Microsoft Outsources OpenAI's Ambitions to SoftBank
The headline is perhaps not entirely fair. But it's also seemingly directionally accurate. At least when it comes to the go-forward compute needed for OpenAI's ongoing quest to achieve AGI and onwards from there towards "superintelligence".
SoftBank is the new sugar daddy.
The announcement around 'The Stargate Project' already made this pretty clear at the highest level. From the fact that it was Masa Son (and Larry Ellison) standing alongside Sam Altman to announce the project at the White House. Microsoft, after all, was to be a "technology partner" in the endeavor, but not a capital partner. It was notable enough that it required a separate press release from Microsoft to announce that they were "evolving" the partnership with OpenAI, to allow them to see other cloud providers – notably, Oracle, which was building the first Stargate project (before it was even the Stargate project).
Microsoft wanted to make it clear: they were not fully breaking up with OpenAI, they were just evolving into an open relationship. And it wasn't hard to read between the lines as to why: all the drama aside, Microsoft no longer wanted to be the sole provider of resources to OpenAI going forward. But riding just underneath that notion was the bigger one: they didn't even want to be the main provider any more. At least not for where OpenAI aimed to go from here.
That was driven home in the updated OpenAI internal projections that Cory Weinberg, Jon Victor, and Anissa Gardizy were able to scoop for The Information. The numbers are fascinating in relation to the old projections – mere months old – but also by themselves. They point to the reality above: what was a tight coupling of OpenAI and Microsoft will now be a tight coupling of OpenAI and SoftBank.
It's obviously more nuanced than that. Because, again, Microsoft remains a key partner in many regards and will even if/when Stargate is built out. And because of Microsoft's extremely prescient early investment in OpenAI, they will remain – or rather, become – a large shareholder if/when OpenAI is able to convert into a for-profit entity. In fact, they'll likely be the largest shareholder, depending on how much equity the for-profit has to give to the non-profit to make it "whole" in the eyes of the law.1 But make no mistake, Microsoft is nudging OpenAI out of the nest. Perhaps shoving.
And it's hard to blame them! Satya Nadella has a public company to run. Sure, Microsoft's stock has been bolstered by the OpenAI affiliation, but that wasn't going to continue in perpetuity. At some point, Wall Street is going to start hitting all the tech companies for their AI spend after rewarding them for the very same thing. It's inevitable. We saw hints of this with the DeepSeek situation. But the real reckoning is still to come.
And all of Big Tech knows this. So it's s delicate dance of spending what it takes to stay in the race but not getting too far ahead of the song. That's why they're all currently coalescing around the same CapEx numbers – well, aside from Apple, but that's another story. The problem for Microsoft is that if they gave OpenAI everything they wanted in terms of spend, their CapEx number would probably more than double on a yearly basis. And may grow from there!
How do I know this? Well, just look at the "promised" spend on Stargate. $500B over 4 years. To be clear, this is clearly a made up big round number to please a certain President who loves made up big round numbers. But still, it's probably safe to assume this is directionally where OpenAI is aiming. And much to Elon Musk's chagrin, they do appear to have line-of-sight to at least that first $100B. Thanks largely to... SoftBank.2
And that's why you had Nadella on CNBC humorously noting that he was "good for my $80B" this year. At the time, it was interpreted as a shot at OpenAI and the ridiculous-sounding Stargate spend. But it was much more of a signal to Wall Street that they were sticking with their CapEx projection and figured out a way to offload the additional spend that OpenAI required.
Said another way: Microsoft was keeping the CapEx at $80B, not going to $180B – a year – to accommodate Stargate. A project which they were initially working on with OpenAI, by the way, until a little coup got in the way...
You can also hear all of this in Nadella's recent sit down with Dwarkesh Patel. While ostensibly about their new quantum project, the podcast kicks off talking about AI. And Nadella is pretty clear in his thinking that not only is all the talk of AGI overhyped, but more so that increasingly, AI spend needs to be tethered to reality. Meaning, it needs to be tethered to actual business growth. Is it any wonder that Wall Street loves this guy? And he delivers!
