OpenAI Georgia Data Center Plan Raises Its Compute Bet to $750 Billion

OpenAI committed $20 billion to start a massive Georgia data center while reportedly raising its projected compute spending through 2030 to about $750 billion. The OpenAI Georgia data center would eventually draw 3.2 gigawatts, turning access to electricity into a central part of the company’s strategy.

The project, called Project Camellia, would sit near Savannah in Effingham County. OpenAI expects Georgia Power to deliver electricity in phases between 2028 and 2032. Several hundred megawatts could become available during 2028, but the full campus will take much longer to complete.

This is not simply another leased cloud facility. OpenAI will lead the site’s design and development, giving it more direct control than at partner-operated campuses. That shift puts the company against a difficult opponent: its infrastructure commitments are expanding faster than the evidence supporting their long-term economics.

OpenAI Georgia Data Center Starts With 3.2 Gigawatts

Project Camellia turns OpenAI from a major cloud customer into the lead developer of one of America’s largest proposed AI campuses.

OpenAI announced the project on July 22, 2026. The site is planned for the Savannah Gateway Industrial Hub, an existing industrially zoned development in Effingham County. That location places the campus near Savannah rather than Georgia’s established data center clusters around Atlanta and Augusta.

The company has contracted with Georgia Power for 3.2 gigawatts of electricity. Delivery will occur in stages from 2028 through 2032, according to OpenAI’s community commitments. A gigawatt measures one billion watts of electrical power and signals the maximum scale envisioned for this campus.

OpenAI has committed $20 billion to begin the development. Sachin Katti, its vice president of compute strategy, said the completed project will likely require more than $30 billion. That larger estimate reflects the difference between the initial commitment and the cost of building the entire campus.

The site’s eventual shape remains unsettled. OpenAI has acquired the land, but it is still selecting a partner to build and operate the facility. The company has not publicly specified the number of buildings, their dimensions, or the final chip configuration.

Those missing details matter because 3.2 gigawatts describes contracted power, not operational computing capacity. Generating electricity, extending transmission infrastructure, constructing buildings, and installing servers are separate processes. Delays at any stage can leave contracted capacity unavailable to AI workloads.

Project Camellia also marks a change in OpenAI’s operating model. At other sites, the company generally rents computing capacity from providers such as Oracle and Amazon Web Services. In Georgia, OpenAI intends to lead the design and development while choosing another company to handle construction and operations.

That arrangement gives OpenAI more influence over power systems, cooling, networking, and deployment schedules. It also exposes the company more directly to permitting, financing, construction, and community relations. Problems can no longer be attributed solely to a cloud landlord.

OpenAI says the campus will use closed-loop cooling, which recirculates water instead of continuously withdrawing and discharging it. The company claims ongoing water demand will resemble an office building serving the same number of employees. That statement has not yet been independently verified at full-campus scale.

Electricity carries a similar promise. OpenAI says Georgia households will not subsidize the project and that it will cover required infrastructure and service costs. Georgia Public Service Commission rules prevent those project-specific costs from being transferred to existing customers, according to the company.

The campus is also designed to reduce its consumption during periods of high demand. Industry reporting says this flexible demand response could cover up to one gigawatt. Flexible demand response means a large customer temporarily cuts electricity use to reduce pressure on the grid.

OpenAI has paired those utility commitments with an $80 million community-benefits program. Proposed uses include schools, public safety, health care, workforce training, housing, veterans’ services, and small-business support. The company also promises up to $71 million in Codex credits for students at eligible Georgia institutions.

These measures show that electricity and local consent have become inseparable parts of AI infrastructure planning. A signed power agreement can secure a project’s technical foundation. It cannot guarantee regulatory approval, construction progress, or lasting political support.

Project Camellia therefore creates the article’s central tension. OpenAI is securing resources years before the workloads exist, based on expectations about future demand and model development. The campus succeeds only if those expectations become durable revenue before its obligations become overwhelming.

