Starlink Takes on Telecom Giants, OpenAI’s Stargate Compute Hurdles, Nvidia Earnings Preview

14:07Gil Luria:What took place then? What happened?

14:09Akash Pasricha:So those three companies, Oracle, SoftBank, and OpenAI, had to come to the table and figure out how are they actually going to structure this effort? the announcement itself had come together within a matter of days. You know, it was kind of cobbled together. And then they had to figure out what is this joint venture? Who has ownership of it? How is this joint venture going to control data centers? Inside OpenAI, they had a few visions for what the Stargate was going to be. You know, one of those options that we reported was they were thinking about spinning out Stargate as a separate entity. and employees who work on Stargate today at OpenAI were, you know, they kind of thought that they might be going to that separate entity.

14:53Akash Pasricha:OpenAI also thought that Stargate could just be a financing vehicle where, you know, it's a separate, you know, structure that investors could invest in and the money would go solely toward OpenAI data centers. But, you know, the three parties were not, you know, they did create this LLC, but in terms of like actual projects. They did not pursue any. So Oracle, OpenAI, SoftBank, all were kind of scouting, you know, sites in the US to find out where might we actually want to build these data centers. But that work, importantly, was happening in like a disjointed way. They, you know, OpenAI would visit a site, Oracle would visit a site, and they couldn't really give developers a clear answer on who was going to be doing what.

15:36Gil Luria:And I just want to make something clear, though. I mean, at the outset though i mean was all this this the 10 gigawatts of compute i mean all of these players were involved in the effort was this all to be used by open ai even at the outset for sure um it

15:51Akash Pasricha:was it was for open ai's computer i mean we've reported a little bit on opening high's plans to maybe sell computing in the future if it has extra but um yeah it was it was all for open ai this was you know open ai's answer to how do we get enough compute because they couldn't get enough from Microsoft at the time.

16:08Gil Luria:Okay, so we have these sort of disjoint efforts here visiting sites. Ultimately, what you reported in your story is that OpenAI then decided they were going to do this on their own then without the partners at some point.

16:21Akash Pasricha:Yeah, OpenAI expected that a large portion of the 10 gigawatts would be owned and operated data centers. And they weren't sure where Oracle or SoftBank would play in that. But as we reported in the spring, OpenAI realized that it would be very hard for it to itself raise billions of dollars to build out a data center. You know, it's as we've noted in several different projects, the way to secure the best project financing is to have an investment grade tenant. And so in the spring, the Stargate team at OpenAI ended up having another meeting with Oracle, the executive team. And they said, look, what if we actually did 4.5 gigawatts together instead of what we're currently doing, which is a project by project conversation.

17:13Akash Pasricha:And so OpenAI told Oracle, let's do 4.5 gigawatts together. And one benefit was that they both viewed that by doing a large 4.5 gigawatt deal, they could kind of spread out the risk that projects might get delayed or they would come online at different times and they could sort of play around with different timelines. So that's kind of how they came to the idea that they would do one massive 4.5 gigawatt deal.

17:41Gil Luria:So this was OpenAI saying, hey, you know, it's a little disorganized here at the start. We think maybe we can build these data centers on our own. They go out, they try to pull together some kind of an effort, and then they decide this is too big a lift. And then they come right back to Oracle and the original parties with Stargate. And they say, you know, maybe let's not start with 10 gigawatts. Let's go to 4.5. And so now we're sort of back at sort of a smaller version of what the original Stargate was supposed to be.

18:14Akash Pasricha:Yeah, essentially. I think, you know, 4.5 is still a really large number. And they were signing other compute deals as well. But yeah, there was a decision to go to Oracle for the 4.5. And then at the same time, OpenAI still wanted to do its own data centers. So they were still at that time, as we get to 10 gigawatts, what can we do on our own?

18:39Gil Luria:Okay, so they haven't given up that idea of doing data centers on their own entirely. They've just said, we can't only do that. we have to lean on Stargate to some extent to get this compute. That's right. Okay. So now they're dual tracking it. They're saying, we're doing Stargate, we're doing our own data centers. You know, timelines for this are still fuzzy. Do we have any updated timelines on when we're expecting any of these data centers to be constructed, or are they being built right now?

