{
  "register": "The Orion Register",
  "subject": "IOTA, Bittensor subnet 9, every Orion training run with its numbers",
  "compiled_by": "Mikyö Clark, from Macrocosmos's published record. Not affiliated.",
  "updated": "2026-10-04",
  "labels": {
    "stated": "Macrocosmos said it, in the source given",
    "verified": "checked against the code, the chain, or a live page on the date given",
    "secondhand": "reported by someone else; not yet said by Macrocosmos in writing",
    "inferred": "reasoned from stated facts; the reasoning is given",
    "unknown": "the published record does not fill this cell"
  },
  "rule": "A run is added when it ends. Every cell is filled or marked unknown. Pilots are added only if Macrocosmos publishes them.",
  "deck": {"value": "The 25-page stage deck from the Exploit Summit (28 Sep 2026, Will Squires and Steffen Cruz) is Macrocosmos's own text, so its figures are labelled stated with a page number. The full talk is public on the summit's replay page (stream.vidaio.io), and spoken figures are cited to its minute. A stage figure enters the runs table only where it names a run and a statistic.", "label": "stated"},
  "corrections": "connect@mikyo.one",
  "runs": [
    {
      "run": "Orion-100B",
      "status": {"value": "ended; a viability run, stopped at 1.1B tokens on cost grounds", "label": "stated", "source": "Macrocosmos on X, 25 Sep 2026: 'It ran as a viability proof rather than to a finished model, and it held.'"},
      "announced": {"value": "2026-06-01", "label": "stated", "source": "substack"},
      "run_dates": {"value": "about two days; calendar dates not published; May 2026 by the essay's own timeline", "label": "inferred"},
      "parameters_b": {"value": 100, "label": "stated"},
      "architecture": {"value": "modified Llama 3.2 family, 90 transformer blocks", "label": "stated"},
      "hardware": {"value": "48 Nvidia A100 80GB, one per peer, non-colocated, across five datacenters in the United States, several providers", "label": "stated"},
      "network": {"value": "median upload 856 Mbps, download 1322 Mbps", "label": "stated"},
      "pipeline": {"value": "16 stages, 3 replicas", "label": "stated"},
      "replica_size_gpus": {"value": 16, "label": "stated"},
      "inner_steps": {"value": 10, "label": "stated"},
      "tokens_trained_b": {"value": 1.1, "label": "stated"},
      "throughput_tokens_per_s": {"value": 9000, "label": "stated"},
      "mfu_avg_pct": {"value": 30.8, "label": "stated"},
      "mfu_peak_pct": {"value": 38, "note": "sustained over six hours without churn", "label": "stated"},
      "speed_vs_colocated_pct": {"value": 65, "peak": 82, "label": "stated"},
      "phase_minutes": {"training": 100.2, "synchronization": 22.2, "label": "stated"},
      "training_utilization_pct": {"value": 81.8, "label": "stated"},
      "activation_compression": {"value": "ResBM, 64x; per-stage transfer 140.6 MB to 2.2 MB", "label": "stated"},
      "cost_per_replica_hour_usd": {"value": 20, "basis": "16 A100s at about $1.25 an hour", "label": "stated"},
      "cost_per_run_hour_usd": {"value": 60, "basis": "48 A100s at $1.25; arithmetic on stated figures", "label": "inferred"},
      "churn": {"value": "peak measured without churn; bandwidth-limited peers slowed weight uploads and the run continued; event count not published", "label": "stated"},
      "dataset": {"value": "fineweb-edu-score-2", "label": "stated"},
      "checkpoint_available": {"value": false, "label": "verified", "note": "No public checkpoint found (Hugging Face and the IOTA site, 28 September)."},
      "when": "Viability run · announced 1 Jun 2026",
      "entry_cost_vs_peer": {"value": "about 2.5x below a Covenant-class 8xB200 peer (about $50 an hour) per replica", "label": "stated", "source": "the 1 June write-up; the 25 September post says the same"},
      "cost_vs_reserved": {"value": "3x cheaper than reserving a node in a datacenter; 'equivalent performance to centralised' at 65% of its speed", "label": "stated", "source": "the stage deck, p. 10; Steffen Cruz said the 3x line on stage (The TAO Daily's clip)"},
