IOTA (IOTA)

Cross-border paperless trade in Asia Pacific still sits at around 42%.
The appetite is there.

The next step is shared infrastructure every party can trust - the kind IOTA is built to provide.

A more detailed answer to @travisk's objection on the latest @theallinpod that decentralized training of AI is not doable because of bandwidth:

Traditional data-center training uses high-bandwidth interconnects (NVLink/Infiniband) and keeps everything co-located because communication is a huge bottleneck.

IOTA (Bittensor Subnet 9) makes decentralized training feasible through several innovations:

- Memory scaling: No single device needs the full model in VRAM. Capacity scales with the sum of participants’ memory, not limited by the weakest one.

- Communication efficiency: Heavy activation compression + efficient all-reduce + concurrent compute/communication + stochastic routing for fault tolerance. The run achieved >30% MFU (peak sustained ~38%), roughly 65% of comparable co-located datacenter speed.

- Fault tolerance & decentralization: Handles churn, variable bandwidth/latency, and unreliable peers via thresholds, redundancy in all-reduce, and randomized pathways. An orchestrator coordinates high-level triggers, but actual compute and much of the data flow is peer-to-peer.

- Incentives: Bittensor’s blockchain layer rewards contributors fairly and granularly based on verifiable work, discouraging hoarding or free-riding.

- Practical results: Orion-100B trained ~1.1 billion tokens over ~2 days on these 48 distributed GPUs at a much lower cost than centralized.

This demonstrates that frontier-scale training over commodity internet is becoming technically and economically viable.