AI Compute Rent Has Jumped From $12B to $50B per Gigawatt in Two Years

AI compute economics have changed dramatically in just two years.
The value of available compute capacity is estimated to have risen from roughly $12 billion per gigawatt to as much as $50 billion per gigawatt.
The core thesis is that model capability and token demand are improving faster than the physical supply of chips, power and data-center infrastructure can expand.
If that gap persists, scarce compute capacity could remain extremely valuable.
Contract structure also matters.
$CRWV reportedly signs longer five-year agreements that can generate around $20 billion per gigawatt, while $SPCX can pursue shorter-duration deals at much higher annualized pricing, potentially around $30 billion to $40 billion per gigawatt.
Shorter contracts preserve more exposure to rising market prices, but they also carry greater downside if compute supply eventually catches up.
SpaceX may have an additional advantage because excess capacity could potentially be redirected toward internal AI workloads rather than immediately becoming stranded supply.
That reduces some of the forced-selling risk faced by pure-play infrastructure providers.
The bullish case is straightforward: if AI demand keeps compounding faster than chips, power and data centers can be built, compute rent could continue moving higher.
The main risk is that efficiency gains, new chip supply or slower AI spending eventually narrow that gap.