OpenAI’s $750 Billion Bet Is Financed by Georgia Ratepayers Who Never Agreed to It

The Spending Projection That Is Not a Plan

OpenAI’s projected $750 billion in cloud computing spending through 2030 commits the company to an infrastructure scale its revenue cannot plausibly cover fast enough, creating the most over-leveraged capital structure in the AI industry. The numbers have moved sharply in a single year: Sam Altman’s earlier claim of $1.4 trillion in planned spending alarmed investors; CFO Sarah Friar walked it back to $600 billion; the figure has since ratcheted to $750 billion, according to a person with knowledge of the company’s projections. That sequence — the walkback-ratchet pattern — is the single most revealing structural feature in the entire story. Each outward projection functions as a release valve for investor concern rather than a binding constraint on actual commitments. The stock of contractual obligations — $138 billion to Amazon Web Services over eight years including 2 gigawatts on Amazon’s Trainium chip, $250 billion incremental to Microsoft Azure, $20 billion for Project Camellia, and tens of billions more to Oracle and Georgia Power — ratchets upward even when the public number is temporarily pulled down.

This is the bet Altman is making: that the Compute-Capability Reinforcing Loop — computing spend increases model capability, which grows product revenue, which justifies further compute spend — will dominate the CFO Solvency Balancing Loop, which has fired exactly once, producing the $1.4 trillion-to-$600 billion walkback, and has already been overridden. The $600 billion ceiling held for months; the current $750 billion projection shows R1 reasserted past it.

The problem is that R1 has a lag problem that the CFO loop does not.

What the Structural Dynamics Reveal

A systems-dynamics analysis identifies four feedback loops that govern OpenAI’s trajectory. The two reinforcing loops drive spending acceleration; the two balancing loops impose correction only reactively, after projections are made public. The asymmetry is operative on the timescale of the current spending cycle.

The Compute-Capability Reinforcing Loop. Computing spend increases model capability, which grows product revenue or user growth, which increases projected compute demand, which drives further computing spend. This is Altman’s thesis. The contracts with Oracle (6 gigawatts), AWS ($138 billion over eight years including 2 gigawatts on Trainium), Microsoft Azure ($250 billion incremental), and Georgia Power (3.2 gigawatts for 2028–2032) all evidence the loop in operation. The escalation from $600 billion to $750 billion within the year evidences its current dominance.

The Infrastructure-Control Reinforcing Loop. A larger spending programme creates a need for dedicated infrastructure leadership, which generates hires and promotions within the compute organization, which increases operational capacity to execute larger projects, which enables a larger spending programme. This loop imported Brent Mayo from xAI — one of the architects of the Colossus supercomputer in Memphis — as head of data-center build and delivery, reporting to Uday Ruddarraju, OpenAI’s new chief technology officer of computing capacity, himself a former xAI Colossus engineer. The pipeline of xAI infrastructure talent into OpenAI’s compute organization strengthens this loop at the expense of a competitor’s organizational capability.

The CFO Solvency Balancing Loop. Computing spend increases revenue coverage risk, which triggers CFO constraint assertion, which produces a spend-cap adjustment or public walkback, which slows or decelerates computing spend. This loop has fired exactly once: Altman’s $1.4 trillion projection triggered investor concern; Friar walked it back to $600 billion. The subsequent ratchet to $750 billion shows the first reinforcing loop reasserted past the CFO-imposed ceiling. This loop’s corrective force depends on Friar’s organizational standing — and the article reports tension between her and Altman ahead of the planned IPO. It is not structurally guaranteed; it is leadership-dependent.

The Capital-Market Constraint Loop. Spending commitments increase investor or IPO risk perception, which tightens financing terms or compresses IPO valuation, which decreases capital available for further commitments. This loop is latent — no investor walk-away is reported yet. Its activation depends on whether the IPO occurs before or after the heavy-spending years (2028 onward). If the IPO precedes the spending ramp, public-market capital strengthens the first reinforcing loop; if post-IPO scrutiny activates, it amplifies the CFO loop.

The temporal mismatch between the reinforcing and balancing loops is the system’s most structurally exposed interval. The Georgia Power contract covers 2028–2032 — a two-to-six-year lag between commitment and operational capacity. Utility generation build-out imposes a multi-year delay. The $138 billion AWS deal spans eight years; Oracle’s 6 gigawatts are mostly still under development. These are multi-year commitments that compound now, while the revenue to cover them must materialize on a nearer horizon. From 2028 onward, when Georgia Power begins delivering capacity, the payment streams for rental capacity (Oracle, AWS, Azure) and for the self-build capacity (Camellia’s operating costs after construction) arrive simultaneously — rental obligations do not retire when owner payments begin. The combined period of fixed obligations for both rented and owned capacity needing to be serviced together, with no compute output from the new self-built site yet contributing to revenue, is the financial structure’s most exposed interval.