Why OpenAI Raised Its Compute Spending Forecast

The reported jump from roughly $600 billion to about $750 billion reflects new cloud agreements and a broader race to reserve scarce capacity.

OpenAI now projects approximately $750 billion in compute spending through 2030, according to reporting based on a person familiar with its forecasts. The new estimate exceeds the roughly $600 billion target reported earlier in 2026.

The change is substantial, but the figures require careful interpretation. They describe projected spending over several years, not cash already paid. They can include cloud services, data center capacity, chips, networking, and other infrastructure obligations with different delivery schedules.

In February, a source familiar with OpenAI’s plans told Reuters that the company expected roughly $600 billion in compute spending through 2030. That projection followed concerns about Sam Altman’s earlier discussion of $1.4 trillion in infrastructure commitments.

The $600 billion figure appeared to narrow OpenAI’s message. It gave investors a shorter time horizon and a clearer distinction between expected spending and broader commitments. Five months later, the reported forecast has climbed by roughly $150 billion.

New cloud agreements help explain the increase. OpenAI has contracted with Oracle for six gigawatts of data center capacity, most of which remains under development. It also expanded a multiyear agreement with Amazon Web Services and made an additional cloud commitment to Microsoft Azure.

Those agreements represent different forms of access rather than one unified construction program. Some involve rented chips, while others reserve power or future cloud capacity. The OpenAI Georgia data center adds another model, with OpenAI directing the development and selecting an operating partner.

This diversification reduces dependence on any single cloud company. It also makes OpenAI’s infrastructure portfolio harder to evaluate. Each contract can have different payment terms, utilization requirements, delivery conditions, and cancellation protections.

The company is reserving capacity because advanced AI requires large clusters for two distinct tasks. Training builds or updates models using extensive computing resources. Inference runs those models whenever users send prompts, generate media, or delegate work to an agent.

Inference can become more expensive as usage grows, even when each individual request becomes cheaper. Lower costs can attract more users and encourage more complex workloads. That effect can raise total demand rather than reducing it.

OpenAI’s strategy therefore depends on more than building larger training runs. It anticipates that AI products will handle a wider range of recurring work. Coding agents, research systems, enterprise assistants, and media-generation tools can consume computing resources continuously.

The company also wants capacity ready before demand arrives. Data centers cannot be ordered and delivered like ordinary software. Power generation, transmission connections, specialized cooling, and server deployment can take years.

Waiting for proven demand would reduce financial risk but create a supply risk. A competitor with reserved electricity and installed chips could train models sooner or serve more customers. OpenAI has chosen to accept greater financial exposure to reduce that operational constraint.

That decision pressures the entire AI supply chain. Cloud providers must finance construction tied to long-term contracts. Utilities must add generation and transmission. Chipmakers must plan production around capacity that may not become operational for several years.

The pressure also reaches OpenAI’s customers. Infrastructure costs ultimately require revenue from subscriptions, enterprise contracts, advertising, developer usage, or other services. If demand does not scale quickly enough, providers can raise usage charges, limit expensive features, or prioritize higher-margin customers.

A June 2026 analysis from J.P. Morgan treated frontier-lab forecasts for positive cash flow as speculative and subject to revision. That warning captures the problem surrounding every large compute projection.

OpenAI can negotiate access to power years ahead. It cannot contractually guarantee that customers will generate enough revenue to support the resulting capacity. The move from $600 billion toward $750 billion makes that gap harder to ignore.

The Real Contest Is OpenAI’s Ambition Against Its Economics

OpenAI is betting that future AI usage will grow fast enough to justify infrastructure commitments made before demand, financing, and construction are settled.

It is tempting to frame the project as another contest between OpenAI, Anthropic, Google, Meta, and xAI. Those companies certainly compete for chips, electricity, engineers, and customers. Yet that comparison misses the more important opponent.

OpenAI’s primary challenge is converting an unprecedented infrastructure build-out into sustainable economics. A rival can delay a product or lose a benchmark. OpenAI cannot easily reverse a power contract, construction schedule, or long-term cloud obligation after partners have committed capital.

