Stargate's Five Gigawatts: The Real Math of AI Data Center Buildouts

01What Stargate Actually Is

The Stargate Project is an American artificial intelligence infrastructure joint venture announced at the White House in January 2025, with OpenAI, SoftBank, Oracle, and the Emirati investment fund MGX as its anchor partners. The headline commitment was attention-getting by design: up to 500 billion dollars invested over four years, with an initial deployment of roughly 100 billion dollars, to build data centers for AI training and inference across the United States. SoftBank's Masayoshi Son chairs the venture, with SoftBank handling financing, OpenAI as the operational anchor tenant, and Oracle and NVIDIA among the technology and silicon suppliers.

The context was political as much as technological. The announcement positioned American AI buildout as a national project, with the new administration taking credit for private capital that — as critics quickly noted — was largely money already committed or already spent elsewhere. That blending of private investment and public relations is part of why the project has attracted a persistent strain of skepticism from its first day, a skepticism this article treats as a serious analytical position rather than a fringe complaint.

It is worth being precise about what the venture is not. Stargate is not a single building, not a single power contract, and — as of mid-2026 — not a fully financed, fully sited, fully energized five-gigawatt fleet. It is a portfolio of announced projects with varying degrees of commitment: some under construction, some leased and planned, some little more than a location and a press release.

02The Announced Numbers: Sites, Dollars, Gigawatts

The scale figure OpenAI uses is about five gigawatts of planned compute capacity once the announced portfolio is complete — roughly the power draw of a mid-sized city's worth of data centers. The anchor site is Abilene, Texas, a former oil field town where Oracle is building out a campus reported at around 1.2 gigawatts of capacity for OpenAI's workloads, with the first buildings energized in 2025 and additional phases planned. Beyond Abilene, the announced list includes Mesa, Arizona; Doña Ana County, New Mexico; and Lordstown, Ohio, alongside additional Texas sites in the Houston area and elsewhere.

The money side is structured in tranches. The initial 100 billion dollar deployment is real in the sense that construction and hardware purchase orders exist against it, but the full 500 billion figure is a target contingent on financing, demand, and — most consequentially — on power. SoftBank has raised capital against the venture, and OpenAI has separately signed multibillion-dollar compute deals with NVIDIA and AMD, some of which are effectively circular: payments flowing between the anchor partners with equity stakes in each other.

Announced megawatts, in other words, are not the same animal as energized megawatts. The gap between the two is where most of the interesting engineering — and most of the interesting skepticism — lives.

03Skepticism and the Hoax Framing

The framing video for this article, from the Wall Street Millennial channel, is titled bluntly: the Stargate data center hoax. That is the sharp end of a skepticism that ranges from Wall Street analysts to energy journalists to semiconductor supply chain watchers. Their argument, reduced to its strongest form: the announcement was engineered for political theater — a headline about 500 billion dollars of investment released alongside an executive order on AI, inflating money that was already committed elsewhere and dressing a portfolio of speculative sites as a shovel-ready national project.

The specific charges are worth listing honestly. Announced megawatts vastly exceed the capacity that utilities have actually agreed to interconnect. Some sites in the original fanfare had no power agreements, no permits, and in some cases no signed land deals at announcement time. The financing structure leans on partners who are also each other's customers and vendors, which critics argue inflates the apparent scale of the commitments. And the demand side rests on OpenAI's projections of compute hunger from models that do not exist yet, backed by revenue that does not exist yet.

The counter-argument deserves equal honesty. Stargate is not vaporware: buildings are standing in Abilene, construction crews are working in multiple states, and OpenAI signed a landmark multi-gigawatt capacity agreement with one of the largest US utilities in 2025 to route dedicated power to data centers. The reasonable synthesis is not "real" or "fake" but a maturity question: which parts of the portfolio are financed and energized, which are merely leased, and which are a site name on a slide. The video's hoax framing overcorrects; the boosters' renderings overpromise; the truth is a spreadsheet with very few green rows.

04The Power Constraint Nobody Can Negotiate With

Every gigawatt of AI compute must arrive as electrons, and the American grid is the one counterparty in this story that does not read press releases. The numbers that frame the squeeze: US data centers consumed roughly 176 terawatt-hours in 2023 — about 4.4 percent of national electricity use — and the Department of Energy's Lawrence Berkeley National Laboratory, in its December 2024 report, projected that share could reach 6.7 to 12 percent by 2028 depending on the scenario. Grid planners across multiple states now name data centers as the primary source of forecast demand growth.

