Verdict: OpenAI has raised its projected compute and cloud infrastructure spending to roughly $750 billion through 2030, a 25% jump from the $600 billion target it set in February 2026 and a figure roughly equal to Sweden's 2025 nominal GDP ($662 billion, per IMF estimates). The bump is driven by a new wave of cloud-computing contracts and a strategic pivot toward owning — not just renting — the data centers that train and run its next models. The single largest new commitment is Project Camellia, a self-designed, 3.2-gigawatt campus in Effingham County, Georgia, where OpenAI has committed $20 billion up front and could spend more than $30 billion at full scale. It is a visionary bet that compute capacity will decide who wins the next decade of AI — and it is also a financial bet so large that OpenAI's own CFO has privately raised concerns about whether revenue can keep pace with the commitments.
TL;DR — Last verified: 2026-07-29
- OpenAI's projected spending on compute and cloud rose from ~$600B (Feb 2026) to ~$750B through 2030, first reported by the Wall Street Journal on July 22, 2026.
- Project Camellia in Effingham County, Georgia: $20B committed, $30B+ at full 3.2 GW, with Georgia Power delivering power in phases from 2028 to 2032.
- The spend trajectory is not a straight line: Sam Altman floated $1.4 trillion, CFO Sarah Friar reset it to $600B, and it has since crept back up to $750B.
- Existing cloud deals — Microsoft Azure ($250B), Oracle ($300B over 5 years for 4.5 GW Stargate), AWS ($138B / 9 years), CoreWeave ($22.4B) — represent the bulk of the cumulative commitment.
- What it means for builders: this is the reason your API bill isn't collapsing, and a reason to think hard about inference-cost hedging now.
- Pricing, capacity, and deal terms change often — last checked July 29, 2026.
Why is OpenAI spending $750 billion on infrastructure?
OpenAI is spending three-quarters of a trillion dollars on compute because frontier AI models require exponentially more training and inference horsepower than the previous generation — and securing land, energy, and GPUs years in advance has become the company's single highest strategic priority. As Sachin Katti, OpenAI's vice president of compute strategy, reportedly told the Journal, the money is earmarked for two things: new cloud service agreements with existing providers (Microsoft Azure, Oracle, AWS, Google Cloud, CoreWeave) and, increasingly, building out OpenAI's own data center footprint so the company controls the design, timeline, and cost.
The shift matters because it reverses OpenAI's previous "rent everything" posture. Until 2025, OpenAI's compute came almost entirely from Microsoft Azure. The new playbook — typified by Project Camellia — has OpenAI acting as the principal designer and developer of its own campuses, hiring infrastructure-execution talent away from rivals, and even assembling land and power before it has named a launch date for a single GPU. The strategic logic: in a world where compute is the binding constraint on intelligence, controlling the factory beats leasing time on someone else's. Or, as Sarah Friar put it in a blog post, "Compute is the scarcest resource in AI."
What is Project Camellia and why does 3.2 gigawatts matter?
Project Camellia is OpenAI's first ground-up, self-designed AI data center campus — a 1,400-acre site within the 2,600-acre Savannah Gateway Industrial Hub, in Rincon, Effingham County, Georgia. OpenAI has contracted with Georgia Power for 3.2 gigawatts of electricity delivered in phases between 2028 and 2032 under a 25-year agreement, and has committed at least $20 billion in the county to qualify for a local incentive package; at the full 3.2 GW build-out, total project costs will exceed $30 billion, per Katti.
To put 3.2 GW in human terms: one gigawatt is roughly the output of a typical nuclear power plant and enough to power about 750,000 U.S. homes. So 3.2 GW is the equivalent of three nuclear plants or roughly 2.4 million homes — reserved for a single, privately-funded AI campus. It is also enough electricity to reshape local grid politics: OpenAI has agreed to provide up to 1,000 MW of "flexible demand response" (letting Georgia Power throttle its draw during peak grid stress), pay the full cost of the electrical infrastructure (state rules bar passing that to existing ratepayers), use a closed-loop cooling system that recirculates water, and commit $80 million in community benefits plus up to $71 million in Codex credits for eligible Georgia college and technical students.
A campus of this size also puts OpenAI in direct competition with other energy-hungry industries — most visibly Bitcoin mining — for the same grid capacity and favorable power contracts, a dynamic already playing out in Texas. A parallel pattern is visible internationally: the data center arms race is reshaping national infrastructure commitments, as seen in India's RMZ–Colt hyperscale deal covering 1.25 GW of Visakhapatnam data centre capacity.
How did OpenAI's spending target get to $750 billion?