To bound how exposed: the $138 billion AWS commitment, spread over eight years, implies roughly $17 billion annually if evenly distributed. That is the AWS line alone, before Oracle, Azure, Georgia Power, and the engineering headcount to run the models. OpenAI’s revenue has not been confirmed, but published estimates have repeatedly placed it in the vicinity of $5 billion to $15 billion annually for 2025 and projected 2026, without formal disclosure. If even the upper end of that range holds, it does not cover a single year’s payment on the AWS contract. Friar’s concern — that OpenAI might not be able to pay for future computing contracts if revenue does not grow fast enough — is not an abstract caution. It is a statement about a gap that can be approximated but not closed without the revenue figures the company has chosen not to publish.

Who Bears the Risk That Isn’t on the Contract

The stakeholder mapping’s most consequential finding is the one the source article leaves invisible: Georgia ratepayers carry the largest contingent financial exposure in this deal, and they have zero information rights at contract-design stage.

Here is the mechanism. Georgia Power is a regulated monopoly. OpenAI paid a meaningful amount — undisclosed — to reserve 3.2 gigawatts of power for 2028–2032 delivery, an amount the utility uses to justify building new generation capacity. Once that generation is built, it enters the rate base. Ratepayers service the cost of that capacity through their electricity bills for decades, regardless of whether OpenAI scales down its commitment, restructures its contracts, or collapses entirely. If The Spiral scenario unfolds — revenue shortfall combined with build-out failure — Georgia Power’s stranded generation costs are not borne by OpenAI’s shareholders or creditors. They are socialized across captive consumers who had no voice in the decision to build.

This is not a footnote. It is the structural core of the deal. OpenAI is using Georgia’s regulated utility structure as a captive financing mechanism: the company pre-pays a fraction of the capacity cost to trigger generation build-out; ratepayers service the remainder through regulated tariffs; OpenAI gets the infrastructure on terms no unregulated private contract would offer. The risk-transfer architecture is invisible precisely because no consumer advocate, Public Service Commission representative, or ratepayer-group spokesperson is cited in reporting on the project.

The cloud providers — Oracle, AWS, Microsoft Azure — and Georgia Power occupy the high-power, low-interest quadrant of stakeholder positioning: they care about being paid, not about OpenAI’s internal governance. Their leverage is contract-specific. Microsoft’s untimed $250 billion commitment gives it the most flexibility; it can slow-walk capacity delivery without breaching a delivery date. Oracle’s mostly undeveloped 6 gigawatts gives it the most exposure — it needs the anchor tenant to validate its AI infrastructure investment path. AWS sits between, partially shielded by Trainium-specific hardware lock-in.

Absent from the contracting table entirely: Georgia ratepayers, OpenAI rank-and-file employees, AI safety advocates, environmental stakeholders, and future IPO investors. The investors who will price the IPO rely on Friar’s private assurances rather than independent disclosure — the largest single financial stake in the analysis concentrated on a party with no information rights and no seat at the negotiating table. This pattern — capital commitments reported as good news for the locale, while the risk-transfer architecture stays invisible — is a structural feature of AI-infrastructure reporting, not an oversight of this article alone.

Why Stargate Realized Is the Least Probable Scenario

A scenario-planning framework maps two critical uncertainties — revenue growth (low or high) and build-out execution (failed or on-schedule) — into four structurally distinct outcomes. The evidence already in the public record argues against the optimistic diagonal.

Stargate Realized (strong revenue, on-schedule build) requires: quarterly revenue more than 10 percent above analyst consensus for two consecutive quarters; Camellia receiving construction permits by the third quarter of 2026; Georgia Power delivering 3.2 gigawatts on schedule; Mayo’s team replicating Colossus-style rapid deployment at a larger scale; the IPO succeeding at a premium valuation; and the Altman–Friar tension resolving through revenue validation rather than organizational restructuring. That is six independent conditions, each with failure modes, and the worst-case ones compound. A single permit delay, a single revenue miss, a single supply-chain disruption collapses the whole diagonal.

Lean Builder (weak revenue, on-schedule build) is more probable because it requires fewer things to go right: only the build-out must deliver. Revenue can disappoint — the company secures financing from SoftBank or other partners on less favorable terms, Friar’s influence grows, cost oversight tightens, build proceeds in phased tranches. The infrastructure comes online, but the financial model is strained. This is the structural default: the walkback-ratchet pattern shows the CFO loop cannot hold permanently, but the first reinforcing loop’s dominance does not guarantee full financing.

