Shortly before publishing this newsletter, I spoke with Gil Luria, Managing Director and Analyst at D.A. Davidson, and asked him whether the capital was there to build the 17 Gigawatts of capacity that OpenAI has promised.
He said the following:
No of course there isn't enough capital for all of this. Having said that, there is enough capital to do this for a at least a little while longer.
There is quite literally not enough money to build what OpenAI has promised.
A few days ago, NVIDIA and OpenAI announced a partnership that would involve NVIDIA “investing $100 billion” into OpenAI, and the reason I put that in quotation marks is the deal is really fucking weird.
Based on the text of its own announcement, NVIDIA “intends to invest up to $100 billion in OpenAI progressively as each gigawatt is deployed,” except CNBC reported a day later that “[the] initial $10 billion tranche is locked in at a $500 billion valuation and expected to close within a month or so once the transaction has been finalized,” which also adds the important detail that this deal isn’t even god damn finalized.
In any case, OpenAI has now committed to building 10 Gigawatts of data center capacity at a non-specific location with a non-specific partner, so that it can unlock $10 billion of funding per gigawatt installed. I also want to be clear that it has not explained where these data centers are, or who will build them, or, crucially, who will actually fund them.
OpenAI, SoftBank and Oracle Are Misleading Us About The Progress Made In Building Stargate Data Centers, And The Media Is Falling For It
The very next day, OpenAI announced five more data centers planned “as part of the Stargate initiative,” “bringing Stargate’s current planned capacity to nearly 7 gigawatts,” which is when things get a little confusing.
Altman said back in July that Oracle and OpenAI were “committed to delivering 10GW of new compute capacity for Stargate,” and they were adding an additional 4.5GW of capacity in the US on top of the 1.2GW that was already planned in Abilene Texas. In fact, the Shackelford data center that is allegedly “new” is the 1.4GW facility I talked about a few weeks ago, though I see no mention of the Wisconsin one tied to the $38 billion loan raised by Vantage Data Centers.
This announcement involves a site in Doña Ana County, New Mexico and an “undisclosed location in the Midwest,” which I assume is Wisconsin, and eager members of the media could, I dunno, look it up, look up any of this stuff, use the internet to look up the news, all the news is so easily found.
But here’s my favourite part of the story:
The two remaining sites are being helmed by OpenAI and SB Energy, a SoftBank subsidiary that develops solar and battery projects. These are located in Lordstown, Ohio and Milam County, Texas.
To be clear, the Lordstown, Ohio site is not a data center, at least according to SoftBank, who said it will be a “data center equipment manufacturing facility.” The Milam County data center is likely this one, and it doesn’t appear that SB Energy has even broken ground.
It Takes 2.5 Years and $32.5 Billion Per Gigawatt Of Data Center Compute
Anyway, I want to get really specific about this, because the rest of the media is reporting these stories as if these data centers will pop up overnight, and the money will magically appear, and that there will, indeed, be enough of it to go around.
Based on current reports, it’s taking Oracle and Crusoe around 2.5 years per gigawatt of data center capacity. Crusoe’s 1.2GW of compute for OpenAI is a $15 billion joint venture, which means a gigawatt of compute runs about $12.5 billion. Abilene’s 8 buildings are meant to hold 50,000 NVIDIA GB200 GPUs and their associated networking infrastructure, so let’s say a gigawatt is around 333,333 Blackwell GPUs at $60,000 a piece, so about $20 billion a gigawatt.
So, each gigawatt is about $32.5 billion. For OpenAI to actually receive its $100 billion in funding from NVIDIA will require them to spend roughly $325 billion — consisting of $125 billion in data center infrastructure costs and $200 billion in GPUs.
If you’re reporting this story without at least attempting to report these numbers, you are failing to give the general public the full extent of what these companies are promising.
According to the New York Times, OpenAI has “agreements in place to build more than $400 billion in data center infrastructure” but also has now promised to spend $400 billion with Oracle over the next five years.
What the fuck is going on? Are we just reporting any old shit that somebody says? Oracle hasn’t even got the money to pay for those data centers! Oracle is currently raising $15 billion in bonds to get a start on…something, even though $15 billion is a drop in the bucket for the sheer scale and cost of these data centers. Thankfully, Vantage Data Centers is raising $25 billion to handle the Shackelford (ready, at best, in mid-to-late 2027) and Port Washington Wisconsin (we have no idea, it doesn’t even appear Vantage has broken ground) data center plans, allowing Oracle to share the burden of data centers that will likely not be built until fucking 2027 at the earliest.
Anyway, putting all of that aside, OpenAI has now made multiple egregious, ridiculous, fantastical and impossible promises to many different parties, in amounts ranging from $50 million to $400 billion, all of which are due within the next five years. It will require hundreds of billions of dollars — either through direct funding, loans, or having partners like Oracle or NVIDIA take the burden, though at this point I believe both companies are genuinely failing their investors by not protecting them from Clammy Sam Altman, a career liar who somehow believes he can mobilize nearly a trillion dollars and have the media print anything he says, mostly because they will print anything he says, even when he says he wants to build 1 Gigawatt of AI infrastructure a week.
Today, I’m going to go into detail about every single promise made by Sam Altman and his cadre of charlatans, and give you as close to a hard dollar amount as I can as what it would cost to meet these promises.
To be clear, I am aware that in some of these cases another party will take on the burden of capital — but these dollars must be raised, and OpenAI must make sure they are raised.
I’ll also get into the raw costs of running OpenAI, and how dire things look when you add everything up. In fact, based on my calculations, OpenAI needs at least $500 billion just to fund its own operations, and at least $432 billion or more through partners or associated entities raising debt just to make it through the next few years.
And that's if OpenAI hits the insane revenue targets it's set!