      "what_it_proved": "Pipeline-parallel pretraining at 100B scale across five US datacenters, one A100 per peer at about $1.25 an hour, at roughly two-thirds of co-located speed.",
      "sources": [
        {"title": "Orion-100B: Distributed pretraining arrives at hundred-billion-parameter scale, Steffen Cruz, 1 Jun 2026", "url": "https://macrocosmosai.substack.com/p/orion-100b-distributed-pretraining"},
        {"title": "IOTA on X, 24 Sep 2026 (five datacenters, MFU, speed)", "url": "https://x.com/IOTA_SN9/status/2103175140699934764"},
        {"title": "Macrocosmos on X, 25 Sep 2026 (replica cost, viability proof)", "url": "https://x.com/MacrocosmosAI/status/2103537520558661976"},
        {"title": "The Exploit Summit stage deck, 28 Sep 2026, pp. 9 to 10 (on file, not online)"}
      ]
    },
    {
      "run": "Orion-16B, first run",
      "status": {"value": "finished; 'one finished model'", "label": "stated"},
      "live_from": {"value": "2026-07-27", "label": "stated"},
      "tokens_100b_passed": {"value": "2026-08-14", "label": "stated"},
      "end_date": {"value": "before 2026-09-23; date not published", "label": "inferred"},
      "parameters_b": {"value": 16, "label": "stated"},
      "architecture": {"value": "Llama-3.2-16B, 44 layers, embedding 5120, context 2048, bf16", "label": "verified", "source": "IOTA dashboard, archived run 28265 (orion-16b-v8), run config read in a browser 2026-09-28", "url": "https://iota.macrocosmos.ai/dashboard/overview"},
      "hardware": {"value": "a series: 256 heterogeneous GPUs at launch (2026-07-27); c.180+ (2026-08-05); 225 across three continents (2026-08-20); at the finish (2026-09-23) RTX 4090, RTX 5090, A100, L40S, A6000; permissionless providers", "label": "stated"},
      "accessible_pool": {"value": "about 18 B200-equivalents at FP16", "label": "stated"},
      "pipeline": {"value": "10 pipeline stages; replica count not published", "label": "verified", "source": "Macrocosmos on X, 27 Jul 2026; dashboard run config (10 pipeline splits)", "replicas": {"value": null, "label": "unknown"}},
      "tokens_trained_b": {"value": "more than 100 by 14 Aug", "label": "stated", "dashboard_archived_run": {"value": 219.05, "label": "verified", "source": "IOTA dashboard, archived run 28265 (orion-16b-v8), tokens tile read in a browser 2026-09-28"}},
      "compute": {"value": "about 10^22 FLOPs by 14 Aug", "label": "stated"},
      "throughput_tokens_per_s": {"value": 80000, "label": "stated"},
      "mfu_avg_pct": {"value": 20, "label": "stated", "alternate": {"value": 24, "statistic": "median MFU, per-worker p75", "label": "stated", "source": "the stage deck, p. 12 ('16B at 24% median MFU'; footnote 'iota: per-worker p75'); first reported live by Subnet Summer on X. A different statistic from the 23 September post's 'roughly 20%' average, not a revision of it; the deck does not say which 16B run"}},
      "speed_vs_colocated_pct": {"value": null, "label": "unknown"},
      "cost_per_replica_hour_usd": {"value": null, "label": "unknown"},
      "churn": {"value": "nodes dropped, throughput swung, machines ran inconsistently; training carried on without direct intervention; event count not published", "label": "stated"},
      "dataset": {"value": "fineweb-edu-2", "label": "verified", "source": "IOTA dashboard, archived run 28265 (orion-16b-v8), run config read in a browser 2026-09-28"},
      "checkpoint_available": {"value": false, "label": "verified", "note": "No public checkpoint found (Hugging Face and the IOTA site, 28 September)."},
      "gpus_series": [
        {"date": "2026-07-27", "value": 256, "label": "stated", "source": "Macrocosmos on X, launch post"},
        {"date": "2026-08-05", "value": "c.180+", "label": "stated", "source": "Deep Dive 2"},
        {"date": "2026-08-20", "value": 225, "note": "three continents", "label": "stated", "source": "Steffen Cruz on X"},
        {"date": "2026-09-23", "value": "4090s, 5090s, A100s, L40s, and A6000s", "note": "at the finish; card types named, no count given", "label": "stated", "source": "Macrocosmos on X, final post"}
