Oracle Cloud Infrastructure Revenue Doubles, Inside Blackstone’s Bid to Rule Over AI Financing

0:13Welcome everyone to The Information's TITV. My name is Akash Pasricha. We've got a great show lined up for you today. Before we get started, we have some news here at The Information. You probably know my colleague Rocket Drew pretty well by now. He is on here. every week sharing all of his in-depth reporting on the technical debates shaping the AI sector. And personally, I have always hoped that those conversations would go on a lot longer. I have so many more questions for him. And lucky for us, the information is launching a new weekly show that Rocket will be hosting. It's called AI Deep Dive.

0:51The first episode is going to air on Monday at 12 o 'clock Pacific, 3 p.m. Eastern with a special guest. He is bringing on Noam Brown, a top research scientist at OpenAI. I want to play a short clip for you from that conversation. Here is Rocket. There's been a lot of discussion recently about the future of chains of thought, in part because of an article The Information published about a new technique that Astra is using where more thinking can happen sort of in the model's head. It can sort of keep more of its thinking to itself and do less thinking out loud, at least if this technique were to be scaled up in the future.

1:31So I'm curious what you made of this whole discussion. I think that idea really touched a nerve with some people. I'm sure it was all over your Twitter feed. It was all over my Twitter feed. What did you make of that whole dialogue? For chain of thoughts, Astra does have observable chain of thought and it is something that we can monitor it's basically it's behavior um i i do think chain of thought monitoring first of all it's a real gift like we were very lucky that this ever existed um and and it is it is fragile um we've seen situations where you know companies have there's the experiments that show if you basically if you punish the model for thinking bad thoughts, that it will learn to think bad thoughts in a way that's not observable.

2:19And so there's always a temptation, you know, for example, during reinforcement learning, you can say like, well, the model is thinking about hacking and we should punish it for thinking about hacking. And actually, like as counterintuitive as it sounds, you actually don't want to do that. Like you want to punish it for taking actions that are observable. Like if it takes actions to hack, then yes, punish it for hacking. But you don't want to punish it for its thoughts because then you will simply lose the ability to monitor what it's thinking make sure to tune in here on the informations youtube channel or wherever you get your podcast for ai deep dive today on this show we're going to unpack oracle's quarterly results top line growth accelerated but the company is burning cash we'll then bring on the co-founder of base 10 to talk about their latest acquisition.

3:08We've also got exclusive reporting on Blackstone's Neocloud joint venture with Google. And to close out the week, we have the editor's cut. We'll get some opinions on the iPhone duo. And the big question of the week, will AI end humanity? It's going to be a great show, so let's get right on into it. Oracle reported its quarterly results. Revenue is expected to grow up to 34 % in the current quarter. That is up from 30 % in the most recent quarter and 21 % in the quarter prior. I want to bring on Rishi Jaluri, I'm Managing Director of Software Equity Research at RBC Capital Markets. Rishi, welcome back to the show.

3:45It's great to have you here. Great to be here. Thanks for having me. Okay, so lay it on us. The good, the bad, the ugly, the not so ugly. What did you think? Yeah, look, I think overall it was probably, I would say, a better than feared quarter. There were concerns in there going about - More than fear. My gosh, that is a low bar. Yeah, yeah. I mean, you've seen the action in Oracle. I mean, look, I think the OCI numbers were really strong. The RPO growth was good. But I think when I was talking to clients, the worry was, are they going to have to raise their capex? Is there going to be an issue of some of the delays in some of the sites?

4:20And they kind of reaffirmed their cash capex and total outlay. And they reaffirmed their ability to maintain the guide or actually exceed the guide without expecting these New Mexico and Wisconsin sites that have been delayed to go online. So, you know, I think that's why you're seeing you initially had the stock up in the aftermarket. I think that's why you're seeing the stock more flat today is just everyone's digesting all these different moving pieces. So on the call last night, they got the question about the data centers and the whole dialogue that there were delays. And the answer that they gave was we are not putting all of our eggs in one basket.

4:55Also, there are some sites that are ahead of schedule in their construction. What did you make of that answer? And how do you compare that to maybe what you're hearing from your own research? Yeah, look, I'm hearing, you know, some similar sort of things that they're trying to diversify. They're trying to, you know, that's why you also have some of the efforts they're doing, like the bringing your own hardware to try to accelerate some of these. Some of these sites are able to get on ahead of time. But I think that's also why you saw them delivering 850 megawatts. of capacity in the quarter, not the one gigawatts that I think the street was expecting.