19:08Akash Pasricha:I mean, in total, this is a multi-year effort. And so So Stargate, you know, ideally was kind of time to President Trump's term. But yeah, you're right. You know, they have all these things going on at the same time. And after that big Oracle deal, then they started working with SoftBank on another kind of data center relationship. So lots of moving parts. But yeah, to your question, this kind of capacity, even though it's committed and they're working on contracts for it, it is not yet deployed.

19:41Gil Luria:Okay. Okay. Now tell me, what is the net impact of all of these disjointed efforts? Maybe not disjointed, maybe they're coordinated now. It's kind of hard to tell who's talking to who, but clearly there's something happening. What's the net impact on the product roadmap that OpenAI has, and then ultimately its finances? Because all of this was in pursuit eventually of generating free cash flow, and we know that revenue hasn't really been the problem. You know, how are the two tied?

20:12Akash Pasricha:Yeah, I mean, what we found in our reporting, which I think is wildly known, is that OpenAI's revenue is quite tied to compute and their ability to train advanced models is also directly tied to how much compute they have. And so, you know, I think it's fair to say that as Stargate kind of dragged on last year and they were pursuing multiple different ideas, that did cost them some time. You know, one thing we did report was that by the end of 2025, they were supposed to have 10 gigawatts with Oracle and SoftBank and they had 7.5. Now, 7.5 is still many, many gigawatts. And so, you know, it's hard to say that that's a bad thing.

20:52Akash Pasricha:But I think the time it took to figure this out, you know, they were probably, as we reported, they were scrambling to sign deals with other companies like Amazon and Google. And so I think on the whole, OpenAI still has access to a lot of compute, and they got very close to that 10 gigawatts. But every company in this space is constrained by compute. So I'm sure that OpenAI competitors like Amproopic will be studying the playbook and trying to make sure that they can do the same thing very quickly.

21:24Gil Luria:And so what should we be watching for here in terms of this story moving forward? Should we be watching for the data center developers to make announcements saying, hey, we've got one, two, three gigabytes, gigawatts, gigawatts up and running? Is it going to be OpenAI saying, hey, you know, we've secured this amount of financing? I mean, we know they're raising funding, but, you know, what should we be looking for?

21:51Akash Pasricha:Yeah, I think one thing that I'll be looking for in particular is how OpenAI decides to structure some of the chip purchase agreements. Because on that 4.5 gigawatts, Oracle has already signaled to investors that it might not be on the hook itself to buy the chips. Its customer might bring its own chips or there could be interesting leasing arrangements. So, you know, this is just kind of a scale we haven't seen before. And so I'm sure that OpenAI will be, you know, talking to NVIDIA, talking to Oracle about some creative way to structure the financing side of things. Because, you know, that many gigawatts is going to require hundreds of billions of dollars on chips and other data center infrastructure.

22:35Akash Pasricha:So we'll be very interested to see what happens on the financing side in the future. Great.

22:41Gil Luria:Anissa, I want to thank you for coming on. That is Anissa Gardizi, our cloud and compute reporter here at The Information. NVIDIA reports its quarterly results this Wednesday after the market closes. Shares have been flat this year so far. Over the past six months, it is up around 5%. There are a number of headlines people will be watching for guidance on. Chip sales in China, impacts of the memory chip supply chain crunch, updates on the Vera Rubin family of chips. To break it all down, I want to bring on Gil Luria, Managing Director and Head of Technology Research at DA Davidson. Gil, welcome back to the show.

23:15Gil Luria:It's great to have you here. Thanks for having me. What are investors going to be watching for this week with NVIDIA earnings?

23:22Anissa Gardizy:Yeah, I think those are the main items. First of all, investors want to see that NVIDIA is still on track. The growth is still out of this world. There's a company growing from$200 billion to over$300 billion. So everybody's going to want to make sure that we're still on track. Then we're going to ask some more specific questions around China. How many of these H200 chips have we sold? Do we still have inventory? There'll be questions about memory and the impact of the much higher prices for memory chips on NVIDIA. How much of it is going to impact margins? How much are they going to pass on to consumers?

24:00Anissa Gardizy:And then investors are going to want to hear about next year and the year after that. But I doubt we're going to get much on that. But that'll still be an area of interest. And then finally, there may be questions about open AI and the investments there. There appears to be some confusion about that. But it is a very important venture for NVIDIA to invest that much in open AI. So there are likely to be questions on that as well.

