OpenAI's Project Camellia Puts a $30 Billion Price Tag on the Race for Compute Sovereignty
A planned 3.2-gigawatt data center campus near Savannah, Georgia crystallizes OpenAI's drive to own its compute infrastructure — even as its CFO expresses uncertainty about whether the company can finance the plan.
OpenAI's announcement of Project Camellia — a data center campus near Savannah, Georgia, carrying a reported price tag of $20 billion to $30 billion and a power capacity of 3.2 gigawatts secured under a long-term agreement through 2032 — is the most concrete expression yet of a company betting its future on owned compute. The project sits alongside OpenAI's expanding Stargate partnership with Oracle, now scaled to 7 gigawatts of planned capacity and more than $400 billion in cumulative investment commitments. Taken together, the two programs signal a fundamental strategic pivot: from dependence on hyperscaler cloud providers toward controlling the company's own compute destiny.
That pivot accelerated sharply over the past year. The Wisconsin Stargate site — a $15 billion facility that turned local farmland into sudden wealth for some landowners — was an early proof of concept. Now, with Georgia under development, a reportedly 10-gigawatt Ohio project in negotiation (where Nvidia is said to be guaranteeing lease terms, though that arrangement has not been officially confirmed), and Cerebras committing 200 megawatts of European compute capacity partly to serve OpenAI's inference workloads by 2027, the buildout has acquired a geographic breadth that resembles national infrastructure planning. OpenAI has simultaneously deepened its silicon strategy: in July 2026 it unveiled a custom AI accelerator called Jalapeño, developed with Broadcom, alongside a stated $200 billion infrastructure investment commitment. The same month, it hired Uday Ruddaraju — formerly a compute executive at xAI — as a dedicated Chief Technology Officer for Compute, a role that signals infrastructure has become a first-class engineering discipline at the company, not merely a procurement function.
The ambition comes with a balance sheet that raises legitimate questions. OpenAI revised its 2030 compute spending forecast upward to $750 billion in late July 2026 — a figure so large that Fortune reported the company's own CFO expressed uncertainty about OpenAI's ability to finance the plan and service the resulting debt. The concern is not purely forward-looking: the company was reportedly spending $1.69 for every dollar it earned as of mid-2026, with some analysts projecting that cash reserves could be exhausted by mid-2027 if that burn rate persists. SoftBank's $40 billion loan, due March 2027, adds a hard liquidity deadline. These pressures have already propagated to counterparties: S&P Global downgraded Oracle's credit rating to BBB-minus — one notch above junk — citing OpenAI as a "key credit risk" on an $85 billion compute contract exposure. Oracle separately faces a reported $100 million annual obligation to guarantee power commitments at the Wisconsin facility.
The financial tightrope grows harder to walk given the competitive environment. Anthropic reportedly reached $47 billion in annualized revenue as of July 2026 — a figure that, if accurate, would represent a genuine revenue challenge to OpenAI's premium position — and its pre-IPO valuation has climbed to approximately $96.5 billion, narrowing the gap with OpenAI's capitalization considerably. Meanwhile, China's Moonshot AI released Kimi K3 in July 2026, claiming performance parity with leading Western models; whether those claims survive independent benchmarking is disputed, but the episode has renewed debate over whether Western compute scale still confers the decisive advantage it once appeared to. Across the industry, AI inference pricing is under sustained downward pressure, compressing the margin economics that underpin the entire infrastructure investment thesis.
Inside OpenAI, power has consolidated around co-founder Greg Brockman, who took on expanded operational authority as the company prepares for a prospective IPO — currently targeted, per multiple reports, for 2027 at a valuation approaching $1 trillion. The departure of co-founder and chief scientist Simo in July 2026 marked another senior leadership transition in an eventful period for the founding cohort. Whether the governance structure emerging from these shifts is suited to simultaneously managing a frontier research organization and a multi-hundred-billion-dollar infrastructure program is a question investors will need to form their own view on.
The bull case for OpenAI's strategy is straightforward: proprietary compute infrastructure, once built, should lower per-unit inference costs below what rented hyperscaler capacity can offer and buffer against supply constraints that have repeatedly throttled output — illustrated most recently in July 2026 when the company was forced to restore usage limits on ChatGPT and Codex following a traffic surge. The bear case rests equally on verifiable data: the company currently loses money on every dollar of revenue; the financing of a $750 billion decadal program is, by the CFO's own account, uncertain; and the competitive moat that supported its original valuation premium is under pressure from well-capitalized Western rivals and from Chinese labs that reportedly achieved comparable model performance at a fraction of the compute spend. Three signals will sharpen the picture over the next twelve to eighteen months: whether OpenAI achieves a credible IPO roadmap that provides the equity capital its infrastructure plan requires; whether the Georgia and Ohio sites reach their stated power capacities on schedule, validating the execution capability of its newly constituted infrastructure leadership; and whether inference pricing stabilizes at a level that makes owned compute economically superior to the on-demand model it is displacing.