(Preview) The End of OpenAI and Microsoft, Risks and Rationale of the Stargate Project, DeepSeek-R1 and Bitter Lessons for the Future

On their 200th episode of the show(!), Ben and Andrew discuss the Stargate Project and what clarifies about the dynamics between OpenAI and Microsoft. Then: the risks inherent to Stargate investments, the rationale for the corporate structure announced this week, PhD-level agents in 2-3 years, and various lessons from the success of DeepSeek and its latest models.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Andrew Sharp?

  Hello and welcome to a free preview of Sharp Tech.

• 0:08Andrew Sharp?

  Hello and welcome back to another episode of Sharp Tech. I'm Andrew Sharp and on the other line, Ben Thompson. Ben, how you doing?

• 0:19Ben Thompson?

  Doing okay, Andrew. How are you? How is Washington DC?

• 0:22Andrew Sharp?

  Washington DC beginning to calm down. There are fewer military checkpoints in my midst as everybody makes their way back to wherever they came from a week ago. So I have no complaints. It's still unbelievably cold here, but you probably call me a pansy because you know real cold in Midwest Wisconsin. But other than that, you know, life is good.

• 0:46Ben Thompson?

  Good. I'm glad to hear. I'm glad you're, you did not freeze to death. The trials and tribulations you bear, we all respect it.

• 0:55Andrew Sharp?

  That's right. Well, and you know what? Our colleague Duman, before we came on to record, informed me that this is our 200th episode of Sharp Tech, which is kind of amazing. 200 big ones. Do you have any deep thoughts to mark the occasion or should we just keep it rolling?

• 1:11Ben Thompson?

  No thoughts at all. Sorry. I've got nothing for you.

• 1:15Andrew Sharp?

  Okay. My only thought is that the next 200 episodes will probably be much crazier than the last 200 episodes, as exemplified by this week's news. So Ben, on that note, here we go. I pulled a number of thought provoking tweets for today's rundown. We're going to start with a pair here.

• 1:34Ben Thompson?

  What's a tweet?

• 1:35Andrew Sharp?

  One, a post, a post on X is what I pulled for our rundown. Thank you. I'm going to improve the nomenclature for the next 200 episodes together.

• 1:46Ben Thompson?

  We could call it tweets. We'll call it tweets forever.

• 1:48Andrew Sharp?

  That's right. We're boomers here on this podcast. OpenAI on X Tuesday afternoon posted, "The Stargate Project is a new company which intends to invest $500 billion over the next four years, building new AI infrastructure for OpenAI in the United States. We will begin deploying..." A hundred billion dollars immediately. The in this infrastructure will secure American leadership in AI, create hundreds of thousands of American jobs and generate massive economic benefit for the entire world. This project will not only support the re-industrialization of the United States, but also provide a strategic capability to protect the national security of America and its allies.

• 2:33Andrew Sharp?

  The initial equity funders in Stargate are SoftBank, OpenAI, Oracle and MGX. SoftBank and OpenAI are the lead partners for Stargate with SoftBank having financial responsibility and OpenAI having operational responsibility. And then a follow-up tweet from a user named Tane Jaidpuriya. He says, wow, Stargate project will invest $500 billion over the next four years. That's about 0.4% of US GDP over that period. For comparison, the inflation-adjusted dollars spent on other large undertakings. The interstate highway system caught a cost about $650 billion.

• 3:15Andrew Sharp?

  The Apollo program cost about $280 billion. The Manhattan project cost about $35 billion. So Ben, Stargate is a fun name for a massive AI project, so I support it on that basis. But from a practical standpoint, I'm not clear on what's trying to be achieved with Stargate and why it's altogether different than what these companies have already been pushing toward for the last few years.

• 3:46Ben Thompson?

  I mean, well, that is a good point, which is the this partnership between Oracle and OpenAI is not new. It was sort of announced last summer and it was notable at the time because OpenAI partnering with an infrastructure provider that was not Microsoft. And I wrote about it, not exactly at the time, but but sometime in the fall. It ended up being that Microsoft was going to like run or they had some sort of role. It would still be like an Azure Cloud or something along those lines. But there was a clarifying OpenAI. At the time, which is, oh, all our training loads continue to run on Microsoft infrastructure. And, you know, I got to pull up my inner pedant, pedant, pedant. I don't know how to say that word.

