Soundtrack — Brass Against — Karma Police 

It was January 21, 2025. Per The Information, Larry Ellison, CEO of Oracle, had just flown to Washington DC from Florida, and had to borrow a coat “...so he wouldn’t freeze during an interview he did on the White House lawn, according to two people who were involved in the event.” He was there to announce a very big — some might even say huge — new project standing next to SoftBank CEO Masayoshi Son and OpenAI CEO Sam Altman.

“Together, these world-leading technology giants are announcing the formation of Stargate, so put that name down in your books, because I think you’re gonna hear a lot about it in the future. A new American company that will invest $500 billion at least in AI infrastructure in the United States and very, very quickly, moving very rapidly, creating over 100,000 American jobs almost immediately,” said President Donald Trump.

After he was done, Ellison stepped to the podium. “The data centers are actually under construction, the first of them are under construction in Texas. Each building’s a half a million square feet, there are ten buildings currently being built, but that will expand to 20.”

Following Ellison, SoftBank’s Masayoshi Son added that Stargate would “...immediately start deploying $100 billion dollars, with the goal of making $500 billion dollars within [the] next four years, within your town!” turning to Donald Trump with his hands extended. It was unclear what town he was referring to.

Altman added that it would be “an exciting project” and that “...we’ll be able to do all the wonderful things that these guys talked about, but the fact that we get to do this in the United States is I think wonderful,” though it’s unclear what “the wonderful things” or “this” refers to.

It’s been 15 months, and Stargate LLC has never been formed. SoftBank and OpenAI have contributed no capital to the project, other than SoftBank’s own acquisition of a former electric vehicle manufacturing plant in Lordstown, Ohio that it intends to turn into a data center parts manufacturing plant with Foxconn, which is best known for effectively abandoning a $10 billion factory in Wisconsin back in 2021. Oh, and Project Freebird, a SoftBank-built project that exists to funnel money to its subsidiary SB Energy, though I can’t imagine how SoftBank actually funds it.

No government money was ever involved, no funding ever left anyone’s bank account, no "initiative" ever existed, and OpenAI, Oracle and SoftBank have, in my opinion, conspired to mislead the general public about the existence and validity of a project for marketing purposes. 

The “data centers actually under construction” referred to a 1.2GW project in Abilene Texas that had been under construction since the middle of 2024, and had originally been earmarked by Elon Musk and xAI, except Musk pulled out because he felt that Oracle was moving too slow. While Ellison said that there were ten buildings under construction with plans to expand to twenty, only eight were actually being built (each holding around 50,000 GB200 GPUs across NVL72 racks), with the extension up in the air until March 2026, when Microsoft agreed to lease 700MW — so another seven buildings — that were meant to go to OpenAI. These buildings will not make Oracle any money, as Oracle is, despite spending so much money, leasing whatever land it uses from Crusoe.

Sidenote: Previously-unknown information from the Wall Street Journal published this week shows that the reason why Microsoft ended up buying the additional capacity at Abilene was because lenders were uncomfortable with providing additional funding to provide compute that was ultimately destined to go to Oracle. 

As far as those eight buildings go, only two are actually online and generating revenue, though sources with direct knowledge of Oracle’s infrastructure have informed me that work is still being done on both buildings despite CNBC reporting that they were “operational” in September 2025. 

Let’s break this down. Based on a presentation by landowner Lancium from May 2025, the Stargate Abilene campus was meant to have 1.2GW of AI data centers online by year-end 2025.

Based on reporting from DatacenterDynamics, the first 200MW of power was meant to be energized “in 2025.” As time dragged on, occupancy was meant to begin in the first half of 2025, had “potential to reach 1GW by 2025,” complete all 1.2GW of capacity by mid-2026, be energized by mid-2026, have 64,000 GPUs by the end of 2026, as of September 30, 2025 had “two buildings live,” and as of December 12, 2025, Oracle co-CEO Clay Magouyurk said that Abilene was “on track” with “more than 96,000 NVIDIA Grace Blackwell GB200 delivered,” otherwise known as two buildings’ worth of GPUs. 

Four months later on April 22, 2026, Oracle tweeted that “...in Abilene, 200MW is already operational, and delivery of the eight-building campus remains on schedule.” It is unclear if that’s 200MW of critical IT capacity or the total available power at the Abilene campus, and in any case, this is only enough power for two buildings, which means that Oracle is most decidedly not “on schedule.” 

Sources familiar with Oracle infrastructure have confirmed that while construction has finished on building three, barely any actual tech has been installed. It also appears that while construction has begun on a power plant of some sort, it’s unclear whether it’s the 360.5MW gas power plant or 1GW substation. In any case, Abilene needs both to turn on the GPUs, if they ever get installed.

Abilene is, for the most part, the only part of the Stargate project that’s anywhere near complete.

I say that because the other data centers — Shackelford, Texas, Port Washington, Wisconsin, Doña Ana County, New Mexico, Saline, Michigan, and Milam County, Texas — are patches of land with a few steel beams, if that. To be explicit, every single Stargate data center is funded by Oracle and its respective financial backers.

