on iota (sn 9):
iota = core infrastructure for the future of distributed ai training. you don't get to train a competitive decentralized model over the open internet if you cannot orchestrate compute across the world in the cheapest way and with the highest MFU performance. that’s the thesis
they’ve already shown early signs that they’re able to train at the lowest cost/FLOP ever in both centralized and decentralized setting. I operate under Bayesian principles, so I’ve been buying more iota as the results they’ve demonstrated increase the probability that the thesis plays out
for context, iota’s 100B model is trained using single A100s at a cost of $1.4 per hour/participant using spot instances. The previous large decentralized training run (sn 3) was a 70B model and used 8xB200 at $50 per hour
the market is currently pricing iota at $45M next 12 months adjusted valuation
Going to aggressively differ with the characterization of decentralized AI as “full of high flying, fundamentally lacking, narrative driven tokens”.
Yes, for better or worse, people created such tokens in the past.
However, decentralized AI at its innovative core is about decentralizing the AI model supply chain. The fundamental technology for that is decentralized AI training, which is an immensely difficult state-of-the-art advancing problem *in AI*.
The companies that work in this field have to publish counter-consensus papers in prestigious AI conferences, staff frontier PhD level researchers, and by sheer impressiveness push back against global AI noise and skepticism.
Not to mention that they have an insane opportunity: we’re looking at a $500B-1T foundation model revenue market over 5 years. Chinese models are closing and the gap between the closed model and open model frontier is widening. Decentralized AI networks have the opportunity to capture large portions of training and inference infrastructure, while democratizing ownership and revenue distribution of AI.
RWA and stablecoins are worthy fields for investment in crypto in 2026, but their technical innovation pales in comparison to the advancements being performed every month in decentralized AI, as decentralized model parameter count keeps growing.
Dismissing the “AI” portion of crypto as memecoins is wholly insufficient.
you need to own the new means of production. own some mac minis or the new RTX spark in your home. plug them into actual computer (sn 95) and build an internet compute cluster with your friends that also own actual computers. then turn on earning mode on targon (sn 4) to passively monetize your cluster. connect your cluster to iota (sn 9) or pluralis to help train a 100B+ parameter model so you and your friends own a piece of a large decentralized model. then host the model on openrouter so you and your friends get paid to use the model
welcome to post agi
read the world for sale. own the new means of production
I swear am so plugged into this markets that I no longer need to check the prices to know whether we’re up or down on the day. can’t explain it. never before seen levels of locked in. not healthy
@0xSunRun upping the ante after reading what they've already achieved is the only logical next step. bayes would agree and I should stake more tao into sn9
> IOTA (subnet 9) ran a 100B parameter model using 48 single A100-80GB GPUs that were distributed (non-colocated, across multiple providers and multiple datacenters) and connected only over the internet (no datacenter fabric)
> achieved 30.8% average MFU, which is impressive because frontier labs run at 38-50% MFU on average
> achieved roughly 65% of the effective training speed of a comparable co-located datacenter setup
> did it at a 2.5x cheaper per replica than high-end datacenter cluster equivalents
training a 100B model with pipeline parallelism across 16 stages over the open internet at 30% MFU is very impressive. it is one of the more impressive things to come out of the bittensor ecosystem so far
that said, I’d like to see them scale this to heterogeneous/permisionless hardware, run it for much longer and produce an actual competitive model while keeping the cost advantage over frontier models
Today, we are launching the first stage of Project Orion.
Our early pre-training run of Orion-100B achieves upward of 65% of data-center training efficiency on hardware costing a fraction of the price.
Orion-100B is the first proof point for a simple idea: that underutilized compute around the world can be turned into frontier training capacity.
We believe that this work presents, for the first time, an economically compelling case for training large models using distributed approaches.