5:29So there really is some impact of these delays. How it translates to revenue, it's a complete black box that I think none of us have any visibility into, but it's still something that we're paying very close attention to is delays on any of these data center sites. How do you compare Oracle's approach to their CapEx right now to the other hyperscalers? Yeah, look, I think if we were to really just analyze the capex of oracle it's probably the most aggressive roadmap right in terms of where they were to where they're going and and it makes sense you have this massive uh open ai deal um that kind of got announced with such fanfare a little bit over a year ago and they talked about peak capex in you know 2027 2028 before theoretically it kind of tapers down from there but now you're layering in power cost component price inflation etc with a little bit of offset by BYOH, but it is telling of this is a company that was in the fourth place, right, relative to the other hyperscalers, and they are trying to play catch up.

6:27And that's why I think you're seeing a more aggressive ramp in their capex. And everyone's aggressive in capex, but I think if you look at the shape of the ramp for Oracle, it's larger than the others. Right, right. Let's talk about the cash flow projections. I mean, so they're burning cash and they did not comment last night on when they expect to turn cash flow positive. They kind of dodged that question. Have you modeled that in yourself at all? Or do you have any expectations or predictions on when the company could return to free cash flow? Yeah, look, we don't have them being cash flow positive anytime soon, right?

7:04Because if CapEx is going to go up again next year and potentially again in 2028, it's hard for me to see how they raise cash. I know they've done some efforts in terms of doing a RIF. They've had multiple rifts now. And so maybe that brings up margins in the rest of the non-cloud, non-AI business. But then also growth is going to decelerate further. So it's hard for me to see a path for them to get to free cash or break even anytime soon. They did reiterate they don't need to raise any more capital in this calendar year. But calendar year 27, calendar year 28 are still big open questions. So then, I mean, help me understand here the argument to invest in Oracle.

7:44And I'm forgetting now if you have a buy or sell rating on the stock. So maybe you have a sell rating. I don't know. But I mean, what would be the arguments to buy Oracle stock when you could just invest in one of the hyperscalers? They're not burning cash. It's the same business. They've got their data centers more built, not entirely built. What is the argument here, do you think? Yeah, and I am sector perform rated and have been kind of all throughout the boom and bust of this one. My argument has actually been pretty similar to that, which is I would rather own Microsoft at these levels over Oracle, especially given the breadth of Microsoft's portfolio and the fact that they are still cashflow positive.

8:26Having said that, if I were to make an argument, why would someone be bullish on Oracle? I think number one, you're saying there's more torque to that AI infrastructure trade. It's definitely a lot more incremental, and we see that. I think number two, you're saying because this is more modern architecture, because they're kind of building up in a scale-up way rather than kind of your traditional scale-out, that that maybe leads to better overall unit economics. And then you've got the just impact of this massive opening ideal that becomes a lot more incremental to the business. Is there anything you think the company could be doing better?

8:59Look, I just think maybe we'll find out at the analyst day next month. But I think the one thing that a lot of investors are having questions around still is what does a long-term trajectory look like? Do we have to kind of revisit some of the very bullish assumptions that they gave us at the analyst day last year? You know, I think exactly your point, they were kind of dodging the questions on when will they get back to free cash or break even kind of what do things look like a few years out? It's a lot of stay tuned. And so that's one thing that I think the market would want to see a lot more is just a little bit more certainty around what does that even three-year roadmap look like here?

9:32And all the discussion last night about the ways that they are tempering their expenditures and their investment. I mean, they're asking customers, they're saying you could bring your own chips. They're also asking for prepayments on some of these deals. I mean, do you see those as smart, sustainable strategies in the long run, or is this just a short-term strategy to really manage their cashflow here? It feels like a little bit of a band-aid solution to be completely honest. No, no, don't get me wrong. Offering customer flexibility is the right thing. but I think this idea of bring your own hardware, there's a reason the other hyperscalers aren't doing it to such a great extent, right?

10:12I think Oracle is in a position that they're trying to be able to do whatever they can to limit their capital fundraising needs, especially given where the stock is today and where interest rates are and CDS and all that kind of stuff. So I would say it's more of a little bit of a band-aid solution. I think the real long-term solution is to actually show real margins on these AI revenue. And it's still very early. They've talked about it kind of being about 30 % to 40 % gross margin. But if they can actually demonstrate that as this OpenAI and other AI deals scale, that's what actually starts to make this look a lot more attractive.

