OpenAI will spend $20 billion on its first self-developed data center campus near the Port of Savannah, taking on construction risk, utility negotiation, and local permitting directly rather than renting capacity from cloud providers. The decision reveals a more fundamental pressure: time is the most expensive ingredient in the company's infrastructure plan, and renting capacity no longer buys enough of it.
The project, dubbed Project Camellia, spans 4.4 million square feet across four buildings within a 2,600-acre industrial hub. Georgia Power will supply 3.2 gigawatts under a 25-year agreement, with the ability to curtail up to 1 gigawatt during peak demand. OpenAI committed to paying full infrastructure and electric service costs, though the dollar amount was not disclosed. The company also secured a 50% tax abatement from the Effingham County Industrial Development Authority.
This is not a real estate deal. It is a time arbitrage. OpenAI's $750 billion spending plan through 2030 means the company cannot wait for hyperscalers to build capacity on their own schedules. By developing directly, OpenAI controls the timeline for power delivery, construction sequencing, and operational readiness. The trade-off is balance sheet exposure and execution risk that a pure tenant would avoid.
The capital implication is straightforward. OpenAI is converting future operating expense into current capital expenditure. Renting capacity from a cloud provider would spread costs over time and keep leverage off the corporate balance sheet. Self-development concentrates risk in a single construction cycle, with no guarantee that power delivery or permitting will hold to schedule. The 3.2-gigawatt commitment from Georgia Power is a 25-year bet on the utility's ability to deliver, and OpenAI's agreement to pay full infrastructure costs means the company is underwriting generation capacity that may not exist yet.
The utility's own numbers illustrate the strain. Georgia Power has proposed $16 billion in new generation capacity to add 9.9 gigawatts over five years. OpenAI's 3.2 gigawatts represents nearly a third of that planned expansion. The utility faces a wave of hyperscaler requests, and the ability to curtail up to 1 gigawatt during peak demand suggests that even with dedicated infrastructure, the grid's constraints will shape operations.
The 50% tax abatement is not the story. It is standard incentive packaging for a project of this scale. The story is that OpenAI accepted the full cost of infrastructure and the full risk of development because the alternative—waiting for someone else to build—was more expensive in time. That calculation matters for every capital partner, lender, and developer watching the data center market.
For lenders, the question is whether they can underwrite construction risk on a tenant that has never developed before. OpenAI has not selected construction or technology partners. The company is soliciting interest from local suppliers. The execution timeline is undefined. A construction lender would need to underwrite not just the sponsor's credit but its ability to manage a process it has never managed before. That is a different risk than financing a build-to-suit for a credit tenant.
For developers, the signal is that the largest AI companies are willing to internalize development functions that were previously outsourced. That reduces the addressable market for third-party development platforms but creates new opportunities for joint ventures, construction management, and power procurement advisory. The developer who can deliver power certainty may be more valuable than the developer who can deliver the cheapest shell.
For owners of existing data center capacity, the implication is that the supply pipeline is about to get much larger and much more direct. OpenAI's $20 billion project is one campus. The company's $750 billion plan implies many more. If every hyperscaler follows the same playbook, the market for leased data center capacity may face a structural shift in supply dynamics.
The project also raises a question about capital allocation discipline. OpenAI's CFO Sarah Friar has reportedly expressed concern about the spending pace ahead of a planned IPO. A $20 billion self-developed campus concentrates risk in a single asset class, a single geography, and a single execution timeline. Public market investors may reward the growth story or penalize the balance sheet exposure. The tension between CEO Sam Altman's spending ambition and the CFO's capital discipline is the kind of conflict that determines whether a company's infrastructure strategy creates or destroys value.
The most revealing detail in the announcement is not the $20 billion figure. It is the 25-year power agreement. OpenAI is locking in electricity costs and supply for a generation. That is not a real estate decision. It is a bet on the long-term cost of compute, the reliability of the Georgia grid, and the company's own ability to operate at a scale that justifies the investment. If the bet pays off, OpenAI owns the cheapest compute capacity in the industry. If it does not, the company owns a very expensive power contract and a half-built campus.
Time is the ingredient that cannot be bought at any price once it is gone. OpenAI is spending $20 billion to buy more of it. The market should watch whether the company can spend it wisely.