Prestige Fumble (strong revenue, failed build) is equally probable for the opposite reason: revenue can grow strongly, but the company has never built a data center as lead developer at this scale. The Stargate joint venture — a $500 billion joint venture with SoftBank and Oracle announced at the White House in January 2025 — struggled to get off the ground. The 2028–2032 power delivery window is long enough for regulatory friction, grid upgrade delays, supply-chain failures, and organizational turnover to compound. Even strong cash flows cannot buy construction speed beyond what the physical world allows. If Georgia Power announces grid delays of six months or more, or Camellia’s partner selection cycle extends beyond its original window, this is the scenario.

The Spiral (weak revenue, failed build) is the outcome Friar’s private warnings forecast. Two consecutive quarters of revenue missing internal targets by more than 20 percent, combined with Camellia pause or downsizing, triggers a board-level spending cap that curtails Altman’s authority. The $750 billion projection becomes the liability she warned about. Contracts are renegotiated or cut. The Stargate-era vision is retrospectively reframed as the peak of a spending cycle.

The most probable trajectory is a compound of Lean Builder and Prestige Fumble: revenue grows but disappoints against projections, build-out hits delays but does not collapse, the company proceeds in phased tranches under tighter fiscal discipline, and the Georgia ratepayer risk-transfer mechanism absorbs some of the stranded-cost exposure that the corporate structure cannot. This is not a catastrophic outcome — OpenAI survives — but it is not the infrastructure-led competitive advantage Altman is betting on.

The Wild Card That Collapses All Four Scenarios

A competing lab — DeepMind, xAI, or a startup — might develop a training method that reduces compute requirements per capability level by a factor of ten or more. Mixture-of-experts routing, post-training distillation, algorithmic sparsity: any of these could decouple infrastructure scale from model progress. The $750 billion spending projection would become a stranded commitment; physical infrastructure built for current-generation models would be over-engineered for the next generation. The entire scenario matrix, which assumes compute demand remains structurally high, would be invalidated.

This wild card sits outside the matrix because the matrix assumes that infrastructure scale and AI capability remain tightly coupled — that more compute leads to better models. A step-change efficiency breakthrough would decouple the two, making the question of whether OpenAI can build fast enough irrelevant if the frontier requires far less hardware. The first reinforcing loop’s core link — computing spend increasing model capability — would weaken its positive polarity. Altman’s capacity lock-up strategy would face obsolescence from the efficiency side, not the financial or execution side.

The indicator is specific: a competitor publishes a peer-reviewed result or live demo achieving a fivefold or greater reduction in training compute for a given benchmark score, with reproducibility verified by a third party. If that signal appears before 2028, the entire $750 billion projection should be evaluated under the assumption that it will not be fully spent.

What to Watch

For a reader following this story, three observable signals determine which scenario unfolds.

Is revenue consistently beating consensus? If quarterly revenue exceeds analyst consensus by more than 10 percent for two consecutive quarters, Stargate Realized’s revenue condition is met. If revenue misses internal targets by more than 20 percent for two consecutive quarters, The Spiral’s trigger fires. Between those thresholds, Lean Builder is the default.

Are Georgia Power’s grid upgrades on schedule? A delay of six months or more is a Prestige Fumble indicator. A delay of 12 months or more combined with a revenue miss pushes toward The Spiral. The schedule is measurable through Georgia Power’s regulatory filings and the Effingham County permit process.

Is the CFO’s influence growing or shrinking in the public record? If Friar’s spending authority shifts to a board committee with explicit quarterly caps, Lean Builder’s governance condition is met. If Friar takes public credit for cost-control measures, The Spiral is unfolding. If Altman and Friar publicly align on the spending plan in an earnings call, Stargate Realized’s governance resolution is signaled.

The scenario that actually unfolds will be determined not by the spending projection — which is a release valve, not a plan — but by the revenue trajectory, the Georgia Power delivery timeline, and the organizational balance between Altman and Friar. The first of those is the one the company can control least. At the sector level, watch whether cloud providers report data-center vacancies or defer new builds, and whether utilities in Georgia and elsewhere report slower-than-expected hook-up demand from AI tenants — if those signals appear, the structural environment shifts for every major AI buyer simultaneously.

Analytical techniques used in this piece

This analysis applies the methods below. Each links to a short, plain-English explainer you can read and reuse.

Scenario Planning

Builds a small set of distinct, plausible futures to plan against.

Stakeholder Mapping

Charts the parties to a situation — their interests, power, and alignments.

Systems Dynamics (Structural)

Maps a system’s structure — stocks, flows, and the architecture that shapes its behavior.