On July 22, 2026, OpenAI announced Project Camellia: a long-term data center project in Effingham County, Georgia, with 3.2 GW of power contracted from Georgia Power Company, delivered in phases between 2028 and 2032, and a commitment to entirely private funding. For readers tracking AI infrastructure, the announcement is worth a close read not for its scale but for how it converts “how does a large-scale data center earn its social license” into an auditable checklist — power rates, the grid, water, community benefits, and audits each come with a mechanism rather than a press-release promise.
3.2 GW and the Two Layers of “No Pass-Through”
The power contract itself is plain: 3.2 GW, phased delivery, service between 2028 and 2032. The real design work sits in the price guarantee, and it has two layers. The first is contractual: OpenAI will pay the full cost of the infrastructure and electric-service costs required to serve the project, rather than letting the existing rate structure absorb them. The second is regulatory: under Georgia Public Service Commission rules, those costs cannot be passed on to existing ratepayers in the first place. Stacked together, “Georgia households will not see their electricity bills rise because of this project” stops being a verbal pledge and becomes a rule-backed statement.
The symmetry is worth noticing. OpenAI did not ask for a special rate or a subsidy; it pays its own way to buy certainty. Compared with the tension between other large loads and the grid, this is a deliberately cooperative posture — the cost is internalized, and the benefit is that the most common objection (higher bills) is dismantled at the source.
Peak-Shaving by Design: The Data Center as a Grid Participant
The second power commitment is rarer. OpenAI describes Project Camellia as one of the first data centers designed to proactively reduce its own power consumption during high-demand periods — curbing load before residential customers are affected, in support of Georgia Power’s grid reliability. Directionally, this redefines the facility from pure load into a participant the grid can lean on: cooling schedules and the timing flexibility of deferrable batch work are both levers that can move.
The operational details — what triggers a curtailment, by how much, and how it is verified — are not yet specified in the announcement. That is exactly the first question the annual audits should track: a design aspiration to reduce load and a verifiable operating obligation are very different things.
Closed-Loop Cooling: No Water Competition with the Community
The water answer is equally concrete. The cooling system uses a closed-loop design that recirculates water rather than continuously withdrawing and discharging it — the announcement’s analogy is a car radiator: you top off losses, you do not drain and refill. OpenAI also commits that water used will not be diverted from local communities; daily demand shrinks to workplace scale — restrooms, kitchens, cleaning and maintenance — comparable to an office building with the same headcount. Set against the evaporation-heavy cooling controversies that follow data centers, this removes the largest source of water-footprint argument at the drawing board.
$80M in Community Benefits and Up to $71M in Codex Credits
The community package is where the announcement gets most concrete (all figures below are OpenAI’s own claims or estimates, not independently verified as of the announcement date):
| Item | Content | Amount / scale |
|---|---|---|
| Community benefits | Schools, public safety, healthcare, housing, workforce training, veterans, small businesses | $80M (over the project’s life) |
| Codex education credits | Georgia college, community college, and technical school students | Up to $71M |
| Taxes | State and local; expected to be the county’s largest taxpayer | Hundreds of millions (estimate) |
| Jobs | Construction and permanent on-site roles; local contractors prioritized | Thousands (estimate) |
The Codex credits deserve a second look: eligible students receive $100 each through their ChatGPT accounts, usable on Codex, OpenAI’s agentic coding tool built into ChatGPT. For OpenAI it is education spending and ecosystem spending at once — raising the next generation of developers inside its own toolchain; for Georgia students it is a zero-cost entry ticket to hands-on agent work. Beyond the money, OpenAI also commits to prioritizing local contractors and businesses, keeping construction-period value inside the county.
Accountability: Annual Independent Audits and a Community Compact
Commitments only have teeth if something verifies them. OpenAI pledges a publicly available audit every year by an independent firm — so whether the four core promises (rates, grid support, water, community benefits) are kept gets a checkable annual baseline. The procedural side is participatory governance: after the July 23 public open house, resident input and the conversations that follow will shape a Georgia Community Compact that translates community priorities into concrete commitments and accountability measures.
The sequencing matters: hold the open house, gather input, then write the input into a document — rather than finalize first and explain later. The Compact’s final content and its practical force remain to be seen, but the combination of a documented compact plus external annual audits already goes further than most peer projects.
Three Takeaways for AI Infrastructure Watchers
First, the siting logic is itself an argument. The site sits inside the Savannah Gateway Industrial Hub, a pre-existing industrially zoned warehouse development; the comparison OpenAI draws is not “data center vs. empty land” but “data center vs. the warehouses that would otherwise occupy the same site” — less traffic, less water, stronger long-term tax revenue, more lasting on-site jobs. Choosing warehouses as the counterfactual stacks the comparison in the data center’s favor.
Second, the funding structure comes first: the project is entirely privately funded, and OpenAI cites its Abilene, Texas campus already operating at scale as the existence proof, softening skepticism that the plan is more promise than delivery.
Third — the limits. The project is still in early development, with significant work remaining on infrastructure, phasing, design, financing, and the operating model; every figure above is an OpenAI claim or estimate, and “largest taxpayer in the county” is explicitly an expectation. For builders the practical lesson is not the numbers but the pattern: as the politics of siting AI compute keep tightening, writing rate guarantees, water design, education investment, and public audits into the first announcement is shifting from differentiator to entry ticket. If Abilene and Camellia keep delivering, this checklist becomes the baseline future projects are measured against.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.