OpenAI’s $30B Georgia Campus and the Compute Crunch Ahead

OpenAI’s Project Camellia — announced July 22, 2026 — is a 3.2 gigawatt data center campus in Effingham County, Georgia, with an initial disclosed investment of at least $20 billion and a reported full build-out cost above $30 billion. It is the clearest single data point yet for an uncomfortable thesis: frontier AI compute is the binding constraint of this era, and the supply doesn’t arrive until 2028 at the earliest.

The same week, The Wall Street Journal reported that OpenAI has raised its planned computing-infrastructure spending through 2030 to approximately $750 billion — up from roughly $600 billion earlier in 2026. Anthropic answered the same scarcity from the other direction, adding a fourth silicon platform to its compute base. And on the demand side, GPU rental prices and frontier API pricing have both moved in one direction this year: up.

This post connects those threads into a single argument. We cover what OpenAI actually announced in Georgia, why the three headline dollar figures describe three different things, what the 25-year power agreement reveals about timelines, and — most practically — what agencies and businesses buying AI capacity should do while compute stays scarce through roughly 2028.

On July 22, 2026, OpenAI announced Project Camellia: a data center campus at the Savannah Gateway Industrial Hub in Effingham County, Georgia, near the town of Rincon. The site spans roughly 1,400 acres and is planned as four buildings. It sits inside OpenAI’s broader Stargate infrastructure program — the multi-hundred-billion-dollar US compute-capacity initiative announced in early 2025.

The staffing signal is worth noting: OpenAI hired Brett Mayo, who previously oversaw construction of xAI’s Colossus facility, to lead Camellia’s data center construction. Colossus became the industry’s reference case for compressed build timelines, and hiring its construction lead is a statement about how fast OpenAI wants this campus standing. Even so, no GPU-activation date has been announced, electricity is not expected before 2028, and TechCrunch characterizes the broader Stargate buildout as having faced delays elsewhere — a journalistic read rather than an OpenAI admission, but consistent with how gigawatt-scale construction tends to go.

The most interesting structural detail is the demand-response commitment: OpenAI agreed to reduce its own power draw by up to 1 gigawatt — 1,000 megawatts — during grid peak periods, described in coverage as one of the largest single-facility demand-response commitments in the US. A frontier lab volunteering to shed nearly a third of its campus load on peak days tells you how sensitive utility regulators have become to AI’s grid impact — and how much labs are willing to concede to get interconnection approved at all.

Most coverage of Camellia blurs three dollar figures into one headline. They describe different scopes, and the difference matters if you’re trying to reason about what’s actually committed.

The $20 billion is the initial disclosed investment in the Georgia campus — the number OpenAI put on the announcement itself. The $30 billion-plus is the reported full build-out: OpenAI’s VP of Compute Strategy, Sachin Katti, reportedly told Bloomberg that Camellia’s total cost is expected to exceed $30 billion once financing, infrastructure, and operational costs are fully accounted for. Those are different levels of scope — initial capex versus all-in lifecycle cost — and both are real numbers in the coverage. Keep them distinct.

The ~$750 billion is something else entirely: The Wall Street Journal’s July 22 report that OpenAI has raised its projected computing-infrastructure spending through 2030 to approximately $750 billion, up from roughly $600 billion earlier in 2026, driven by new deals with cloud-computing providers. Coverage notes the figure is roughly equal to Sweden’s GDP. It is planned spending — reported planning guidance, not signed contracts — and the fact that it was revised upward by 25% within a single year is itself the most informative thing about it.

None of that means the buildout won’t happen — the Georgia Power agreement is a signed 25-year contract, and land, tax abatements, and construction leadership are in place. It means the shape of the commitment is asymmetric: the near-term campus is real and funded; the through-2030 aggregate is a moving planning envelope that has grown every time it has been reported. For anyone modeling the compute supply curve, the honest takeaway is that even OpenAI doesn’t know the final number — only that every revision so far has been upward.

The power agreement is where the announcement stops being a real-estate story and becomes an energy-market story. Georgia Power — the Southern Company subsidiary — received Public Service Commission approval in December 2025 to build 9,885 additional megawatts of capacity. OpenAI’s 3.2GW deal represents roughly one-third of that entire approved expansion. One company, one campus, a third of the state’s newly approved dispatchable power.

