Elon Musk told a SpaceX earnings call this month that the company will build 10 gigawatts of AI computing capacity by the end of 2027, all of it running on Nvidia's forthcoming Vera Rubin chips. Ten gigawatts is roughly the output of ten nuclear reactors. SpaceX runs about 1.4 today. The number is staggering, and that is exactly the problem: in the AI buildout, the announced figure and the delivered figure have quietly become two different things.
A smaller, more revealing version of the same story played out this month in Armenia. Firebird opened an AI factory in Hrazdan, a genuine site of Nvidia Blackwell hardware, financed like national infrastructure with a $300 million loan from six Armenian banks and anchored by a five-year, $25 million government compute contract. Yet as the tech site TECHi documented, almost every number attached to it refuses to line up. The company markets the first hall as 6,144 GPUs across 15 megawatts; a June government brief described more than 6,000 GPUs and 18 megawatts. The headline promise is more than 70,000 GPUs and 300 megawatts by the end of 2027. Armenia's own prime minister has said the country no longer has the electricity surplus to power it.
That gap between "operational," "installed," and "available" is the real story of AI infrastructure right now. A data center is not a product when the ribbon is cut. Chips have to arrive, clusters have to be commissioned, grid power has to become firm, and customers have to pay for sustained workloads. Each of those is a separate milestone, and the announcements tend to collapse them into one.
The money, at least, is real. Cathie Wood's Ark Invest spent roughly $30 million on Nvidia and TSMC shares in late July, timed to Meta's earnings, where the company raised its 2026 capital budget to as much as $145 billion and called itself "supply-constrained." Investors are treating hyperscaler spending as the surest bet in technology, and the logic holds: more Meta capex mechanically becomes more Nvidia orders, which becomes more TSMC wafers. Demand is not in doubt. What is in doubt is how fast the physical world can absorb it.
The binding constraint is power, and it is starting to bite in ways the public can see. In May, the North American Electric Reliability Corporation issued a rare top-tier alert after events in which more than 1,000 megawatts of data center load dropped off the grid in seconds, the machines riding through a voltage dip on their own batteries while the grid took the hit. Sean James, an Nvidia engineer who spent 25 years building data centers at Microsoft, wrote this month that his real career goal is to "disappear the data center," to make it as unremarkable as a substation. His current designs run coolant hotter than a hot tub to eliminate cooling towers, and treat the buildings as grid assets rather than liabilities.
That is the honest version of the buildout: incremental, power-constrained, and far less cinematic than a 10-gigawatt orbital compute constellation. The gigawatt has become the currency of AI ambition, but the unit that matters is deliverable compute, energised and sold. Firebird could settle its own story tomorrow by publishing four numbers each quarter: commissioned GPUs, energised load, contracted capacity, and the date each was measured. That almost no one in the industry does so is the tell. When the announced number and the real number diverge this far, the announcement is doing work the megawatts cannot yet do.