      ],
      "compression": {"value": "bottleneck dimension 64", "label": "verified", "source": "IOTA dashboard, archived run 28265 (orion-16b-v8), run config read in a browser 2026-09-28"},
      "loss": {"value": 2.85, "at_tokens_b": 218, "label": "stated", "source": "the stage deck's convergence chart, p. 11 ('iota 16B · v8, ~218B tokens'); the dashboard's archived tile read 2.8567 at 219.05B tokens on 28 Sep, which agrees", "chart_peers": "Pluralis-13B 3.01, Pluralis-8B 2.44, OLMo 2 13B (centralised) 2.46, on the same chart"},
      "what_it_proved": "A 16B run survives mixed hardware and node churn across continents with no reserved cluster anywhere in it.",
      "sources": [
        {"title": "Macrocosmos on X, 27 Jul 2026 (launch: 16B, 256 GPUs, 10 pipeline stages)", "url": "https://x.com/MacrocosmosAI/status/2081791889674772618"},
        {"title": "Deep Dive 2: The Technical Requirements for Liquid Training, Cruz and Cohen, 5 Aug 2026", "url": "https://macrocosmosai.substack.com/p/deep-dive-2-the-technical-requirements"},
        {"title": "Steffen Cruz on X, 14 Aug 2026 (100B tokens)", "url": "https://x.com/macrocrux/status/2088298605912174989"},
        {"title": "Steffen Cruz on X, 20 Aug 2026 (225 GPUs, three continents)", "url": "https://x.com/macrocrux/status/2090544176274223467"},
        {"title": "Macrocosmos on X, 23 Sep 2026 (final: 80k tokens a second, about 20% MFU)", "url": "https://x.com/MacrocosmosAI/status/2102808976442458457"},
        {"title": "The IOTA dashboard, archived run, read 28 Sep 2026", "url": "https://iota.macrocosmos.ai/dashboard/overview"},
        {"title": "The Exploit Summit stage deck, 28 Sep 2026, pp. 11 to 12 (on file, not online)"}
      ]
    },
    {
      "run": "Orion-16B, second run",
      "status": {"value": "training", "label": "stated"},
      "seen_training": {"value": "2026-09-28", "label": "verified", "note": "on the IOTA dashboard; start date not published"},
      "parameters_b": {"value": 16, "label": "stated"},
      "hardware": {"value": null, "label": "unknown"},
      "tokens_trained_b": {"value": null, "label": "unknown"},
      "mfu_avg_pct": {"value": null, "label": "unknown"},
      "throughput_tokens_per_s": {"value": null, "label": "unknown"},
      "cost_per_replica_hour_usd": {"value": null, "label": "unknown"},
      "churn": {"value": null, "label": "unknown"},
      "loss": {"value": 2.75, "label": "stated", "source": "the stage deck's convergence chart, p. 11, labels the 'iota 16B · v9, most recent' curve 2.75; the token count at that point is not legible in the export", "note": "the dashboard tile read 1.1296 at 86.62B tokens on 28 Sep; not the same measure, and which is the run's loss is unknown"},
      "what_it_proved": "Nothing yet. The run is in progress.",
      "dashboard": {"value": "the IOTA dashboard lists this run as orion-16b-v9, 10 stages, 11 replicas", "label": "verified", "date": "2026-09-28", "url": "https://iota.macrocosmos.ai/dashboard/overview"},
      "sources": [
        {"title": "Subnet Summer on X, live from the Exploit stage, 28 Sep 2026", "url": "https://x.com/SubnetSummerTAO/status/2104591996459360637"},
        {"title": "The IOTA dashboard, read 28 Sep 2026", "url": "https://iota.macrocosmos.ai/dashboard/overview"},
        {"title": "The Exploit Summit stage deck, 28 Sep 2026, p. 11 (on file, not online)"}
      ]
    }
  ],
  "earlier_runs": [
    {"period": "June 2025", "what": "launch run: a 15B target model across 5 pipeline stages, permissionless; 'fell over almost immediately' under load and adversarial attacks", "label": "stated"},
    {"period": "August 2025 to April 2026", "what": "the 1.5B, 3-layer testbed: more than 700 controlled experiments, almost 15 trillion tokens in total, throughput up about tenfold", "label": "stated", "source": "the Orion-100B write-up, 1 Jun 2026 ('over 700 controlled experiments on our 1.5B/3L testbed'); said again on Hash Rate, Ep. 176, 23 Jun 2026, at 26:09 (local transcription)", "note": "the stage deck, 28 Sep 2026, p. 9, gives '750 small-scale experiments in a year'; the write-up's 750 counts 1.5B to 21B and its 700 the testbed alone. Both figures are Macrocosmos's; the 700 is kept for the testbed"},
