I think scarce power plus contracted capacity ultimately holds the strongest pricing power.

Nebius shows that machine hours can command extraordinary prices when GPU capacity is tight. Management even says it could sell all its 2027 capacity at current terms, while Q2 revenue surged 454%. But compute pricing is vulnerable as GPUs improve and competitors add capacity.

CoreWeave’s $104bn backlog offers better visibility, but it still carries enormous capex, financing and customer-concentration risk.

Riot is the interesting third model. A 20-year, 191 MW contract worth about $9.1bn locks monetisation to something AI cannot easily manufacture: power-connected data-centre capacity.

My ranking: power/capacity > contracted compute > spot machine hours for durable pricing power. GPUs depreciate quickly; megawatts in the right place increasingly look like the scarce commodity.