OpenAI's Compute Bill Is Now $750 Billion and Stargate Has Stalled
OpenAI's infrastructure commitments through 2030 have reached $750 billion, up 25% from earlier 2026 estimates. The newest piece is Project Camellia: a $20 billion data center campus on 1,400 acres in Effingham County, Georgia, northwest of Savannah, drawing at least 3.2 gigawatts from Georgia Power with electricity expected to start flowing in 2028 and capacity coming online through 2032.
Read the utility side, because that is where the constraint actually lives. Georgia Power intends to serve this with roughly 5.8 gigawatts of new natural gas generation plus grid batteries and solar, expects all of it contracted by end of 2026, and OpenAI alone is about a third of that. Effingham County handed over a 50% property tax abatement for fifteen years. The bottleneck for frontier AI stopped being GPUs a while ago and became interconnect queues, gas turbines and county commissions.
The timing detail everyone will skip past: Stargate appears to have stalled, and OpenAI just hired Brett Mayo away from xAI to run data center construction. Announcing a new $20B campus while the flagship program stumbles and simultaneously poaching a builder is a company discovering that the hard part is not the deal, it is pouring concrete on schedule. First power in 2028 means this capacity serves models that do not exist yet, trained on architectures nobody has published.
For anyone building on agents, the number that matters is not the headline, it is what a commitment this size implies about pricing. $750 billion of fixed cost has to be recovered from inference, and agent workloads are the highest-volume inference there is β a single agent run consumes what a hundred chat turns used to. Capacity coming online in 2028 does not help your token bill in 2026. Between now and then, the people building infrastructure and the people building agents are competing for the same scarce gigawatt.
https://techcrunch.com/2026/07/22/openais-ai-spending-spree-has-ballooned-to-750b/
← Back to all articles
Read the utility side, because that is where the constraint actually lives. Georgia Power intends to serve this with roughly 5.8 gigawatts of new natural gas generation plus grid batteries and solar, expects all of it contracted by end of 2026, and OpenAI alone is about a third of that. Effingham County handed over a 50% property tax abatement for fifteen years. The bottleneck for frontier AI stopped being GPUs a while ago and became interconnect queues, gas turbines and county commissions.
The timing detail everyone will skip past: Stargate appears to have stalled, and OpenAI just hired Brett Mayo away from xAI to run data center construction. Announcing a new $20B campus while the flagship program stumbles and simultaneously poaching a builder is a company discovering that the hard part is not the deal, it is pouring concrete on schedule. First power in 2028 means this capacity serves models that do not exist yet, trained on architectures nobody has published.
For anyone building on agents, the number that matters is not the headline, it is what a commitment this size implies about pricing. $750 billion of fixed cost has to be recovered from inference, and agent workloads are the highest-volume inference there is β a single agent run consumes what a hundred chat turns used to. Capacity coming online in 2028 does not help your token bill in 2026. Between now and then, the people building infrastructure and the people building agents are competing for the same scarce gigawatt.
https://techcrunch.com/2026/07/22/openais-ai-spending-spree-has-ballooned-to-750b/
Comments