    {"period": "April to May 2026", "heading": "67x in a month", "what": "model size up 67x and pipeline stages up 10x in a single month; more than 750 experiments in all, from 1.5B to 21B, before Orion-100B", "label": "stated", "source": "the Orion-100B write-up, 1 Jun 2026, for the 750 ('validated through over 750 experiments from 1.5B through to 21B parameter model scale')"},
    {"period": "18 months since launch, as summed up on stage", "heading": "The science first", "what": "'1st live pipeline-parallel training run, ever'; '750 small-scale experiments in a year'; ResBM 'state of the art for pipeline parallel'; experiments 'in excess of 250,000 GPU hours' behind the 16B convergence chart", "label": "stated", "source": "the stage deck, pp. 9 and 11"}
  ],
  "said_on_stage": {
    "session": "Project Orion: Introducing Liquid Training",
    "where": "OTF stage, Exploit Summit, Montreal",
    "when": "2026-09-28 10:50 local",
    "speakers": ["Will Squires", "Steffen Cruz"],
    "replay": {"value": "no full replay published; a 38-second highlight titled 'Announcing the Iota SDK and Liquid Compute Platform'", "url": "https://stream.vidaio.io/highlights/d1eaa7b9ebc6e77f433884f1c1ad2cfa", "label": "verified", "note": "as read on 28 Sep 2026; superseded by the full recording found on 3 Oct, see recording and spoken"},
    "recording": {"value": "full recording of the talk on the summit's replay page, 30 min 54 s, English and French subtitles; the 29 Sep debate 'Research vs Revenue', 39 min 46 s, on the same page", "url": "https://stream.vidaio.io/vod.html", "label": "verified", "date": "2026-10-03", "note": "the debate's listing is stated, not verified", "source": "research/iota/2026-10-03-iota-since-the-stage.md, finding 5"},
    "announced": {"value": "the iota SDK and Liquid Compute; Liquid Compute is 'our disaggregated compute platform'; the SDK 'powers any training workload across it, as if it were one cluster'; go-to-market 'in the coming weeks'; early access by registration", "label": "stated", "source": "Macrocosmos on X, about 15:30 local, 2026-09-28"},
    "deck_definitions": {"value": {"iota SDK": "Fragmented, interruptible compute that trains like a centralised cluster. Drop-in for PyTorch.", "Liquid Compute": "Harnessing disaggregated, interruptible, heterogeneous compute into a global supercomputer.", "title": "Project Orion. iota. Introducing Liquid Training."}, "label": "stated", "source": "the stage deck, pp. 1 and 15"},
    "mechanisms": {"value": ["heterogeneous compute: mismatched hardware, one job; NVLink or InfiniBand in the datacentre, iroh over the internet", "fault tolerance: nodes drop and rejoin mid-run; training continues at 80% quorum (iota.fault_tolerant(Relative(0.8))) and averages stay correct", "orchestration: DiLoCo, SparTA, pipelining and data loading built into the SDK"], "workloads": "pretraining, RL and post-training, inference, fine-tuning (frozen base on every node, small adapters train, merge), 'anything else'", "code_change": "dist.init_process_group(backend='nccl') becomes iota.use_mesh(mesh, group='all') and dist.init_process_group(backend='iota')", "label": "stated", "source": "the stage deck, pp. 16 to 18"},
    "sdk": {"value": "'anyone can train models using globally distributed, heterogeneous and unreliable compute with just a few lines changed from pure PyTorch'; two years of R&D distilled into communication primitives", "label": "stated", "source": "Steffen Cruz on X, 2026-09-28"},
    "cost_quotes": {"value": ["The most impressive thing about Orion-100B is the cost of doing it compared to colocated compute.", "We were able to train it 3x cheaper than if we had simply reserved a node in a data center and trained it there.", "That's because we can access the abundant compute that exists everywhere, in the corners of the world."], "speaker": "Steffen Cruz", "label": "stated", "source": "the stage deck, p. 10, prints '3x cheaper'; the wording as spoken is The TAO Daily's transcription from a 3:21 clip, 2026-09-28", "note": "the 1 June write-up puts entry cost per replica 2.5x below a Covenant-class 8xB200 peer; the 25 September post says the same"},