The generation mix behind the deal is primarily natural gas — roughly 5.8GW of gas capacity — supplemented by grid-scale batteries and solar. OpenAI has stated it will “pay the full cost of the infrastructure and electric-service costs,” per TechCrunch — self-funding the grid buildout rather than spreading it across Georgia Power’s ratepayers generally, and separately claiming local residents won’t see higher electricity bills because of the project. Both are vendor-stated claims that regulators and local press will test over the contract’s 25-year life.

The timeline is the strategic tell. Power arrives in phases between 2028 and 2032. That means one of the largest compute announcements of the summer adds zero usable capacity for at least two years — and doesn’t finish phasing in for six. When the industry’s marginal supply is on a 2028–2032 delivery schedule while demand compounds monthly, the arithmetic for the next couple of years is already settled: scarcity.

The announcement landed hard locally. On July 23 — one day after the news, with residents given roughly 24 hours’ notice — nearly 1,000 people attended a community open house in Effingham County. Protesters carried signs reading “You can’t drink data” and “Say no to data centers,” citing water use, noise, and cost-of-living concerns. Anger was compounded by process: OpenAI acknowledged the deal was negotiated privately with local officials before public disclosure, which residents at the meeting said fueled the backlash.

OpenAI’s responses — a closed-loop water system it describes as using “minimal” water, the self-funded grid costs, the tax base the campus brings — are the standard data-center reassurance package, and they may well hold up. But the pattern is now national: gigawatt-scale AI campuses generate organized local opposition within days of announcement. For the compute supply curve, that’s not a footnote. Community resistance, permitting friction, and utility politics are exactly the forces that turn a 2028 phase-in into a 2029 one — and every quarter of slippage extends the scarcity window this post is about.

Why would a company reporting a $38.5 billion annual loss plan three-quarters of a trillion dollars of infrastructure spend? The answer is visible in this year’s market data: compute scarcity stopped being a forecast and started showing up in prices. The two-year trend of falling frontier-API prices ended around April 2026, when GPT-5.5 launched at $5 per million input tokens and $30 per million output tokens — double GPT-5.4’s per-token pricing. Analysts read the reversal as the market pricing in a real compute-supply constraint rather than a margin play; our Q2 2026 inference pricing matrix tracks how that squeeze already shows up provider by provider.

The rationing is qualitative as well as priced. CoreWeave raised prices over 20% in late 2025 and began requiring three-year contracts from smaller customers — capacity going to whoever commits longest, not whoever bids highest today. One industry analysis claims Anthropic limited access to its newest model to a few dozen vetted partner organizations during a capacity-constrained window this year; that specific characterization comes from a single secondary source and we can’t independently confirm it, but it is directionally consistent with the measured uptime strain and contract-term hardening across the market.

How long does this last? Camellia’s own schedule is the honest answer: meaningful new supply phases in from 2028. Bank of America analysts are cited projecting AI compute demand exceeding supply through at least 2029. Our working assumption for planning purposes is scarcity as the base case through roughly 2028 — which means capacity rationing and price volatility on frontier APIs are features of the next two to three years, not a passing phase.

The same week Camellia was announced, the industry’s other frontier lab showed the opposite answer to the same scarcity — Anthropic added a fourth silicon platform through a multi-gigawatt AMD partnership, which we break down in full in our analysis of the AMD–Anthropic compute-diversification play. Set side by side, the two announcements define the strategic spectrum for securing frontier compute.

Neither strategy relieves the crunch before 2027 at the earliest — Anthropic’s AMD capacity arrives in stages from 2027, and Camellia’s first power lands in 2028. That’s the projection worth internalizing: every gigawatt announced this summer is a 2027–2032 story. Between now and then, the industry runs on the capacity that already exists, plus whatever efficiency gains labs can wring from it. If demand keeps compounding the way OpenAI’s own throughput numbers suggest, the gap between announced capacity and usable capacity is the defining market fact of the next two years — and it lands on buyers as exactly the rationing and price volatility we’re already measuring.

Most businesses reading this aren’t building data centers — they’re buying tokens. If scarcity, rationing, and price volatility are the base case through roughly 2028, the passive strategy of picking one provider and assuming prices only fall is the one position that loses in every scenario. Four moves cover the practical response.

We’ve published the operational detail behind these moves: the routing and cost-discipline playbook covers the routing tiers and caching mechanics, and our buy-vs-rent-vs-cloud GPU decision guide walks the break-even math that rising rental prices keep moving. For teams that want this architected rather than described, our AI transformation engagements build multi-provider routing, spend governance, and capacity planning into production stacks — the boring plumbing that makes a compute crunch a line-item problem instead of an outage.