The same H100 graphics card can cost you $0.80 or $4.20 an hour in 2026. Same card, same performance. The only thing that changes is where you run it. There are four paths. On-demand cloud is the default and the priciest, around $3 to $4 an hour, and you pay that premium even in the hours your hardware sits idle. Specialist GPU clouds rent the identical card for roughly $1.49 to $3.00. GPU hosting gets you dedicated capacity on a committed term, cheaper because you are not paying for flexibility you never use. Buying and colocating is cheapest per hour, typically 40 to 60% below hyperscaler cloud, with break-even around 6 to 18 months. The rule is simple. Running hard and steady for a year or more? Own it. Spiky or still experimenting? Rent it, and do not let anyone talk you out of that. Here is the part most people miss: the real bottleneck is not chips, it is power. A normal rack handles about 5kW. A modern AI rack can need 100kW or more with liquid cooling. Most facilities simply cannot do it. Bitcoin miners spent a decade building exactly that kind of capacity, which is why the same buildings now run AI.