Oracle is taking on a massive amount of debt to build these data centers, working with a labyrinthine network of financiers and construction partners to pull together the capacity necessary to get paid for its five-year-long $300 billion compute deal with OpenAI. 

Oracle has also, per Bloomberg, deliberately raised money using “project financing” loans that are repaid using the projected cashflow, allowing it to keep the massive amount of debt off of its balance sheet. This is remarkable — and offensive! — because it’s borrowing over $38 billion to fund construction of its Wisconsin and Shackelford data centers (the largest debt deal of its kind on record) and said debt will now effectively not exist despite its massive drag on Oracle’s cashflow, which sat at negative $24.7 billion in its last quarterly earnings.

Based on estimates ($30 million in critical IT and $14 million in construction per megawatt) from TD Cowen’s Jerome Darling, the total cost of Oracle’s 7.1GW of data center capacity will be somewhere in the region of $340 billion to build.

All of these data centers are being built for a single tenant — OpenAI — which expects, per The Information, to lose over $167 billion (assuming it hits annual revenues of over $100 billion) by the end of 2028, and as a result does not actually have the money to pay Oracle for its compute on an ongoing basis.

In addition to its commitments to Oracle, OpenAI has also made commitments to spend $138 billion on Amazon over eight years, $250 billion on Microsoft Azure over an unspecific period, $20 billion with Cerebras over three years, $22.4 billion with CoreWeave over five years, and a non-specific amount with Google Cloud. 

All of this is happening as Oracle’s core businesses plateau, even after Oracle reshuffled them in Q3 FY25 to represent Cloud, Software, Hardware and Services segments, the latter three of which have barely moved in the last 9 months as low-to-negative-margin cloud compute revenue grows. 

In other words, Oracle’s only growth comes from a segment requiring hundreds of billions of dollars of compute. 

To make matters worse, every single one of these data centers is behind schedule. Stargate Abilene was meant to be done at the beginning, middle, and now the end of this year, yet sources tell me there’s no way it’s finished before April 2027.

Bloomberg also reported late last year that Oracle had delayed several data centers from 2027 to 2028, but here in reality, every other Stargate data center is somewhere between a patch of dirt, a single steel beam, multiple steel beams, or less than half of a shell of a single building. Considering it’s taken two years for Stargate Abilene to build two buildings, I don’t see how it’s possible that these are built before the beginning of 2029.

And at that point, where exactly will we be in the AI bubble? What GPUs will be available? What other kinds of silicon will exist? What will the demand be for AI compute?

I don’t think that OpenAI exists for that long, and even if it does, it will have to raise at least $200 billion in the space of three years to possibly keep up with its commitments.

I’m surprised that nobody (outside of JustDario, at least) has raised the seriousness of this situation.

Stargate, as it stands, will kill Oracle, outside of OpenAI becoming the literal most-profitable and highest-revenue-generating company of all time within the next two years. Even then, by the time that Abilene is built, its 450,000 GB200 GPUs will be two-years-old, and entirely obsolete far before its debts are repaid. A similar fate awaits whatever GPUs are put in the other Stargate data centers.

Today’s newsletter is a thorough review and analysis of the ruinous excess of Stargate, a name that only really means “data centers being built for OpenAI in the hopes that OpenAI will pay for them.” Oracle is mortgaging its entire future on their construction, and even if it gets paid, I see no way that the cashflow from OpenAI’s compute spend can recover the cost before its GPU capex is rendered obsolete, let alone whether it can cover the debt associated with the buildout.

I’m Larry Ellison — Welcome To Jackass.

Coming Up In This Week’s Where’s Your Ed At Premium…

• The total estimated cost of Oracle’s Stargate capacity is around $340 billion.
• OpenAI needs to make, in total, $852 billion in both revenue and funding through the end of 2030 to keep up with its compute costs with Oracle, Amazon, Google, CoreWeave and Microsoft.
• Oracle cannot afford to pay for the cost of construction and equipment out of cashflow, and has had to take on over $100 billion in debt and sell $20 billion in shares.
• Across a potential 7.1GW of planned Stargate capacity, Oracle stands to make around $75 billion in annual revenue.
  • Abilene is expected to generate around $10 billion a year in revenue on completion for a project that will likely cost in excess of $58 billion.
• Stargate Abilene is extremely behind schedule, and likely won’t be finished until Q2 2027.
• Oracle estimated in 2024 that Abilene would cost it $2.14 billion a year in colocation and electricity fees.
• Oracle has spent over $5 billion in construction costs on the first two buildings of Abilene, with sources saying that it will likely spend over $10 billion to finish them, suggesting an overall cost of around $48-per-megawatt.
• Oracle’s remaining Stargate sites are barely under construction, and will likely not be finished before the end of 2028.
• Even if Oracle builds the data centers and OpenAI pays for them, the incredible upfront cost and NVIDIA’s yearly upgrade cycle will render much of the GPU capacity worthless within the next ten years. 
• And if OpenAI fails to pay, Larry Ellison likely has over $20 billion in personal loans collateralized by over $60 billion in Oracle shares, meaning that margin calls will follow with the collapse of Oracle's stock.

Welcome to the end of Oracle, or Sell The Compute To Who, Larry? Fucking Aquaman?