www.youtube.com/watch?v=YmxqgTXfpDo" }, { "said": "\"An individual with a subnet done right can become a trillion-dollar company.\"", "medium": "YouTube", "venue": "VirtualBacon interview", "date": "2026-03-26", "url": "https://www.youtube.com/watch?v=YmxqgTXfpDo" } ], "recurring_words": [ "market", "loss", "king of the hill", "digital commodity", "permissionless", "largest decentralized training", "Templar" ] } ] }, "largest_model": { "value": "**Teutonic-II-110B**, the same: about 110B total, a MiMoV2-style mixture-of-experts with 256 routed experts, 8 selected per token, about 7.3B active. The genesis checkpoint is random weights, not a pretrained base.", "label": "verified", "parameters_b": 110, "active_b": 7.3, "note": "Hugging Face model card, dendriteholdings/teutonic-II-110B-genesis, read 29 September 2026. One shared expert per MoE layer as well." } } ], "references": [ { "name": "SWARM Parallelism, 2023", "who": "Ryabinin, Dettmers, Diskin, Borzunov", "showed": "Temporary randomized pipelines over preemptible T4s at 200 Mb/s, rebalanced on failure, trained a 1B shared-parameter model.", "iota": "IOTA's nearest ancestor. IOTA adds a central orchestrator that assigns, scores and pays on chain.", "url": "https://arxiv.org/abs/2301.11913" }, { "name": "Petals and Hivemind, 2020 to 2022", "who": "learning-at-home; Borzunov et al.", "showed": "Volunteer machines found over a DHT; Petals served and fine-tuned BLOOM-176B by splitting layer blocks across consumer GPUs.", "iota": "The same shape as IOTA's slices, with trusted peers and nobody paid.", "url": "https://arxiv.org/abs/2209.01188" }, { "name": "DiLoCo, 2023; Streaming DiLoCo, 2025", "who": "Douillard et al., Google DeepMind", "showed": "Data-parallel replicas take many local steps and sync through an outer optimizer, matching synchronous training with 500x less communication; streaming subsets cut bandwidth two more orders.", "iota": "The method behind Prime, Nous, Covenant and Flower. IOTA shares the slow-links premise, not the data-parallel shape.", "url": "https://arxiv.org/abs/2311.08105" }, { "name": "DeMo, 2024", "who": "Peng, Chen, Su, Quesnelle, Kingma, Liu (Nous)", "showed": "Decoupled momentum with a sparsified transform sends up to 85x less data per GPU than AdamW-DDP at 300M to 1B.", "iota": "Compresses gradients between replicas; IOTA compresses activations between slices. Different bottleneck, could stack.", "url": "https://arxiv.org/abs/2411.19870" }, { "name": "Decentralized training in heterogeneous environments, 2022", "who": "Yuan et al.; Together's founders", "showed": "Scheduling for geo-distributed runs across three continents, 4.8x over Megatron.", "iota": "The academic root of the whole field; the company now sells clusters.", "url": "https://arxiv.org/abs/2206.01288" } ], "left_out": [ { "name": "Affine, SN120", "why": "A teacher-anchored distillation contest, not distributed training: each miner fine-tunes its own model to match a frozen teacher; nothing sold. The operator rents Lium pods to serve the leading model. Steeves matters as a person and is on the Teutonic card." }, { "name": "Swarm, SN124", "why": "Drone flight policies, one per miner, trained locally and scored in simulated worlds; no LLM, no compute sold, no fleet. A €370k raise and a solar-park pilot, secondhand." }, { "name": "Bagel Labs", "why": "Now a physical-AI lab. Paris (2025) trained 8 isolated experts with no communication on clouds it rented (AWS, GCP, Runpod, local); no network, nothing sold; Bakery is offline. Toronto; $3.1M pre-seed (2024), secondhand." }, { "name": "EXO Labs", "why": "Local inference on devices you own, sold to businesses as deployments; no training product. Training work is small-cluster and simulated research (SPARTA, 124M, up to 8 nodes; EXO Gym idle since December 2025). No open network." }, { "name": "io.net, Akash, Hyperbolic, Aethir, Render, Together", "why": "None trains