🟢 [E683] Frontier model training will probably always be domain of hyperscalers, but inference, model tuning etc can run very well on single machine or decentralized group. Bittensor network has pro cards like h200s and some b200s. FOSS training being done by Chinese labs.