Every Orion training run, with its numbers, dates and sources. Where the published record is silent, the page says unknown.
Three steps between the first launch and the 100B run.
The launch run, permissionless. It "fell over almost immediately" under system load and adversarial attacks.
A 1.5B, 3-layer testbed; almost 15 trillion tokens in total; system throughput up about tenfold.
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.
The 700 and the 750 are both from the Orion-100B write-up, 1 June 2026: 700 on the 1.5B testbed, 750 from 1.5B to 21B. The 700 was said again on Hash Rate, Ep. 176, 23 June. The stage deck (28 September, p. 9) gives "750 small-scale experiments in a year."
Exploit Summit, Montreal, 28 September 2026. "Project Orion: Introducing Liquid Training," Will Squires and Steffen Cruz. Figures from the 25-page stage deck are Macrocosmos's own, cited to the page; spoken figures are cited to the minute of the full recording, which is public (see the note below). A stage figure enters the runs table only where it names a run and a statistic.
"Powers any training workload across it, as if it were one cluster."
The iota SDK and Liquid Compute, "our disaggregated compute platform." Go-to-market "in the coming weeks"; early access by registration.
Spoken in the Q&A, 27 to 30 min, from the recording's captions: "We actually have a compute provider form."
Macrocosmos on X, about 15:30 Montreal time."Just a few lines changed from pure PyTorch."
Two years of research "distilled into a powerful set of communication primitives."
Spoken, 20 min, from the recording's captions: "three lines of code."verified The deck, p. 17, shows one line replaced by two.
Steffen Cruz on X, the same hour."3x cheaper than if we had simply reserved a node in a data center."
The deck prints "3x cheaper" beside Orion-100B, with "equivalent performance to centralised" at 65% of its speed. The 1 June write-up puts entry cost per replica 2.5x below a Covenant-class 8×B200 peer; the 25 September post says the same.
Spoken, 12 min, from the recording's captions: "approximately three times cheaper than if we had just reserved a node in a data center."verified
The stage deck, p. 10; the wording as spoken from The TAO Daily's 3:21 clip.164 companies "in conversation." 26 "active conversations." "100+" SDK registrations ahead of launch. Three pilot workloads scheduled from October.
16B at 24% median MFU, per-worker p75; the 23 September post's "roughly 20%" is an average, a different statistic. 750 experiments in a year, more than 250,000 GPU hours. 2T dense parameters, which Macrocosmos calls "a design limit, not a completed run"verified (1 October); Steffen Cruz, 2 October: "we don't know for certain."verified ResBMoE toward 10T. NVIDIA Inception since May. October: private beta and first pilots. November: customers onboarding. Year end: full public launch.
Spoken, from the recording's captions. Orion-100B: "a total of fifty GPUs around the world ... for about three days to a billion tokens" (12 min)verified; the write-up says 48 A100s in five US data centres over about two days. Orion-16B: "around two hundred and fifty GPUs, about a quarter of a million GPU hours" (13 min)verified; the deck puts the 250,000 hours on the experiments. Compute "came from Green Compute, it came from Targon," with "around a hundred interruptions" (14 min). The 24%: "fifty percent better than the other state-of-the-art team ... nine providers, three continents, eight different GPU types at one point" (15 min)verified; the captions garble the other team's name. 2T: "a Covenant class node, eight B200s, and we can just stack thirty-two of those" (17 min)verified. Traction: "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" (24 min)verified.
The stage deck, pp. 9 to 13 and 22; the 164, the pilots and the 24% were first reported live by Subnet Summer."Olmo 3 32B, trained on iota: half the price."
Reserved H100 at $3.10 an hour (one-year median across 52 clouds), spot $0.98, idle hosts on Vast $0.18. Ten points of utilisation recovered on a 10,000-GPU site: 8.76m hours and $21.9m a year. Every point of neocloud utilisation orchestrated by iota: $420m a year. The deck's arithmetic on the deck's inputs; none of it is run data. On 3 October only the $3.10 reserved median was found in a public source.
Spoken, from the recording's captions. Utilisation: "from 85% to 95% ... 8.76 million GPU hours a year ... $22 million a year" (22 min)verified. The Olmo case: "it would cost him about a quarter" (23 min)verified; the deck says half the price to the customer.
The stage deck, pp. 19 to 21, citing AIMultiple, FastGPU, GetDeploying and Synergy Research."All of our published results to date are on the legacy code base."
Pretraining "is too small of a market," so the focus now covers RL, fine-tuning and more.
Will Squires on X, 28 September.Recording: the full talk, 30 min 54 s, with English and French subtitles, on the summit's replay pageverified, found 3 October; the 29 September debate, 39 min 46 s, is listed on the same page. On 28 September the only replay was a 38-second highlight. Every spoken line above is from the recording's machine captions, not the audio, so the wording is approximate.