The Stargate Data Center Layer Cake
Last July, I wrote a quick post about a convoluted data center deal in AI. Really, just to help myself wrap my head around it. Well, it's time to do that again. As Tabby Kinder and George Hammond report:
Oracle will spend about $40bn on Nvidia’s high-performance computer chips to power OpenAI’s new giant US data centre, as technology groups race to build the vast infrastructure needed to underpin artificial intelligence models.
The site in Abilene, Texas, has been billed as the first US Stargate project, the $500bn data centre scheme spearheaded by OpenAI and SoftBank, and will provide 1.2 gigawatts of power when it is completed next year, making it one of the largest in the world.
Okay, so far so good. Oracle spending $40B is a lot – perhaps especially for them – but it's the Abilene, Texas site: aka 'Stargate 1' (as far as I know, it's not actually called that, but I'm calling it that – 'SG-1' perhaps?).
But hasn't it been under construction for a while, with a ton of NVIDIA chips that Oracle already bought? Yes, and actually, this was the point of the post last year. At the time, xAI was walking away from a deal with Oracle to use the Abilene site as their 'Colossus' data center. Yes, that's right, the first Stargate data center was nearly an Elon Musk property, until he got impatient with the Abilene build-out speed and set out to build his own data center. Which he did.
Anyway, the Abilene site is now set to be completed in "mid-2026" and this (apparently) new $40B commit is going specifically towards:
Oracle would purchase about 400,000 of Nvidia’s GB200 chips — its latest “superchip” for training and running AI systems — and lease the computing power to OpenAI, according to several people familiar with the matter.
Previously, Oracle, like everyone else, was buying up H100 and H200 chips from NVIDIA to power the data center. But the state of the art shifted. As, I might note, it will again! Anyway, all of this is fairly straightforward so far, but then it starts to get a bit messy:
The site’s owners, Crusoe and US investment firm Blue Owl Capital, have raised $15bn in debt and equity to finance the Abilene project, which will encompass eight buildings. Ground was broken on the site in June last year.
The data centre is expected to be fully operational by mid-2026. Oracle has agreed to lease the site for 15 years. Stargate, which incorporated earlier this year, has not invested in the site.
That's right, Oracle doesn't actually own the Abilene site, they simply have a long-term lease on it. And they're in turn sub-leasing it to OpenAI. But only because xAI walked away. And only then because Microsoft opted not to take on the lease,1 and instead handed such financial burdens over to SoftBank – and thus, the 'Stargate Project', which was also initially meant to be a Microsoft/OpenAI effort, by the way, was born as a SoftBank/OpenAI/Oracle (and friends) effort. With a phone call.
JPMorgan had provided the bulk of the debt financing across two loans totalling $9.6bn, according to people close to the matter, including a $7.1bn loan announced this week. Crusoe and Blue Owl have separately invested about $5bn in cash.
But wait! There's yet another owner of the site – in the form of a bank providing the debt for the built out. Per a report by Anissa Gardizy of The Information a couple days ago:
The bank already lent $2.3 billion to the companies developing the Abilene data center, which funded the initial phase of the project. The new $7 billion loan will round out the funding for the construction of the data center, which would be one of the largest in the world, with eight buildings that cost $1.4 billion each. The data centers can hold a total of 400,000 Nvidia chips.
The project is being developed by a group of companies and will be used by OpenAI. The actual project is being built by data center developer Crusoe, and it is partly owned by asset manager Blue Owl and Primary Digital Infrastructure, an investment fund. Oracle has agreed to lease it for 15 years. Oracle will then rent the chips to OpenAI. At full capacity, the site is expected to have more than 1 gigawatt of power for AI chips.
To recap: Crusoe is developing the site which is partially owned by Blue Owl and Primary Digital Infrastructure, thanks to loans from JPMorgan. That group, in turn, has Oracle signed up for a 15-year lease, and they're putting up the money for the NVIDIA chips that will be housed in the data center. Those chips, in turn, are being sub-leased to OpenAI. Which, by the way, is likely paying for this sub-lease mainly with SoftBank's money.
There are a lot of layers here. And it's Stargate partners nearly all the way down.
Why doesn't, say, OpenAI just buy the chips from NVIDIA directly? Well, they'd still need the money to do so and a place to put them. They have a lot of money, but not $40B to spend. And they do not yet have data center space of their own. And certainly they don't have a half-complete 1.2GW space. And so a lot of extra people are getting paid here.
And let's not even get started on 'Stargate UAE' until they do. Also, I can't believe I wrote a whole post about AI data center sub-leases and NVIDIA chips and debt and didn't mention Coreweave. Another big OpenAI partner!
1 Yes, this means that the chain was nearly Crusoe and Blue Owl owning the site and leasing it to Oracle who in turn was leasing their servers to Microsoft who in turn was giving them to OpenAI. I guess it's good they cut out one middle-man here?