Analysis
OpenAI unveiled Project Camellia on July 22: a 3.2-gigawatt, 1,400-acre data center campus in Effingham County, Georgia, part of the Savannah Gateway Industrial Hub, with $30 billion in total planned spending. OpenAI has already committed $20 billion of that and is now looking for partners to finance and develop the remaining phases, with several hundred megawatts expected online starting in 2028 and full build-out continuing through 2032.
The project follows a now-familiar pattern for OpenAI's infrastructure strategy: pair a headline-grabbing capacity number with an aggressive community-relations push designed to head off the local opposition that has slowed data center projects around the country over water and power concerns. OpenAI is promising 400 long-term jobs, $80 million in direct community benefits, and $71 million in Codex credits for Georgia students, alongside a closed-loop water cooling system and a commitment to throttle its own power draw during high-demand periods before residential customers are affected.
The competitive backdrop is a capex race that shows no sign of slowing. Microsoft's capital spending is up 84% year over year, Amazon is on pace to spend roughly $200 billion in 2026, and hyperscalers collectively account for about 40% of the nearly $489 billion in AI-related debt issued industry-wide this year. Georgia joins a growing list of AI infrastructure hubs, competing directly with Microsoft, Amazon, Google and Meta's own multi-gigawatt campus announcements, and with Nvidia's Vera Rubin NVL72 racks already ramping at CoreWeave, Google Cloud, Azure and Oracle Cloud Infrastructure.
“The competitive backdrop is a capex race that shows no sign of slowing.”
The numbers only make sense against how compute-constrained the industry actually is. Power ordered today doesn't arrive until 2028 at the earliest -- which is exactly why Georgia's site was chosen for pre-existing industrial zoning rather than raw land, and why Google has reportedly had to ration Gemini access to Meta and Moonshot has hit its own capacity limits training Kimi K3. Compute scarcity, not model capability, is now the binding constraint across the frontier lab landscape, and no amount of fresh capital changes that on a sub-two-year timeline.
For VCs, the signal is less about OpenAI specifically and more about where the next wave of infrastructure-adjacent opportunity sits: power procurement, grid interconnection expertise, water-efficient cooling, and the specialty construction and materials suppliers who benefit regardless of which hyperscaler wins the underlying AI race. LPs should also note the financing structure -- OpenAI funding a third of this itself and syndicating the rest signals data center debt and structured infrastructure finance is becoming as important a venture-adjacent category as the models themselves.
The risk is that this is exactly the kind of forward commitment that looks brilliant if AI demand keeps compounding and reckless if it plateaus. OpenAI's compute spending commitments have reportedly surged toward $750 billion in aggregate across its various infrastructure deals; a demand air pocket, even a temporary one, would leave a lot of half-built gigawatts and a lot of debt-financed hyperscalers exposed at the same time.
Watch whether OpenAI successfully lines up outside financing partners for the remaining two-thirds of the $30 billion, whether Effingham County's local opposition groups escalate legal challenges the way Starbase's neighbors have against SpaceX, and whether this accelerates a broader wave of Southeast U.S. data center announcements chasing the same cheap land and power access.