• 4:30Andrew Sharp?

  Pedant.

• 4:32Ben Thompson?

  I wasn't even close. That's unfortunate.

• 4:35Andrew Sharp?

  But you are a pedant. It's not all that inner, actually.

• 4:38Ben Thompson?

  Yeah, well, that's just a sop to all our pronunciation pedants in the audience.

• 4:43Andrew Sharp?

  That's right.

• 4:45Ben Thompson?

  It was in the present tense. It was not in the future tense. And it's hard to sort of talk about this and think about this independent of the Microsoft angle here. Which you go back to what OpenAI used to be, which was this sort of research team with this goal of achieving AGI. They realized they need a lot of money to do so. Microsoft's like, well, we could really use these sort of models along the way. That's cute. You have your API. Startups can use it, but big companies are going to want to sort of work with us. So we'll, we'll also have the API. We'll give you money and you basically are our research team.

• 5:24Ben Thompson?

  And we will, we will reap the benefits. And I was sort of thinking about the course of this relationship. And we've talked, obviously about the OpenAI board drama and everything that went down then. And it's easy to point to that as when this relationship sort of started to fall apart. But I think it's, the more I think about it, it's one of those stories that is very appealing to humans. Cause there's a clear sort of narrative arc. Oh, they had this beautiful partnership and OpenAI sort of couldn't get their crap together. Because they had this weird sort of arrangement and suddenly it starts to come apart.

• 6:03Ben Thompson?

  But a lot of big shifts, I think, end up being more structural than necessarily sort of, you know, cute stories, as it were. And the real split was a split that no one realized was a split at the time, which was ChatGPT. The ChatGPT, and I, I sort of called this pretty early on, really fundamentally changed the trajectory of OpenAI. Like, like I called him the accidental consumer tech company. And I, and I think a lot of this stuff that's happened at OpenAI, not including the board thing, but also the departure of lots of folks is. It's a transformation of this company to suddenly having an opportunity to pursue

• 6:48Ben Thompson?

  the biggest opportunity that exists, which is being one of the big consumer tech companies. Like, there's very few of them, right? There's Apple, it's really Apple, Google and Facebook are it. Microsoft's an enterprise company. Uh, Amazon is a consumer tech company, but it's, you know, it's an e-commerce site. It's sort of, it's not a platform in the way we think of the other ones. Or you just think about OpenAI, you go to ChatGPT and you use it and they develop features and you're on there. And those sorts of companies, sort of by definition, and you think about like my aggregator sort of framework where they marshal demand, they are at the top of their value chains. Everyone that interfaces with them exists to serve them. Apple's developers exist to serve them.

• 7:33Ben Thompson?

  Apple's supply chain exists to serve them. Apple brought in their own sort of processor technology for themselves. Like Google, the websites of the world exist to serve Google. They have special maps on their site, just so Google can sort of map them more accurately. The whole SEO industry is a somewhat perverse version of serving Google, right? Like, all these influencers, all this content is there to serve Facebook. Because those, they, Amazon suppliers, if you want to go there, they exist to serve Amazon. And they get really upset about this and frustrated about this, but it is sort of an inevitable outcome when you sort of control demand.

• 8:13Ben Thompson?

  And in this context, Microsoft is not particularly interested in being a servant. That's not what they do. They, they were there for the technology. They were there for OpenAI to serve them, not the other way around. And, and, and in this sort of context, you look back at ChatGPT, which no one expected to be what it was. OpenAI didn't expect it, Microsoft didn't expect it. It was, again, it was this accidental tech company. And that, once that happened, once this sort of course was set, this sort of long-term dissolving of the relationship, which is It's not, it's not going away. Microsoft still has exclusive access to the OpenAI API, which is what they want.

• 8:56Ben Thompson?

  A lot of inference is still going to run on Azure. Microsoft going forward has the right of first refusal. If OpenAI wants cloud capacity, Microsoft could say, yeah, run that on ours or run it somewhere else. But, but it's not, it's not what it was. And in retrospect, that was inevitable.