24:27Gil Luria:You know, when we've had you on the show previously and other analysts, we've talked about NVIDIA as a bellwether for the AI sector at large. Do you still think that it is that bellwether? I mean, will we see the entire market and cloud companies, AI companies, software companies, everyone used to move based on NVIDIA results. Is that going to happen this week as well?

24:50Anissa Gardizy:probably not i think we're past that stage with nvidia part of it is that they've raised the bar so high that everybody expects them to exceed it so there's a little less tension but they also can't exceed it by as much as they used to because the numbers are so big so it's less of a bellwether plus now there's this notion that google may have won and that comes with tpus not gpus so the the AI trade that was really NVIDIA centric for a while has now expanded to more Google and Broadcom and A &D and the optical companies and the memory companies. And what's really interesting right now is that those stocks are reflecting a more optimistic view of AI, an investment cycle that could last for several years, while NVIDIA's stock is actually trading at a multiple that implies that this may be peak here.

25:46Anissa Gardizy:So at 24 times earnings, NVIDIA is trading at only a slight premium to the market, in spite of the fact that it's growing revenue at 50 % with 75 % gross margins. If NVIDIA can show that it will continue to grow at all into next year, that's the opportunity. Because if you look at the memory chips and optical chips and Google and Broadcom and AMD, those stocks are implying that the AI cycle will continue for a much longer period of time.

26:18Gil Luria:Why do you say that? Why do those stocks indicate that to you?

26:21Anissa Gardizy:Their multiples are significantly higher than they used to be. And to justify the multiple, especially if you're looking at Broadcom or AMD or the memory companies, you'd have to believe that there's follow through in the AI investment cycle that will keep the demand for those products going. Where with NVIDIA, again, this is on the low end of their historical multiple range, meaning this is setting up an expectation for the cycle to peak this year. For you to believe that AMD and Broadcom are worth more than 30 times earnings, you have to believe that the investment cycle will continue and ramp up for them which would only happen if there's more capex next year than than this year amd and broncom are not going to ramp if the the cycle rolls over and yet their valuation implies that it will so now looking at the the recent stock

27:22Gil Luria:price for nvidia we mentioned at the end in the intro it hasn't really i mean it hasn't moved this year, hasn't really moved in the last six months. Maybe this is exactly what you're pointing to, but what will it take then to move the stock higher now? I mean, if it's not moving higher, it's moving lower. And I find it hard to believe the stock's moving lower, but at the same time, I mean, what is the catalyst then, do you think?

27:47Anissa Gardizy:Yeah, I mean, at the very latest, it's as they start looking into 2027 and we start hearing from their customers that they intend to spend more in 2027-26, that'll probably do the trick. But to be fair to NVIDIA, their stock has outperformed the rest of the MAG-6 or MAG-7, depending on how you include that, because really it's all those other companies spending on NVIDIA chips. There's really a wealth transfer from Microsoft, Amazon, Google, and Meta, and Tesla, to a certain extent into NVIDIA. And so they have outperformed that group. But again, the trade is just broad. The AI trade, people got a little born with NVIDIA as the AI trade and expanded it to all those other categories.

28:34Anissa Gardizy:That's really all that's happened when, realistically speaking, there is no AI trade without NVIDIA. Nobody's building data centers without NVIDIA chips as the primary component, not anytime soon. So as we get towards the latter part of the year and start hearing from companies about their CapEx plans for next year, that should allay the concerns that this is peak spending. And at the earliest, that's going to be the July report from Microsoft. Microsoft's on a June fiscal year end, so they're going to give some guidance into their next fiscal year, which will include 2027 on their July call. So as we go through the next couple of quarters, we're going to start getting signs that whether or not this year is peak year.

29:23Anissa Gardizy:And if it's not, that's how NVIDIA outperforms.

29:26Gil Luria:Now, China is always a focus, at least on the earnings call. People are always watching for what kind of commentary the company gives around sales in China, whether or not they're even counting sales in China now, given the current export controls. I think my question is not so much what we're expecting to see, because that will change week by week. And, you know, look, we've got a whole new family of chips coming out with Verorubin. And I think we're going to go through the whole cycle again when people realize that H200, it's better, but it's not even nearly what we could be getting with the Verorubin.

30:02Gil Luria:So my question is not that. My question is, is China more or less important to NVIDIA today than it was 18 months ago?