The company has evidence supporting its confidence. OpenAI’s 2025 revenue reportedly reached $13 billion, above its earlier $10 billion projection. It expects consumer and enterprise businesses to contribute heavily to future growth.

However, revenue and infrastructure commitments operate on very different scales. Even rapid revenue growth does not automatically produce positive cash flow. Model training, inference, staffing, data acquisition, and product operations continue consuming money alongside construction.

Inference creates the most persistent pressure. Every active user can trigger recurring computing costs, unlike traditional software whose marginal delivery cost is often low. More capable models may also encourage longer prompts, deeper reasoning, and more tool calls.

Agents intensify that pattern. A conventional chatbot might answer one request with one response. An agent can search, write code, inspect files, call services, and revise its work across many computational steps.

Enterprise adoption can magnify the workload further. A large organization may run AI across customer support, software development, document analysis, security, and internal operations. Reliability requirements can demand extra capacity, redundancy, and low-latency regional deployment.

OpenAI is therefore not building only for more ChatGPT conversations. It is preparing for AI systems that operate repeatedly across business processes. If that transition occurs, today’s data center commitments may look like early capacity reservations.

If the transition slows, the same commitments become a burden. Customers may resist usage-based costs. Enterprises may limit agents because of security or accuracy concerns. Efficiency gains may also let competitors serve similar workloads with less infrastructure.

The company’s choice to lead Project Camellia’s design adds another layer. Direct control can improve scheduling and technical coordination. It also requires expertise that cloud providers have developed over decades.

OpenAI has recruited leaders with large-scale construction experience. Brent Mayo, who helped oversee xAI’s Colossus facility in Memphis, reportedly joined OpenAI as head of data center build and delivery. Uday Ruddarraju, another former xAI infrastructure executive, oversees computing capacity.

Those hires signal urgency. They also reveal that execution now extends beyond model research and software engineering. OpenAI must coordinate utilities, developers, equipment suppliers, local officials, and financing partners across multiple projects.

The history of Stargate explains why. Announced in January 2025, Stargate initially envisioned up to $500 billion in American AI infrastructure investment over four years. Its early structure relied heavily on partners, including SoftBank and Oracle.

That effort developed more slowly and unevenly than its headline suggested. OpenAI subsequently moved toward greater internal control. Project Camellia is the clearest expression of that adjustment because OpenAI owns the land and leads development.

This is a strategic reversal in operating responsibility, not a retreat from scale. OpenAI still depends on cloud and construction partners. It now wants greater authority over how their capacity is designed and delivered.

Competitors are pursuing their own paths. Google can combine model development with its cloud business, custom chips, and long-standing infrastructure organization. Amazon can spread investment across AWS customers while supporting Anthropic through cloud capacity and custom silicon.

Microsoft has a global cloud platform and an established enterprise distribution channel. Meta funds AI infrastructure using cash generated by advertising. xAI has emphasized fast physical construction and close integration between model teams and data center operations.

OpenAI lacks several of those cushions. It does not own a mature public cloud, a global advertising network, or a diversified consumer-hardware business. Its infrastructure obligations depend more directly on AI products, external financing, and partner relationships.

That does not make failure inevitable. It makes execution unusually sensitive to demand assumptions. OpenAI must grow product usage while improving unit economics and securing funding, all before large portions of its reserved capacity arrive.

The company’s projected compute spending is thus a strategic forecast disguised as a budget. It predicts that AI will become common enough to support industrial infrastructure at a scale associated with major utilities and cloud platforms.

Project Camellia tests that forecast in physical form. Its first power deliveries are expected in 2028, with full delivery scheduled by 2032. By then, the market will know far more about agent adoption, model efficiency, and customers’ willingness to pay.

The financial commitments arrive earlier than that clarity. This timing is the core risk. OpenAI is using capital and contracts to pull future capacity forward while relying on product adoption to catch up.

What the $750 Billion Forecast Does Not Guarantee

A rising compute forecast demonstrates appetite for capacity, but it does not guarantee financing, completed infrastructure, reliable demand, or acceptable community impact.