one model across machines it does not run as one cluster. Akash, Hyperbolic, Aethir and Render rent GPUs and the customer runs the job. io.net (Training as a Service, beta) and Together sell managed fine-tuning on a rented cluster, not pretraining. They rent out other people's hardware." }, { "name": "ChronoLLM SN38; SN39, Basilica's former subnet", "why": "SN38 is ChronoLLM: miners train whole models with a date cutoff and a validator duels them; not distributed training, nothing sold. Emission 3.9% today. SN39 has been deprecated since Covenant's April exit and carries no emission; Basilica now runs off-chain and is on the Covenant card." } ], "disputed_claim": { "who": "Most likely to be raised by either team", "claim": "Which run is the largest. By tokens, Covenant-72B: about 1.1 trillion against Orion-100B's 1.1 billion. By dense parameters, Orion-100B (Teutonic-II has 110B total but is a mixture-of-experts with about 7.3B active, trained one checkpoint at a time). Covenant-72B was a completed permissionless pretraining run; Orion-100B was a viability run on 48 provisioned A100s, stopped for cost; the permissionless run is Orion-16B.", "settles": "The Covenant-72B paper (arXiv 2603.08163, 1.09T DCLM tokens plus annealing) beside the Orion-100B write-up (1.1B tokens, stopped for cost). Both numbers are on the page; the page says which is which." }, "note": "Two pages. The register: every IOTA run with its numbers. The field map: every project training across compute outside a single reserved cluster, placed by its relation to IOTA, mechanism first. Both rendered from JSON, every cell sourced or marked unknown, corrections by mail.", "iota_signals": { "iota-today": [ { "url": "https://iota.macrocosmos.ai/", "read": "2026-09-28", "js_rendered": false, "title": "IOTA | Macrocosmos.ai", "h1": "The home of liquid training.", "first_h2": "Compute isn’t the problem, accessing it is.", "meta_description": "Incentivised Orchestrated Training Architecture from Macrocosmos.ai", "og_description": "Incentivised Orchestrated Training Architecture from Macrocosmos.ai", "we_are_sentence": "IOTA is your disaggregated training infrastructure, turning scattered and underused compute into unified capacity you can train on.", "proof_numbers": [], "audience_named": [ "For teams" ], "cta_verb": "Train with IOTA", "contested_words_present": [ "distributed", "open" ], "token_symbols": [], "dates_on_page": [ "unknown" ] }, { "url": "https://www.macrocosmos.ai/", "read": "2026-09-28", "js_rendered": false, "title": "Macrocosmos.ai", "h1": "Incentivizing intelligence that scales", "first_h2": "Decentralizing AI, delivering fast, flexible, and efficient compute", "meta_description": "Decentralizing AI, delivering fast, flexible, and efficient compute", "og_description": "Decentralizing AI, delivering fast, flexible, and efficient compute", "we_are_sentence": "Permissionless, accessible, and scalable pretraining, harnessing distributed compute to build the largest decentralized model ever trained.", "proof_numbers": [], "audience_named": [ "unknown" ], "cta_verb": "Docs", "contested_words_present": [ "decentralized", "distributed", "permissionless", "open" ], "token_symbols": [], "dates_on_page": [ "unknown" ] }, { "url": "https://iota.macrocosmos.ai/", "read": "2026-09-28", "js_rendered": false, "title": "IOTA | Macrocosmos.ai", "h1": "The home of liquid training.", "first_h2": "Compute isn’t the problem, accessing it is.", "meta_description": "Incentivised Orchestrated Training Architecture from Macrocosmos.ai", "og_description": "Incentivised Orchestrated Training Architecture from Macrocosmos.ai", "we_are_sentence": "IOTA is your disaggregated training infrastructure, turning scattered and underused compute into unified capacity you can train on.", "proof_numbers": [], "audience_named": [ "For teams" ], "cta_verb": "Train with IOTA", "contested_words_present": [ "distributed", "open" ], "token_symbols": [], "dates_on_page": [ "unknown" ] }, { "url": "https://iota.macrocosmos.ai/sdk", "read": "2026-09-28", "js_rendered": false, "title": "SDK | IOTA | Macrocosmos.ai", "h1": "The IOTA SDK is coming soon", "first_h2": "You define the workload. IOTA runs it.", "meta_description": "Incentivised Orchestrated Training Architecture from Macrocosmos.ai", "og_description": "Incentivised Orchestrated Training Architecture from Macrocosmos.ai", "we_are_sentence": "The IOTA SDK will let you configure a workload and run it across disaggregated compute directly, with the coordination, execution, and recovery handled for you.", "proof_numbers": [], "audience_named": [ "unknown" ], "cta_verb": "Train with IOTA", "contested_words_present": [ "distributed", "open" ], "token_symbols": [], "dates_on_page": [ "unknown" ] }, { "url": "https://www.macrocosmos.ai/liquid-compute", "read": "2026-09-28", "error": "HTTP Error 404: Not Found" } ] }, "message_rows_iota": [ { "name": "IOTA today", "id": "iota-today", "url": "https://iota.macrocosmos.ai/", "read": "2026-09-28", "label": "verified", "category": "your disaggregated training infrastructure", "h1": "The home of liquid training.", "promise": "turning scattered and underused compute into unified capacity you can train on", "proof": "unknown", "audience": "For teams", "cta": "Train with IOTA", "words": [ "distributed", "open" ], "tokens": [] } ], "what_they_say": { "paragraphs": [ "**Common words.** \"Open\" is on nine of twelve home pages, including IOTA's. \"Decentralized\" is on six. \"Distributed\" on four.", "**Verification words.** No home page on the map uses \"verifiable\", \"auditable\" or \"trustless\". \"Incentivized\" appears only on Templar's, which is dormant. IOTA's own name already contains it.", "**Numbers on home pages.** Three rivals lead with a number on the home page: Pluralis (\"6B-parameter open pretraining run trained on 500B tokens across 330 contributors\", their home page), Prime Intellect (\"INTELLECT-3: A 100B+ MoE\"), Chutes (\"Trillions of Tokens per Month\"). IOTA's page leads with none; Orion-100B, the largest dense model trained across machines on this map, is not on its own home page." ] }, "about": { "voice_intro": [ "Founder-led marketing means the founder's own account and appearances carry the message and the brand account amplifies them. This section records, for IOTA and for every entity on the map, who actually carries it: the canonical accounts with their size and cadence, the founders' bios and dated statements across X, podcasts, talks, essays and interviews, and the words each person keeps using. Quotes are verbatim, under twenty words, read from the source on 28 September 2026; X timelines were read through mirrors or the browser, so thirty-day cadence is marked where it could not be observed.", "**The pattern across the field.** Three shapes. Founder-led: Steeves (30,759 followers, no brand channel at all), Long (the argument on his account, metrics on the company's), Dare before April. Brand-led: Prime Intellect (84,789 on the company account, every post signed \"Team\"), Gensyn (93,020, retweeting its founders), Nous on launches (275,027), Targon and Flower (no founder byline in 2026). Private founder-led: Fish, who sets Lium's prices by Discord directive and has never appeared on a podcast. The one founder whose reach matches the brand is Teknium (130,042 against 275,027), and he is the product's hands, not its CEO.", "**The words.** Steffen's adjectives: \"heterogeneous, unreliable, permissionless and token incentivized\" (May 2025); \"globally distributed, heterogeneous and unreliable\" (September 2026). \"Trustless\" appears in 2025 posts and the docs and in nothing from 2026. \"Liquid\" replaced \"swarm\" in June; Pluralis's founding scientist uses \"liquid compute\" too (May and September). \"Compute nobody reserved\" is in the IOTA account's own copy." ] } }