The interconnect queue is the choke point. Requests to connect large loads to the transmission system now routinely wait years, and large industrial interconnections in congested regions are the slowest of all. A developer can pour a concrete pad in twelve months; the substation, transmission tie, and utility studies needed to power it take far longer. This is why announcements consistently outrun energization: construction time is measured in quarters, and grid time is measured in years.

There are workarounds, each with costs. Gas turbines can be installed behind the meter, converting a grid constraint into a fuel-logistics and permitting problem — but emissions rules and turbine supply complicate that at gigawatt scale. On-site generation and fuel cells fill gaps but not baseload deficits. Some operators have resorted to diesel backup running during grid stress, which is legal for reliability purposes in some jurisdictions and quietly controversial everywhere. The physics, not the press release, sets the schedule.

05Grid Interconnects, Transformers, and the Multi-Year Wait

The transformer shortage deserves its own section because it may be the single least negotiable bottleneck in the entire buildout. Large power transformers — the massive units that step transmission voltage down to something a campus can use — carry lead times that have stretched to two to four years in recent procurement cycles, with domestic manufacturing capacity limited and global demand far exceeding supply. A data center developer can expedite chips, expedite steel, and expedite labor; a transformer order placed in 2026 may arrive in 2029.

The interconnect process stacks on top. A large-load interconnection study must confirm that the surrounding grid can absorb the draw without violating reliability standards, and where it cannot, the customer pays for network upgrades — transmission lines, substations, reactive power — that must themselves be engineered, permitted, and built. Utilities have grown wary: several have begun requiring data center customers to commit to taking power even if their project is canceled, so that speculative AI demand does not leave ratepayers holding the cost of unused infrastructure.

The result is a widening gap between silicon schedules and power schedules. GPU supply, once the binding constraint, has loosened considerably through 2026 as accelerator manufacturing has scaled. Power is now the long pole — and unlike chips, you cannot fab your way out of it in eighteen months.

06What Gets Built First: Shell vs Silicon

Because power arrives in phases, campuses get built in phases. What goes up first is the shell: site work, foundations, steel, and the mechanical plant sized for a final buildout that may be years away. This is rational — a shell is quick, relatively cheap, and keeps a project's place in the interconnect queue — but it also means that aerial footage of a massive data center frame tells you almost nothing about how much compute is live. Skeptics make much of this, and they are right to: shells are a real form of progress and also a real form of theater.

Silicon arrives on committed schedules. The GPU supply commitments behind Stargate — NVIDIA's involvement, Oracle's multibillion-dollar accelerator orders, OpenAI's separate compute deals — are real contracts with delivery dates, and GPU supply has eased enough that processors are no longer the scarcest input. When power is energized at a site, the racks fill in weeks. The scarce resource being sequenced is electricity, and the project that wins is the one that energized first, not the one that announced loudest.

For Abilene specifically, the pattern is visible in public reporting: first buildings operational in 2025, additional phases planned against additional power delivery, with the full 1.2 gigawatt capacity dependent on successive energization milestones rather than a single switch-throwing ceremony.

07How to Read the Next Data Center Announcement

The Stargate saga suggests a simple checklist for the next breathless headline. Sort any announced capacity figure into three buckets: announced, under construction, and energized. Announced means someone stood at a podium; under construction means concrete and steel are being bought; energized means a utility is delivering power to racks that are drawing it. These are wildly different facts that share the same units, which is precisely what makes the confusion profitable for the announcers.

Ask what the power contract says, not what the render shows. Does the site have a signed interconnection agreement? A transformer on order with a delivery date? A utility willing to put the megawatts in writing? These questions sound mundane, but they are the difference between the 500 billion dollar headline and the energized-gigawatt count — and they are exactly the questions the hoax framing exists to force, even where it overshoots the evidence.

None of this means Stargate will fail. It means Stargate will be slower than its own announcements, because every participant in it — even the most lavishly funded — is queueing behind the same transformers as everyone else. Five gigawatts is not impossible; it is simply a number that will be delivered in installments, against the grid's calendar, with or without a podium.