10:47It's still just early to say whether that's going to happen or not. Right. Great. Well, Rishi, I want to thank you for coming on. That is Rishi Jaluria from RBC Capital Markets here on TI TV. AI inference platform Base10 has acquired Blackshell, an infrastructure startup specializing in secure execution sandboxes for AI agents. For more on that, I want to bring on Amir Hagiha, co-founder and CTO of Base10. Amir, welcome to the show. It's great to have you here. Great to be here. Okay, so just remind us here, before we get to the acquisition, remind us of the difference between Base10 and the inference provider category you play in, and then the classic cloud providers, you know, Oracle, who's talking about, and Microsoft, just so we can make sure we understand the business you're in.

11:34That's right. So a lot of the players in the CSP space or in the NeoCloud space, the main product that they offer is bare metal GPUs. GPUs that you can SSH to and do whatever you want with it. If I could get one massive billboard in San Francisco, it would say GPUs plus some runtime and model weights does not equal production inference. There's a lot more that goes into it to providing a fast, cost-efficient, and reliable, like 99.99 % uptime reliable inference. And that is what we provide. In simple terms, for people who maybe don't have a technical background, what are these things that you might need to make that runtime?

12:22Ultimately, to do inference well, it takes two things. And both of them are necessary but not sufficient. The first one is at the runtime level. Think of the software that runs on the GPU that has the weights of the model loaded and spits out tokens. The optimizations that you do at this layer can make a difference of do you spit out 100 tokens per second or 200 tokens per second? And that has a direct impact on how many GPUs you need to serve that traffic and therefore your cost basis. So that's one thing that it takes to do inference. Well, the optimizations at that layer. But there's a second thing, which is at some point, that one model running on that one GPU gets too much traffic that it cannot handle.

13:04And at that point, you need to horizontally scale it. You need to replicate it. You need to go and find GPUs elsewhere and put the same model and the same runtime on it and split traffic amongst those. That is a different kind of problem. It's not a PyTorch or CUDA problem. It's an infrastructure distributed systems problem. Where do you go and get these GPUs, especially when there is such a shortage? how quickly do you spin up these new replicas so that the traffic doesn't notice that things are slowness in cold start times? Right. That is the second problem that you need to solve as an inference provider.

13:39And then, you know, maybe silly questions here, but is that not also, does that not overlap then with, you know, the software layer that all these Neo clouds are trying to build out? I mean, they're also building their own infrastructure. And then also help me compare or contrast that to the software that the chip companies have? I mean, like CUDA, I mean, do they not overlap here at all? CUDA is the underlying software that all of them use. How you use CUDA is the question. And there's other layers of the stack where there's a lot of optimizations that can be applied that has a massive impact on the cost efficiency without affecting the quality of the model.

14:17On the NeoCloud side, yes, they are going in that direction. And I understand that. Like they rather, you know, provide inference and post-training higher up the stack as opposed to rent out bare metal GPUs. I get that. They are going in that direction. The question is at what pace, with what quality, and with what kind of reliability. One thing that cloud's limitations that they inherently have is that individually they're all unicloud. They're not incentivized to build a multi-cloud strategy because they themselves are the cloud. Where we're sitting, we are very much inherently multi-cloud. And that's important, not just from access to capacity or price negotiations, but poor reliability.

15:05The kind of customers that we go after require four nines uptime, 99.99 % uptime. You cannot provide that with a single cloud underneath you. You have to inherently be multi-cloud. Okay. So now, what is BlackSol? Let's talk about the company you acquired and why did you decide to bring them on? Yeah, let's talk about that. So maybe to get there, I want to sort of draw it out as to the trend that we're seeing in the industry where we're sitting. You know, as an inference provider, we've been doing this for a long time. The company is now seven years old. And really in the past couple of years, year and a half, inference went from being, hey, I'm going to ask my model a question.

15:48then it's going to give me the answer and I'm going to run with it, to I'm going to ask the model to do something that requires multiple steps and multiple reasoning steps and also actions. It really went from answering a question to taking lots of actions. We saw that. We saw that trend, but what we were doing to our customers is just giving back the response of the model, which it was, hey, I need all these tools to be executed. We ignored that. We just gave it back to the customer. And the execution of those tools, of those actions, was left as an exercise to our customers. Unsurprisingly, they kept coming to us and saying, hey, you do the inference layer.