The number has bounced around. Here is the trajectory, by primary source:
| Date | Figure | Context | Source |
|---|---|---|---|
| 2025 (early) | ~$1.4 trillion | CEO Sam Altman publicly floated a $1.4T infrastructure ambition — large enough to spook investors | CNBC, Feb 2026 |
| Feb 20, 2026 | ~$600 billion | CFO Sarah Friar reset the record with investors; revenue projected at $280B annually by 2030 to justify it | Reuters, Feb 20 2026 |
| Jul 22, 2026 | ~$750 billion | Revised upward after new cloud contracts and the Project Camellia self-build announcement | WSJ via TechCrunch, Jul 22 2026 |
The CFO's concerns are not abstract. According to the Wall Street Journal's reporting, Sarah Friar has privately raised concerns that OpenAI may not be able to honor future compute contracts if revenue growth does not keep pace with commitments. That is the financial tension underlying the headline number: every new gigawatt signed in 2026 is a payment due in 2028–2032, backed by a revenue line that OpenAI's own projections say must scale from ~$13 billion in 2025 to $280 billion by 2030 (Reuters, Feb 2026). That is a more than 20× increase in five years.
What cloud deals make up the $750 billion?
The $750 billion is cumulative spend through 2030 across a portfolio of agreements. The largest disclosed commitments, by primary source:
| Provider | Commitment | What it covers | Source |
|---|---|---|---|
| Microsoft Azure | $250B | Cloud services; Microsoft receives a 27% equity stake and right to use OpenAI IP; loses right-of-first-refusal | DCD, Oct 2025 |
| Oracle | ~$300B over 5 years | Stargate data center services across sites in Texas, New Mexico, Wisconsin; 4.5 GW agreement | DCD, 2025 |
| Amazon Web Services | $138B over 9 years | Compute including ~2 GW of Trainium3/Trainium4 chips; AWS becomes exclusive third-party cloud for OpenAI Frontier | Investing.com, Jul 2026 |
| CoreWeave | ~$22.4B | GPU cloud capacity across multiple contract expansions (Mar–Sep 2025) | DCD, Sep 2025 |
| Project Camellia (self-built) | $20B committed, $30B+ at full scale | 3.2 GW Georgia campus, 2028–2032 energization by Georgia Power | Bloomberg, Jul 22 2026 |
The throughline: OpenAI is no longer putting all its compute eggs in one hyperscaler's basket. It is deliberately spreading capacity across Azure, Oracle, AWS, CoreWeave, and Google Cloud, and building its own — both to negotiate better unit economics and to avoid being captive to any one provider's price or roadmap. The same horizontally-shared compute approach is now defining the entire AI infrastructure supercycle, as the recent SK-Nvidia $500B deal and other July 2026 moves illustrate across the industry.
Is the Stargate project still on track?
Partly. Stargate — the joint OpenAI/Oracle/SoftBank program originally announced in January 2025 with a $500 billion envelope — is moving forward at its flagship Abilene, Texas site (target 1.2 GW, with multiple buildings reportedly live), but several expansion plans have stalled or been walked back:
- The Texas Abilene expansion from 1.2 GW to 2 GW was paused in March 2026 when Oracle/OpenAI financing talks broke down. The core 4.5 GW partnership agreement remains intact, and the freed-up 600 MW capacity will reportedly be absorbed at other Stargate campuses.
- The UK Stargate site was put on hold in April 2026, with OpenAI citing energy costs as the driver.
- OpenAI also stepped back from selected projects in Norway and has focused capacity on U.S. and scale-advantaged sites.
So when the new $750B headline lands against a backdrop of Stargate pauses, the honest read is: OpenAI is consolidating and re-routing its infrastructure bet — buying cloud where partners can deliver, self-building where partners are too slow, and trimming international ambition while doubling down on U.S. capacity. The strategy mirrors what Anthropic is doing on the chip-supply side, where the AMD $5 billion Anthropic deal represented an explicit attempt to spread Claude's compute footprint past Nvidia's roadmap.
Who is OpenAI hiring to build all this?
A telling signal of how serious OpenAI is about owning the factory: it has hired Brent Mayo, formerly of xAI, where he oversaw the Colossus data center campus in Memphis. In his new role overseeing data center construction and delivery, Mayo is responsible for keeping OpenAI's cloud partners on construction timelines and contributing to the Georgia facility. He reports to Uday Ruddarraju, who was promoted to chief technology officer of computing capacity — a role Russo had held before leaving xAI in July 2025. The hires read like a deliberate talent raid — pulling infrastructure-execution expertise from a rival's flagship project into OpenAI's own build-out, and a sign that in 2026 the data center is the product as much as the model is.