30:12Anissa Gardizy:Oh, it's a lot less important. It's a rounding error right now. Whether they sell a couple of H200s or don't sell those, those are already in inventory. They've already been accounted for. It's just noise at this point. The much bigger question is, is NVIDIA still in the China market? China used to be at least a quarter of NVIDIA's results. Now it's, again, a rounding error. But in terms of the market in the future, China will continue to be at least a quarter of the global market for GPUs. for XPUs, for whatever combination of those chips you want to see. And NVIDIA wants to be in that market.

30:51Anissa Gardizy:They're caught in the middle of an internal debate in the U.S., of the U.S.-China trade tensions, of the U.S.-China geopolitical tensions. So I don't expect this plot to be resolved anytime soon, but it is very important because it is a quarter of the global market. And if NVIDIA has to see that market, it's going to be Chinese domestic suppliers that take over that market, scale up, and then present a global threat to NVIDIA in other markets. So it is very important for NVIDIA. They're doing what they can, but I don't expect resolution anytime soon.

31:28Gil Luria:Right. And it sounds like what you're saying is it's much less important to the results as they are today and as they're being reported. But it's much more important in the sense that if NVIDIA is going to get that bump that we're all looking for or expand that multiple any more than what it is right now, China is going to have to be a factor there. And so it's actually even more important than it was because this is one of the catalysts that people might be looking for.

31:53Anissa Gardizy:That's right. That's right. And if they don't get that China market back, it takes up a lot of the upside and, again, creates some long-term threat as the Chinese improve and ramp volumes and ramp capacity in the ecosystem that's outside of the ASML to TSMC to NVIDIA ecosystem that is so dominant right now.

32:17Gil Luria:Last question for you, the memory chip supply chain crunch. What are you expecting there in terms of commentary? And when do you think all this gets resolved?

32:25Anissa Gardizy:So for NVIDIA, I wouldn't expect too much of an impact. First of all, they have very high gross margins they can absorb. It's not that much of their componentry. And they can pass along the rest of it to their consumers. I'm a lot more worried about companies that make low-end PCs and low-end handsets. Not only can they not afford to buy memory, they're not even getting allocation. So if you need any of those devices, I'd go to Best Buy right now because pretty soon those devices, if they're even available, will be a lot more expensive.

33:00Gil Luria:Grant, Gil, I want to thank you for coming on. That is Gil Luria from DA Davidson here on TITV. Thank you. AI has very much turbocharged the business of transcription, and Otter AI has had a front row seat to that story. As the company has grown, so too has competition, but the company has had more than 25 million users as of last year. Joining me now to talk about what is next for the business is Sam Liang, CEO of Otter AI. Sam, welcome to the show. It's great to have you here.

33:30Sam Liang:Great. Thank you for having me here.

33:33Gil Luria:So I think a lot of us, certainly journalists, know of Otter AI because we use Otter AI to transcribe all of our interviews. and it seems like a pretty simple product and it works well and, you know, it does the transcription. And I think that business has become a little more popular now as speech to text has taken off. But what is the structure of AutoAI's business today? I mean, is it really just the, you know, consumer subscriptions, enterprise subscriptions, transcription tools, is that it?

34:07Sam Liang:You're right. We are actually probably the first who created this category. Back in 2018, we launched our product with high accuracy, real-time transcription. However, as you mentioned, there are more competitors now. There are more meeting note-takers. We have actually been moving beyond just meeting transcription. I'm glad that you used Otter. to transcribe your interviews. But once you have dozens, hundreds, or thousands of meetings, especially in enterprises, if you think about any enterprise, they have sales team, marketing team, product engineering. Every team has tons of meetings. Today, if they just use a simple note taker, all the notes are fragmented, they're isolated individual notes.

35:11Sam Liang:What we're trying to do is actually connect all the dots. We're building a meeting-centric knowledge base for enterprises to get more productive. Right.

35:22Gil Luria:So now let's talk about the competition there. I'm curious as to how you think about your differentiation strategy, because as you said, transcription is getting more competitive. And the product that you just described, which is the meeting intelligence drawing on thousands of meetings and conversations, I mean, this is something that I'm thinking of even all the customer support companies, customer support AI companies that, you know, they're doing it themselves with their thousands and tens of thousands of conversations. So how do you think about standing out?