16:32Can you also do the tools layer? Can you also not just be the brain, but also be the muscle? Because you're forcing me to go elsewhere to find the muscle. And what is that muscle? It's the sandbox. The sandbox is really the muscle that allows these models to take certain actions. And that's basically this controlled environment where you can test things out and you don't actually have to deploy them into the real world. That's right. You can control things out. You can take certain actions, whether that is run some LLM-generated code, which can be inherently risky. And so therefore, the security of the sandbox is really critical too.

17:10And the third thing that is important is the statefulness of the sandbox. You see the same things again and again. Having to reset up the sandbox from scratch every time has a massive impact on your latencies. That is where BlackSoul shines as a product, where they have built not just an environment where you can execute some code and die, but perpetual sandboxes that are stateful, that can pause and then resume to be very cost efficient, and do so extremely quickly with the right storage and network components. So were you guys thinking about making this acquisition prior to the Hugging Face OpenAI hack?

17:48Because this is where sandboxes really got introduced into what more people understand to be, you know, important to AI agents and, you know, them not going rogue. So did that accelerate your decision to buy them? No, it did not. I have been talking to Paul, CEO of Black Soul, since February. Okay. And we have been internally debating build versus acquire. Those were the only two options. But I mean, is this, so help me understand, Black Soul, would that, I mean, is that sort of these sandboxes? I mean, are they the only way to protect against, you know, these rogue agents? I mean, this week we've got the discussion that first it's rogue agents, then it's end of humanity, right?

18:37And, you know, maybe sandboxes are the only way we can protect it. I don't know. Is that the equation here? I don't think so. Sandboxes are not really a way to protect. Sandboxes are that muscle. And that muscle that takes actions really can be used to do nefarious things. And that's what you saw with the Hugging Face OpenAI saga, is that there were sandboxes being used. and those sandboxes were probably secure at the operating system level. But I think that the tricky part there was that those sandboxes were given access to the internet. Maybe not unfettered access, but still a way for them to be able to get to the internet.

19:18If they couldn't, we would not have news from it at this point. And so that's the critical piece, and the security of sandbox, not just at the operating system level, but also at the network, what they call egress access level? Does it have access to open internet one way or another? And once you can control that, and once you bring the muscle and the brain together, the inference and the sandboxing together, and monitor them very carefully, I think there are ways to... Is this not like, doesn't the sandbox, is it not an issue of, you know, security around the sandbox and making sure that it can't escape?

20:01I mean, is that not how it works? This is how it works. But as models get smarter and smarter, they try to get around the perimeter that you have set for them. And what you need to do is get better at that perimeter. But I think that perimeter is really at the network layers where we can have most of the impact in terms of controlling them. and also at the layer of, again, monitoring both the inference calls and the actions from the sandbox at the operating system layer holistically and being able to see what is happening and be able to quickly stop it. Right, right. Let me ask you one more question.

20:39You guys just did a funding round earlier this year. We reported that you were at$600 million in revenue in Q1. Where are you guys at now? A lot more, but please give me a few weeks and we'll be able to publicize it. And then will you come on our show and talk about it? Awesome. Well, Amir, I want to thank you for coming on. That is Amir Hagihot, co-founder and CTO of Base10 here on TI TV. Thanks, Brian. Blackstone has firmly embedded itself as a bank roller of sorts for the entire AI sector. One move that caught a lot of people's attention was its joint venture with Google, effectively a neocloud for TPU chips, which Blackstone said it would invest$5 billion into.

21:27The information's finance reporter, Dakin Campbell, has new reporting on Blackstone's plans for that venture, which he reveals in a new feature story about the private equity giant. I want to bring on Dakin to share more about what he learned. Dakin, welcome back to the show. It's great to have you here. Yeah, thanks so much. okay so let's talk about blackstone uh you know we have this joint venture with google we know uh how it came to be and it has a name now crux ai which i guess is what they're calling it you know i if we just take a step back here a little bit i mean when this deal was announced initially what was wall street's reaction to a pe giant like blackstone making such a big play into AI?

