Iota SN09 is one of the $TAO Bittensor theses I keep coming back to because the economic question is concrete.
Frontier labs already spend heavily on workload prioritization. Macrocosmos is trying to make that problem permissionless.
The way I understand iota is simple enough. Inference demand creates irregular GPU usage, idle capacity appears, and flexible training can fill part of that gap instead of waiting for perfect dedicated clusters.
To me, Orion-100B is the proof object. A way to inspect the architecture instead of only reading the thesis.
I still want to understand the hard parts. Coordination overhead, data movement, and whether demand repeats after the first big run.
It gives me something to inspect beyond a token chart.
Training spends capital. Inference earns it back. Iota is trying to work between those two clocks.
