A power contract that one of Georgia’s own commissioners called far larger than what the commission normally reviews closed yesterday without a commissioner vote. The Georgia Public Service Commission approved a supply agreement between Georgia Power and OpenAI covering 3,200 megawatts of new generation for a data center campus in Effingham County, in southeastern Georgia near Savannah [DatacenterDynamics, 27 August 2026; Atlanta News First/WTOC, 26 August 2026]. What kind of generation will be built has not been disclosed.

What’s happening

• Delivery is phased between 2028 and 2032. OpenAI pays the full cost of the infrastructure required to serve it and provides up to 1,000 MW of flexible demand response, letting Georgia Power reduce deliveries during statewide peaks. The type of generating capacity “has not been disclosed” [DatacenterDynamics, 27 August 2026; Atlanta News First/WTOC, 26 August 2026].
• Terms are treated as trade secrets. The copy released publicly was fully redacted and stays secret even after execution, per Commissioner Peter Hubbard. Under an April 2025 rule, Georgia Power must file data center contracts above 100 MW thirty days before execution; staff may object, but the elected commissioners do not have to vote to approve [The Current GA, 25 August 2026].
• Staff were prepared to object. At the public committee meeting the day before the 14 August deadline, PSC Director of Utilities Tom Bond told commissioners that “right now, unless the company agrees to certain changes, staff would file an objection.” Georgia Power then extended the review by 12 days rather than let the objection stand and open the contract to debate by the five elected members [Georgia Public Broadcasting, 14 August 2026]. Thirty plus twelve is 42 days of review, desk arithmetic.
• The revised deadline fell on 26 August. The PSC energy committee next meets 27 August, one day later, a sequence Georgians for Affordable Energy founder Patty Durand questioned publicly [The Current GA, 25 August 2026].
• Georgia Power frames the result as a household win. Its large-load contract portfolio is projected to deliver $950 million in annual system savings from 2029, lifting expected savings for a typical residential customer using 1,000 kWh a month to about $15 a month, or $180 a year, up from a previously projected $102 a year [Atlanta News First/WTOC, 26 August 2026].

US angle

Two things happened here and they are worth separating. The first is scale. Hubbard, whose objection triggered the extension, asked where the state would “find three cities’ of Atlanta worth of power customers to backfill a contract like that if it was terminated early” [Georgia Public Broadcasting, 14 August 2026]. The second is process. A load this size, Hubbard noted, exceeds what the commission typically approves inside an integrated resource plan, which is a contested docket with scrutiny and public input [The Current GA, 25 August 2026]. This was not that.

The materials consequence follows from the redaction rather than from the megawatts. Someone is about to order the generating fleet for 3.2 GW, plus the transformers, conductor and interconnection hardware to move it. Gas turbines, new nuclear, and solar paired with storage imply entirely different input baskets and supplier queues, and nobody outside the nondisclosure agreement can size any of them. The procurement gets committed years before an elected regulator, or a supplier planning capacity, learns which one it was.

Brazil angle

Brazil is running the opposite experiment in public. Data center connection requests to the transmission network reached 26.2 gigawatts as of November 2025, more than a quarter of national electricity demand, per the Associação Brasileira de Data Centers and EPE’s transmission volume of the Plano Decenal de Expansão de Energia 2035. Installed data center load in Brazil in 2025 was 800 MW, per a Schneider Electric study with the MDIC [DatacenterDynamics Brasil, 28 May 2026].

That makes this single Georgia contract four times Brazil’s entire 2025 data center load and roughly 12 percent of Brazil’s whole pending connection queue. Desk arithmetic from the figures above.

The contrast that matters is disclosure. In December 2025 the ONS validated 43 Rede Básica access requests totalling 7.3 GW, of which at least 38 were data centers at 7.04 GW, with São Paulo holding 20 requests and 3.9 GW. Decree 12,772 of 5 December 2025 replaced the first-come queue with the Temporadas de Acesso model, requiring financial guarantees, binding schedules and, where demand at a substation exceeds supply, a competitive process whose proceeds return to tariff moderation [DatacenterDynamics Brasil, 28 May 2026]. Brazil also promises ratepayer benefit from large loads, and routes that promise through an auction with published numbers rather than a redacted bilateral contract.

Brazil’s binding constraint is physical rather than procedural. A data center is built in 18 to 24 months; a large transmission line or substation takes 42 to 60 months from tender to energization. ReData, reintroduced as PL 278/2026, conditions its tax relief on 100 percent clean or renewable supply and a water use efficiency indicator at or below 0.05 litres per kWh [DatacenterDynamics Brasil, 28 May 2026], plus at least 10 percent of capacity serving the domestic market and 2 percent of incentivised equipment value going to R&D in Brazil [Capacity, 26 February 2026].

China angle

China already ran the low-scrutiny version and is cleaning up after it. Local governments backed data centers across the country, and Beijing’s own estimate put utilization at 20 to 30 percent. Cancellations of state-backed projects went from 11 in 2023 to more than 100 in the 18 months to July 2025, and MIIT began working with the three state telecoms operators on a national platform to resell surplus compute [DatacenterDynamics, 26 July 2025, citing Reuters]. That is a year-old signal rather than fresh news, but it is the closest available read on what happens when approval is easy and demand verification is not.

What it means

Hubbard’s stranded-asset question is the question China answered expensively. The difference is that Georgia has attached its answer to a single counterparty whose contract nobody can read. The 1,000 MW of flexible demand response is the real hedge, because it is the one term that lets the utility shed exposure without finding a replacement customer. Its price and its triggers sit inside the redaction.

For materials, procurement now runs ahead of disclosure. The Tantalum AI Materials Index stood at 108.8 as of 22 August, with the materials sub-index at 109 and power at 108.3. Those levels reflect orders already placed. Contracts structured like this one are why the order flow behind them is hard to forecast from public filings.

What to watch

1. Generation mix disclosure. Georgia Power has not said what will produce the 3.2 GW. Watch its next integrated resource plan filing and any certification request for new units. Until then the turbine, nuclear and storage supply chains are all sizing against the same undisclosed number.
2. Georgia’s November general election. Two of five PSC seats are to be decided, including Hubbard’s, against former Commissioner Fitz Johnson. Hubbard has said enforcing the cost-shift guarantee “will require three votes.” A Democratic majority would be the first in three decades [The Current GA, 25 August 2026].
3. Brazil’s Temporadas de Acesso first cycle, closing October 2026. It is the filter between mature projects and speculative queue positions in a 26.2 GW pipeline. Whatever share survives is the number that tells you how much of the Global South buildout is real.