22:07Yeah, I think it was intrigue. The numbers weren't huge at the time, but Blackstone partnering with Google was enough to get people talking and people really thinking about what could this mean and what could be the implications of this. You know, there wasn't a ton that we knew about it At the time, they put out a press release. It said, you know,$5 billion and the partnership. And it had some talk about what the, you know, neocloud was going to look like. And it talked about sort of networking and chips and all the things that you would sort of expect a new neocloud to talk about. But there just wasn't a lot of detail.

23:00And what were the big questions for you as a reporter when that was announced good question uh the big question for me was is blackstone starting a new company with google it seems to be what it is that is not the kind of thing that blackstone typically does they typically invest in companies that already exist that already have um cash flows and other questions were, why did Google choose Blackstone? All the financiers and things like that. So we've now learned a little bit more about it. It's got a name, has a CEO. This gentleman who was at Google for a couple decades certainly is very familiar with Google's TPUs use and how to employ them in a data center, neocloud environment.

24:00We also know now that the venture has gotten much bigger. The ambitions for the venture have gotten much bigger. Are they putting more than$5 billion into it? Or what do we know about that? They are. Yeah. So I was able to speak with somebody who said that they're now looking to put multiples of five billion dollars into the venture they wouldn't give me a number or an amount uh of chips uh or of compute capacity that that would relate to but multiples is not you know one and a half times the five billion dollars we're talking north of 10 billion dollars more likely between 10 and 20 billion dollars the other thing i learned uh as i said the press release was a little bit vague the other one of the other things i learned is that the money is going specifically into chip purchases so the five billion dollars it was for five billion dollars worth of chips it wasn't for five billion dollars worth of chips and data centers and wires and land things like that it was specifically just for tpus so uh so the multiples will be multiple more uh TPUs, basically.

25:16Presumably, they're going to need chip purchases. I think they, well, at least for GPUs, they're about 60 % of the cost by some estimates of building a data center. So they're going to need not an insignificant amount of more capital than to build a data center themselves. Yes, that's right. Right. That's right. And from what? Yeah, go ahead. No, you're going to start. I was just going to say, from what we understand, this is money that Blackstone is raising from their funds, from investors. We don't have a great idea yet of exactly where it's coming from, but hopefully we can learn that in the coming months when we learn more about it.

25:59So, you know, as best we can understand, this is going to be, I mean, this is a neocloud play, I guess, similar to the ones that we've seen. And the main difference being the TPUs will be the underlying chips. I mean, can you talk a little bit about the broader impact this deal has had on the private equity sector? Are there other PE companies that have followed suit with their own similar types of deals or investments or joint ventures like this? Yeah, this is a little bit unique, which is why it got people talking in May. But we do see other Blackstone competitors getting into similar things like this.

26:42AKR, for example, partnered with NVIDIA and the Kuwait Investment Authority and Vistra Energy to create, they are in the midst of creating Helix Digital Infrastructure, which is a data center company. I don't think they've talked about it in the terms of it being a neocloud. I think it is more about building data centers and offering them to customers who maybe want to bring in their own chips or put their own software on top of those chips. But that's another example. KKR is building a company from scratch, similar to what Blackstone is doing. And that is a departure from what we've seen private equity firms do in the past.

27:35And that's interesting. Right. Now, you spoke with Jaskara, who is one of the executives at Blackstone, who is leading a lot of their work into these types of AI investments. What surprised you about that conversation? What did you learn from him? uh i think one thing that was really interesting to me was the curiosity he had so blackstone has been around for a long time they were founded in um 1985 you know that makes them uh 40 years old at this point plus. And they have, for most of their life cycle, have done a certain type of investing, which we've talked a little bit already. They have not really done, they've not really built companies from scratch in the past, though they have done a little bit of it.

28:33But he is really, he just moved to Silicon Valley. He is out there taking meetings. He is learning. He is thinking, trying to think creatively about how Blackstone can play a role in the AI build out. And from what my understanding and talking to him, he's open to different ideas and different models. And he is really acting as a, you know, a scout, as a point man for Blackstone. And I think the firm really felt it's huge. It's the biggest private equity firm in the world, $1.3 trillion under management or assets under management. And lots of parts of the firm are playing in the AI space. But I think they felt like they needed somebody to sort of coordinate all of that and to sort of act as almost like a salesman out in the world, sourcing opportunities, meeting people, finding new investment opportunities to make.