The first uncertainty is whether the $750 billion estimate represents a stable plan. OpenAI’s publicly discussed infrastructure numbers have shifted several times. Different figures can cover different periods, agreements, and definitions of compute spending.

That makes direct comparisons difficult. A long-term commitment can include conditional payments or capacity that has not been built. A spending projection can change as contracts, financing, product demand, or delivery schedules change.

The reported increase from $600 billion should therefore be read as a current internal expectation. It is not an audited liability schedule. OpenAI has not published a complete contract-by-contract breakdown supporting the total.

The second uncertainty is financing. Project Camellia will be privately funded, according to OpenAI, but the company has not detailed the campus’s capital structure. It is still selecting the entity that will build and operate the site.

The completed campus could cost more than $30 billion. OpenAI’s initial $20 billion commitment does not settle who supplies the remaining capital, how risk is divided, or when money becomes due.

Cloud providers and infrastructure partners can finance portions of development. Investors can supply equity or debt. Yet every arrangement transfers the economic burden rather than eliminating it.

The third uncertainty is grid delivery. Georgia Power has agreed to provide 3.2 gigawatts between 2028 and 2032. Meeting that schedule requires generation, transmission, substations, equipment, permits, and construction labor.

A contract establishes intent and commercial responsibility. It cannot eliminate supply-chain delays or regulatory disputes. The power must reach the campus at the required reliability before servers can use it.

OpenAI says it paid Georgia Power a meaningful amount to reserve capacity. That payment gives the utility confidence to invest in supporting infrastructure. The company has not disclosed the payment or the conditions attached to it.

The fourth uncertainty involves community protections. OpenAI promises that households will not pay higher electricity rates because of the project. It also says the closed-loop cooling design will keep ongoing water use low.

Those commitments are important, but their implementation will require transparent measurement. An independent auditor is expected to review compliance. Residents will need clear baselines for electricity costs, water withdrawals, tax benefits, and construction impacts.

The OpenAI Georgia data center arrives during growing political resistance to data center development. Communities across the United States have questioned electricity demand, water use, tax incentives, noise, land conversion, and limited permanent employment.

OpenAI appears to recognize that opposition. It began meeting local officials, schools, and community groups before finalizing every project detail. A public open house was scheduled for July 23, one day after the announcement.

The company intends to create a Georgia Community Compact documenting local commitments and accountability measures. That process can produce enforceable expectations. It can also reveal disagreements that polished announcements do not capture.

OpenAI argues that the project will produce less traffic and use less water than the warehouse development previously planned for the site. It also expects higher long-term tax revenue and more durable on-site employment.

Those comparisons remain company projections. The final operating model, staffing, cooling design, and tax treatment will determine actual outcomes. Changes during construction could weaken assumptions presented at the announcement stage.

The fifth uncertainty is utilization. A data center can become operational without operating near full capacity. Chip availability, networking constraints, model schedules, and demand can all affect how much electricity the campus consumes.

Underutilization would damage the project’s economics. Excess demand would create another problem because OpenAI would need additional capacity beyond its contracts. The company is trying to reserve enough infrastructure without committing to more than the market can support.

Model efficiency complicates that calculation. Better algorithms and chips can reduce the resources required for a specific task. However, cheaper inference often encourages more usage, larger models, or new applications.

No one can yet know which effect will dominate through 2030. OpenAI’s forecast assumes total demand will remain extremely high despite efficiency gains. That assumption deserves scrutiny rather than automatic acceptance.

The sixth uncertainty involves commercial behavior. Infrastructure spending can influence product design and pricing. Companies carrying large fixed obligations have incentives to increase usage, create premium services, and move enterprise customers toward recurring consumption.

That pressure does not mean every product decision will serve infrastructure repayment. It does mean customers should evaluate AI services with operating costs in mind. Generous access today may not predict contract terms several years later.

Developers should watch usage limits, latency, model retirement schedules, and token charges. Enterprise buyers should examine portability, data governance, and contract protections. A provider’s infrastructure scale matters less when customers cannot move critical workloads.