    "numbers": {"value": {"tokens_trained_b": 100, "model_16b_mfu_pct": 24, "mfu_statistic": "median, per-worker p75", "experiments": "750 small-scale experiments in a year", "experiment_gpu_hours": "in excess of 250,000", "dense_parameters_today": "2T dense: 'a design limit, not a completed run' (@MacrocosmosAI, 1 Oct 2026); Steffen Cruz, 2 Oct 2026: 'we don't know for certain'", "path": "ResBMoE R&D toward 10T and beyond; the deck's summary slide says ResBM and ResBMoE chart the path to 1T and 10T", "companies_in_conversation": 164, "active_conversations": 26, "sdk_registrations": "100+ ahead of launch", "pilots": "3 workloads scheduled from October", "nvidia_inception": "accepted in May 2026", "roadmap": {"October": "private beta and first pilots", "November": "customers onboarding", "year_end": "full public launch and scaling"}}, "label": "stated", "source": "the stage deck, pp. 9 to 13 and 22; the 164, the pilots and the 24% were first reported live by Subnet Summer on X, 2026-09-28", "note": "the 23 September post gave about 20% average MFU for the first 16B run; the deck's 24% is a median, per-worker p75"},
    "commercials": {"value": {"h100_per_gpu_hour_sept_2026": {"reserved_1yr_median_52_clouds": 3.10, "spot_average_28_clouds": 0.98, "spot_vast_idle_hosts": 0.18, "spread": "spot runs 3 to 17x cheaper than reserved"}, "utilisation_example": "one 10,000-GPU site at 85% utilised; iota recovers 10 points: 8.76m hours and $21.9m a year", "olmo_3_32b_example": "1.05m H100-hours reserved at $3.10 is $3.26m; on iota, 1.3x the hours (1.365m), $1.63m to the customer: half the price", "market": "neocloud GPU revenue outside the hyperscalers about $36bn a year; every point of utilisation orchestrated by iota is $420m a year (revenue divided by 85%); 5 points $2.1bn, 10 points $4.2bn"}, "label": "stated", "source": "the stage deck, pp. 19 to 21, citing AIMultiple (14 Sep 2026), FastGPU and GetDeploying (27 Sep 2026), Synergy Research (Apr 2026)", "note": "the deck's arithmetic, on the deck's inputs; none of it is run data. On 3 Oct 2026 only the $3.10 reserved median was found in a public source"},
    "framing": {"value": ["GPUs around the world are sitting idle. Macrocosmos wants to put them to work training AI.", "bringing spare compute across different machines into AI training"], "label": "secondhand", "source": "Exploit Summit on X, 2026-09-28"},
    "deck_framing": {"value": ["Demand for AI compute is outstripping supply, and will keep doing so", "Distributed networks exceed frontier buildouts today (Colossus 2 at 1.1 GW against Bitcoin at about 18 GW)", "The long tail of compute: 57 GW installed, 84% outside the current generation, consumer GPUs estimated at 37.7 GW; distributed, heterogeneous, idle", "Orchestrating disaggregated compute is the next great opportunity in AI efficiency", "Pipeline parallel is the only path to training frontier scale models distributed", "Distributed training is ready for trillion-parameter scale", "Other systems are built for one run, iota is all-purpose", "iota is the state of the art in distributed training; the iota SDK is the path to disruptive scale; iota is built to be highly profitable"], "label": "stated", "source": "the stage deck, pp. 3 to 9, 13, 17 and 23; the deck cites Citi and OpenAI, the AI Safety Report 2026, Epoch AI, Cambridge CBECI, Jon Peddie Research, and quotes Dylan Patel, Keith Rush, Thomas Kurian and the Fable 5 system card"},
    "why_sdk": {"value": ["All of our published results to date are on the legacy code base.", "pre training is too small of a market, so we've broadened the focus to include many more standard workloads (RL, fine tuning and more)"], "speaker": "Will Squires", "label": "stated", "source": "Will Squires on X, 2026-09-28", "url": "https://x.com/WSquires/status/2104681373533679638"},