What does $750 billion mean in revenue terms?
The math is what makes this number uncomfortable. If infrastructure spending flows through cost of goods sold, and OpenAI targets a 70% gross margin by 2029 (per an independent analysis by Tomasz Tunguz of disclosed spend), the implied revenue OpenAI needs at its target margins is staggering: roughly $577 billion by 2029 and $983 billion by 2030 (Tunguz, modeling disclosed margins and spend). OpenAI's own internal projection — $280 billion of revenue by 2030 (per Reuters, Feb 2026) — is enormous (a 20× jump from 2025) but still only about 28% of what the spend implies at a 70% margin.
That gap — between what OpenAI says it will spend and what its own revenue projections can plausibly absorb — is the financial heart of the "land grab vs. spending spree" question. Either OpenAI wildly outperforms its own revenue projections, margins climb higher than modeled, or the spend is front-loaded capacity that pays off only if AI demand keeps compounding through the decade. It echoes the same dynamic that already became visible in Google's most recent disclosures: Alphabet's Q2 2026 earnings showed record revenue alongside record capital expenditure, making clear that even the largest incumbents are running harder on the compute treadmill to stand still on free cash flow.
What does this mean for you?
If you are a builder, SMB, or anyone paying per-token for AI inference, the $750 billion bet is the most important denominator in your API bill. Three practical takeaways:
Don't bet on collapsing inference prices. When the lab building your models is committing $750B to keep capacity scarce for the rest of the decade, prices may oscillate but the long-term floor is set by the cost of gigawatts of GPUs coming online — not by a price war. The transatlantic AI compute build-out — see the Microsoft–Mistral sovereign European compute partnership — is one signal among many that hyperscalers are pricing compute like a long-scarce utility, not a commodity input.
Hedge with open-weight or routing. If a single vendor's roadmap can quietly move your unit economics, diversify. Open-weight models like Qwen 3.8 or DeepSeek V4 — and routing layers that send each query to the cheapest model that can do the job — are the realistic hedge.
Build your strategy for the infrastructure-readiness gap. Enterprise AI adoption stalls most often not on model quality but on whether the surrounding infrastructure (cost ceiling, latency budget, data residency, fallback) is ready. OpenAI spending $750B to fix its own gap is a reminder to audit yours. Don't treat capacity as an afterthought, even if your team is never going to sign a gigawatt-scale power contract.
FAQ
Q: How much is OpenAI spending on infrastructure through 2030? A: OpenAI raised its projected compute and cloud infrastructure spending to roughly $750 billion through 2030, up from ~$600 billion earlier in 2026 — a 25% increase first reported by the Wall Street Journal on July 22, 2026.
Q: What is Project Camellia? A: Project Camellia is OpenAI's first ground-up, self-designed AI data center campus — a 1,400-acre site in Effingham County, Georgia (within the Savannah Gateway Industrial Hub). OpenAI has committed $20 billion up front and contracted with Georgia Power for 3.2 gigawatts delivered in phases from 2028 to 2032; at full scale the project could exceed $30 billion.
Q: Is the $750 billion equal to Sweden's GDP? A: Approximately. Sweden's nominal GDP for 2025 is ~$662 billion per IMF estimates (with the 2030 projection around $740B). The $750B figure is roughly comparable — slightly above Sweden's 2025 GDP and slightly below its 2030 forecast.
Q: Why did OpenAI raise the number from $600B to $750B? A: Two reasons: new cloud service agreements with existing providers (Microsoft Azure, Oracle, AWS, Google Cloud, CoreWeave) and a strategic pivot toward building its own data centers — typified by Project Camellia — to control design, timeline, and cost rather than relying entirely on cloud partners.
Q: Has the Stargate project stalled? A: Partly. The flagship Abilene, Texas site remains partially operational, but specific expansion plans were paused: the Abilene 1.2→2 GW expansion (March 2026, Oracle/OpenAI financing dispute), the UK site (April 2026, on hold citing energy costs), and OpenAI stepped back from selected projects in Norway.
Q: Can OpenAI afford $750 billion? A: It depends on revenue growth. OpenAI's CFO has privately raised concerns about honoring future compute contracts if revenue doesn't keep pace (per WSJ). OpenAI's own 2030 revenue projection is ~$280B against an independent model that, at a 70% gross margin, implies ~$983B of revenue would be needed — a gap that means the bet only pays off if AI demand compounds aggressively through the decade.
Discussion
0 comments