• 9:14Andrew Sharp?

  Rendered impossible because ChatGPT changed the nature of what OpenAI was going to be and was trying to do. Is that right?

• 9:22Ben Thompson?

  That's right. That's right. It just, it was, it was always going to be sort of, sort of incompatible in that regard. So I think that that's the broader context to sort of think about sort of what's happening here. So, so you have.

• 9:34Andrew Sharp?

  Right. And I want to come back to that though, but just as far as Stargate specifically, like I'm trying to determine what it actually means and whether it's real progress and meaningful to the long-term future of AI in America, or if it's essentially the announcement this week. I don't know why this was announced from the White House because the government doesn't seem to be all that involved in it. Is this the next stage in the never-ending shell game that Sam Altman may or may not be running on the entire world? Like, how seriously should I take that announcement?

• 10:08Ben Thompson?

  Well, well, well, just so, so you have OpenAI that has this grand opportunity. And I was obviously married to grand ambition. The grand ambition has been there for a long time. Sam Altman is anything, is nothing if not grandly ambitious, I think is, is, is a way.

• 10:22Andrew Sharp?

  And good at marketing those ambitions. Yeah.

• 10:25Ben Thompson?

  Absolutely. So you have this, by and large, this sort of inevitable split. And you have still this dream of AGI and ASI sort of beyond that. And what we've, you know, what, what you're sort of seeing, and we're going to talk a sort of about DeepSeek and, and things sort of later on, is we're, we're seeing the manifestation of a lot of the ingredients that were sort of predicted and talked about in what will enable sort of AI to take off. So first and foremost, you have like. You had this model sort of challenge, where's the data going to come from? Once we've sort of harvested the whole internet? Or in all, you know, whatever we can sort of get our hands on? Well, AI can create data for itself, right?

• 11:09Ben Thompson?

  Like, and so you're starting to see this cycle sort of manifest. This seems to be a core component of, you know, you know, of some of these more reasoning models. And, and so the question was, would synthetic data sort of work to build this? Well, it does seem like it's going to work. And so suddenly you're needing huge amounts of capacity, not just to train new models, but to train the models that train the models and, and to have these sort of iterative passes and things along those lines. And you, you have this goal and this ambition married to this entity that it's, it's sort of core supporter, it's core infrastructure provider. It's not that they don't want AGI.

• 11:51Ben Thompson?

  It's that they're sort of governed by the need and desire to serve their existing customers. This goes back to our, our discussion last week. Are those existing customers going to be on the leading edge of all this AI sort of stuff? If you ask them, of course they'll say yes, but the reality of these being human organizations organized around the old way of doing things is that it's going to be sort of a long, a long slog. You have Microsoft talking about, yes, of course, we're going to keep investing, but our rate of investment in training is going to be gated by the inference demand. Like how much, like to what extent do our customers want inference computing is the extent to which we'll, we'll invest in training. And that's a very telling statement because it's sort of backwards looking in a way.

• 12:35Ben Thompson?

  It's, it's where, you know, the OpenAI's view is not that we're governed by inference, what people use today that will guide our investments going forward. It's that we are pushing forward because we see this end in sight and, and we want to accelerate and sort of lean into it. And so there's a bit where this overall, setting aside the numbers and what's involved, it just makes much more sense, this sort of structure and sort of multiple regards. So number one, Stargate is this independent entity. It's not OpenAI. It is a different company of which OpenAI is an equity holder. And it sounds like they're not putting much money in, they're just putting in sort of operational expertise.

• 13:16Ben Thompson?

  Uh, you know, uh, SoftBank's, you know, in charge of getting the money, which is a, a different question.

• 13:23Andrew Sharp?

  All right, and that is the end of the free preview. If you'd like to hear more from Ben and I, there are links to subscribe in the show notes, or you can also go to sharptech.fm. Either option will get you access to a personalized feed that has all the shows we do every week, plus lots more great content from Stratechery and the Stratechery Plus bundle. Check it out, and if you've got feedback, please email us at [email protected].