    "spoken": {"value": [
      {"minute": 12, "topic": "Orion-100B fleet and length", "quote": "a total of fifty GPUs around the world ... for about three days to a billion tokens", "beside": "the write-up says 48 A100s in five US data centres over about two days", "label": "verified", "from": "the recording's machine captions (English subtitle track); wording approximate"},
      {"minute": 12, "topic": "cost", "quote": "approximately three times cheaper than if we had just reserved a node in a data center", "beside": "the deck, p. 10, prints '3x cheaper'", "label": "verified", "from": "the recording's machine captions (English subtitle track); wording approximate"},
      {"minute": 13, "topic": "Orion-16B fleet", "quote": "around two hundred and fifty GPUs, about a quarter of a million GPU hours", "beside": "the deck puts 'in excess of 250,000' GPU hours on the experiments", "label": "verified", "from": "the recording's machine captions (English subtitle track); wording approximate"},
      {"minute": 14, "topic": "suppliers and interruptions", "quote": "came from Green Compute, it came from Targon; around a hundred interruptions", "label": "stated", "from": "the recording's machine captions (English subtitle track); wording approximate"},
      {"minute": 15, "topic": "the 24% MFU against another team; the fleet's spread", "quote": "fifty percent better than the other state-of-the-art team ... nine providers, three continents, eight different GPU types at one point", "beside": "the deck gives 24% median MFU, per-worker p75; the captions garble the other team's name", "label": "verified", "from": "the recording's machine captions (English subtitle track); wording approximate"},
      {"minute": 17, "topic": "2T", "quote": "a Covenant class node, eight B200s, and we can just stack thirty-two of those; in 'any other context', eight hundred to a thousand B200s", "beside": "Macrocosmos on X, 1 Oct 2026: 'a design limit, not a completed run'", "label": "verified", "from": "the recording's machine captions (English subtitle track); wording approximate"},
      {"minute": 20, "topic": "the code change", "quote": "three lines of code", "beside": "the deck, p. 17, shows one line replaced by two", "label": "verified", "from": "the recording's machine captions (English subtitle track); wording approximate"},
      {"minute": 22, "topic": "utilisation", "quote": "from 85% to 95% ... 8.76 million GPU hours a year ... $22 million a year", "beside": "the deck: 8.76m hours and $21.9m a year", "label": "verified", "from": "the recording's machine captions (English subtitle track); wording approximate"},
      {"minute": 23, "topic": "the Olmo case", "quote": "it would cost him about a quarter", "beside": "the deck: half the price to the customer", "label": "verified", "from": "the recording's machine captions (English subtitle track); wording approximate"},
      {"minute": 24, "topic": "traction", "quote": "Over the last eight weeks, we spoke to a hundred and sixty-four companies ... three pilots set to start on October 1st and over a hundred registrations", "beside": "the deck: 3 workloads scheduled from October", "label": "verified", "from": "the recording's machine captions (English subtitle track); wording approximate"},
      {"minute": "27 to 30", "topic": "Q&A", "quote": "We actually have a compute provider form", "label": "stated", "from": "the recording's machine captions (English subtitle track); wording approximate"}
    ], "source": "the full recording of the talk, 30 min 54 s, read 3 Oct 2026 through its English subtitle track; research/iota/2026-10-03-iota-since-the-stage.md, finding 5", "url": "https://stream.vidaio.io/vod.html", "note": "every quote is from the recording's machine captions, not the audio; verified means three verifiers confirmed the captions say it"},
    "rule": "A stage figure becomes a runs-table cell only when Macrocosmos publishes it.",
    "label_on_page": "not yet run data"
  },
  "runs_table_excludes": ["the IOTA SDK", "Liquid Compute", "pilots", "the companies in conversation"]
}