What he's not delivering, at least not anytime soon with OpenAI is superintelligence. I still suspect we'll hear OpenAI talk about "AGI" at some point this year, but the term is so nebulous that it's unclear how much it really matters beyond the PR value. But Sam Altman is good at such things; always has been. Nadella, again, is more skeptical. But seemingly also happy to be along for the ride now that he's not the one footing the bill beyond what they've achieved to date.3
Anyway, back to those internal OpenAI numbers that The Information gathered:
The company has told investors in recent weeks that it expects Stargate, the fledgling data center expansion project to be heavily financed by SoftBank, to support three-quarters of the computing power it will need to run and develop its AI models by 2030. That would represent a reversal from today and the next few years, when its power and data center capacity is mostly handled by Microsoft.
SoftBank powering 3/4ths of OpenAI's computing needs by 2030 would be pretty wild. Because they currently power 0/4ths of it and Microsoft powers something like 4/4ths. Again, it speaks to insane growth in terms of compute needs from here (and yes, Microsoft taking their foot off the pedal).
They go on to note that SoftBank is likely to contribute – wait for it – 3/4ths of the $40B round they're currently raising. And nearly half of that will go towards, what else: Stargate.
Perhaps more incredibly, OpenAI projects that a full 1/3rd of their revenue this coming year will come from – who else – SoftBank. They're investing and buying what they're investing in and everything in between. We're perhaps treading into some gray areas here if SoftBank is the one setting the valuation of the company and if they're valuing it based off of the revenue they're providing but I digress...
To OpenAI's credit, but largely thanks to this SoftBank spend, they now believe they're going to beat their original revenue targets for this year. For a company at this scale, it's impressive no matter the reason:
Overall, OpenAI told investors it expects revenue to more than triple this year, from $3.7 billion to more than $12.5 billion. This year’s expected revenue is slightly higher than what it told investors last fall, thanks to the SoftBank spending. By 2026, OpenAI expects revenue to hit $28 billion. It expects most of its revenue in 2025 and 2026 to come from ChatGPT, with the rest of its sales coming from software developer tools and AI agents.
Also impressive: how much of the revenue is coming from ChatGPT, which could end up more sustainable than, say, an API-heavy business model. All that said, all of this comes at a cost. Quite literally.
OpenAI anticipates that the amount of cash it is burning will grow at a similarly torrid rate. It expects cash burn to grow from about $2 billion last year to nearly $7 billion this year. The company forecasted that its cash burn would grow in each of the next three years, peaking at about $20 billion in 2027, before OpenAI would turn profitable by the end of the decade after the buildout of Stargate.
These are losses that only Masa Son could love. But unlike WeWork and some of his more questionable bets in the past, if they are able to convert this spend into something truly massive (from a business perspective) and profound (from a dent-in-the-universe perspective), it's not entirely impossible that it will be worth it. Again, quite literally. It would make the Alibaba bet look like winning on the penny slots.
There's a long way between here and there though. Obviously! And a lot of capital needed to get there. To say the least! But it appears that OpenAI has successfully navigated from one deep pocket in Microsoft to a nearly unlimited pocket in SoftBank – unless the macro picture shifts substantially, which will happen as it always does. And so the race is on!
OpenAI expects their costs to be $320B between now and 2030 – with "more than half" of that still going towards model training. If I had to guess, based on Nadella's various comments, this is the specific part of the ship from which Microsoft clearly wanted to jump. OpenAI is now not expecting the cost of actually running their models to surpass the training cost until 2030. Again, this suggests that the plan alongside SoftBank is now very much to go for it, pedal the the metal. From AGI to Superintelligence. As hard and as fast as they can. Satya Nadella, though, has a business to run.
Previously, OpenAI thought they might spend a "mere" $200B by the end of the decade. But now by that point, most of that compute will be running on Stargate, which should be made up of several data centers – ideally around the world, much to President Trump's chagrin – assuming they can get them built. All it will take, beyond the money, is a mere 8 gigawatts of power. My god this is going to be something to watch.
And Microsoft will be more than happy to watch, largely from the sidelines, it seems.
Update February 24, 2025: A bit more on this general topic based off of some news today about Microsoft's data center spend...
M.G. Siegler
1 You know, the thing Elon Musk has been trying to muck up.
2 And specifically thanks to their ARM stake, which gives them the leverage for the debt that will be used here.
3 Which, to be clear, is amazing. Last night, I was having a conversation with ChatGPT while I was trying to remember a movie I had seen where someone seeks power only to be totally consumed by it, but enjoying their descent. As it turns out, this is a very common theme in movies. And ChatGPT knew them all, in an instant. It was a fun conversation!