Knowledge workers face a related issue. Greater capacity can make agents faster and more available, but it does not guarantee accuracy or trustworthy automation. Infrastructure expands what a system can attempt, not whether every result deserves reliance.

The $750 billion forecast should therefore be treated as evidence of OpenAI’s conviction, not proof of market demand. Project Camellia gives that conviction a location, a utility contract, and a construction timeline.

Three Signals Will Decide Whether the Bet Works

Power delivery, revenue quality, and binding community accountability will show whether OpenAI’s infrastructure strategy is becoming durable or merely larger.

The first signal is Georgia Power’s delivery schedule. The most important milestone is not another announced campus. It is whether Project Camellia receives its first several hundred megawatts during 2028 as planned.

Before then, readers should watch for approved generation and transmission investments, named construction partners, permits, and a detailed campus design. Each step would strengthen confidence that the 3.2-gigawatt contract can become operating capacity.

Delays would weaken OpenAI’s strategy in two ways. They would postpone access to computing resources while leaving planning and financing costs in place. They could also push OpenAI toward more expensive temporary capacity from cloud partners.

The flexible-demand commitment deserves equal attention. OpenAI says the site can reduce consumption before residential customers face disruption during periods of high demand. Technical filings should eventually explain how much load can be curtailed and for how long.

Independent performance data would make that promise credible. Without it, flexible demand remains a design claim. Grid reliability depends on predictable operating behavior, not broad assurances.

The second signal is revenue quality relative to compute obligations. Reported revenue growth alone will not settle the question. Investors and customers need evidence that OpenAI can improve margins while serving increasingly compute-intensive workloads.

Watch the relationship among enterprise revenue, consumer subscriptions, inference costs, and cash spending. Rising revenue that requires proportionally greater compute does not solve the underlying problem. The critical measure is whether each unit of infrastructure supports improving economic returns.

OpenAI’s planned public offering could bring additional transparency. Formal disclosures might clarify contractual commitments, payment schedules, concentration among cloud providers, and dependencies on external financing. They could also show how management defines the $750 billion estimate.

Those disclosures would strengthen the bet if OpenAI demonstrates manageable obligations and improving unit economics. Repeated revisions without clearer definitions would weaken it. So would evidence that revenue growth depends on deeply subsidized usage.

The third signal is the Georgia Community Compact and its verification process. The final document should convert promises into measurable obligations covering rates, water, employment, taxes, community funding, and demand response.

An independent auditor must have access to meaningful data. Public reporting should use clear baselines and regular schedules. Residents need a way to identify missed commitments and understand the remedies available.

A strong compact would make Project Camellia a useful model for future developments. It could show that large AI campuses can negotiate responsibilities before construction rather than after conflict begins.

A vague or weakly enforced compact would have the opposite effect. It would reinforce concerns that community programs cannot offset electricity, water, and land-use risks. That outcome could slow approvals for OpenAI and its competitors elsewhere.

These three signals are connected. Power delivery determines whether the campus can operate. Revenue quality determines whether OpenAI can afford it. Community accountability determines whether similar projects retain political permission to proceed.

The OpenAI Georgia data center will not resolve those questions immediately. Its planned timeline stretches beyond the current product cycle, and much of its capacity will arrive after today’s models have been replaced.

That delay is precisely why the project matters now. OpenAI is making infrastructure decisions before it knows which products, models, and business structures will dominate in 2030.

The company has chosen scale as protection against a future compute shortage. The risk is that this protection becomes an obligation larger than demand can support.

Developers, enterprise buyers, and AI users should follow delivery milestones rather than announcement totals. Ask whether new capacity improves reliability, lowers real workload costs, and supports products customers continue using without subsidies.

The reported $750 billion compute forecast makes OpenAI’s direction clear. What happens next depends on execution that cannot be measured through headline commitments alone. Watch the first power deliveries, the economics behind usage, and the enforceability of Georgia’s community compact.