35:53Sam Liang:Yeah, we have a new philosophy or very different philosophy. We see that voice is becoming the primary interface for enterprise intelligence. If you think about it today, more and more documents, written documents, are being generated by AI now. We see that maybe in 12 months, you will never need to write anything. Well, reporters still need to write things, but for most people in enterprises, they will rarely write anything because voice is the primary interface. So they spend most of the time, they're already spending, you know, most roles other than engineers who write code, most roles are already spending most of the time in meetings using voice.

36:45Sam Liang:And then, you know, Google document notion, any data in the database, CRM, everything will be written by AI. So what we do is to optimize the voice communication and optimize the synthesis of information, connecting dots across multiple meetings, not just the one meeting, but all the meeting series in your team. A lot of teams have recurring weekly meetings or bi-weekly meetings and connected dots between between meetings between different departments, as I said, sales, marketing, product, engineering, design, to create this meeting context layer. And we're also incorporating data from document, from email, from Slack messages.

37:45Sam Liang:With all this, then we build the organic workflow on top of that.

37:52Gil Luria:So I'm with you, and I think that's a really compelling offering. I guess my question is, again, why does Otter have a leg up here compared to, I don't know, why couldn't Salesforce do this? You mentioned Slack. You mentioned all the other dozens of transcription tools. I mean, what is the moat here for your company?

38:13Sam Liang:Well, there's several moats. One is a data and user moat. You mentioned 25 million. Actually, that number has increased a lot. We have more than 35 million users now. We've been doing the longest. We have the biggest voice data set, hundreds of millions of hours of voice data. And also, again, this product philosophy, I don't see most of the other players doing this, building a meeting-centric knowledge base. Most players still stay at the individual meeting note taker stage. Even Microsoft Zoom, if you look at their product, I don't know if they use their product for meeting notes, they just generate a basic meeting transcription summary, but without a cohesive system that organizes hundreds of thousands of meeting notes that allow you to search, that allow you to create, that allow you to organize based on projects, based on teams.

39:29Sam Liang:So it's just that practical philosophy where we're ahead of everyone. Most people haven't realized the power of that knowledge base yet.

39:40Gil Luria:So you've got 35 million users. How big is the company now in terms of revenue?

39:47Sam Liang:We announced that we were over$100 million in ARR last year. That's when it was 25 million users. Yeah, we're way higher than that. We haven't announced our latest revenue numbers yet.

40:07Gil Luria:I want to ask you a question about the underlying technology here. So which models are you using in the background? What have you learned from all the different models that are coming out? And do you see the strength of all these models converging over time? Are you in that camp? Or are you very much still fundamentally on the side of these models are going to duke it out for the next three, four years, and we have to pick the one that is best?

40:35Sam Liang:For speech recognition and speaker recognition, there are two technologies. Speech recognition is the technology to transcribe the voice, and speaker recognition is technology to separate multiple speakers in the same room and also identify the voice of each speaker. So we have proprietary technologies for both. This allows us to provide higher accuracy. For example, on Zoom, if everyone speaks in a quiet environment, the transcription is relatively easy. But once you have multiple people in the same room, if there are AC running in the background, or if you're meeting someone, for reporters, you could meet someone at a restaurant or Starbucks.

41:38Sam Liang:With that background noise, most companies' transcription will be really bad. We have this special technology to handle the background noise.

41:50Gil Luria:And this is a model that you have built yourself that does this?

41:55Sam Liang:Yeah, we built this over the last 10 years. And secondly, the speaker recognition is really critical. On Zoom today, if you have two or three speakers in one room, Zoom cannot separate them. Zoom cannot identify who is who. We are able to separate them and identify the voice of each person. This is very critical because once you want to build authentic solutions on top of the meetings, you need to know who said what and who is taking on that action items, right? Whose point of view, people may have different opinions, different ideas. You need to attribute that ideas to the right person. So this is why speed correctation is so critical.

42:48Sam Liang:So again, by building this in-house end-to-end technology, it allows us to innovate a lot faster, it allows us to provide a higher quality. And also, you know, we can operate with much lower cost. If you're just using a third-party API, you're going to pay a lot to access that API.

43:10Gil Luria:Last question for you. As you continue to build more and more under the hood and you start to connect the context of all these different applications that you're talking about. How do you ensure privacy through all this? How do customers know that they have the same privacy standards that was the case five, six years ago maybe when there wasn't as much to it? It was just a more simple login transcribe play interface.

43:39Sam Liang:It's not so much different than most of the other cloud-based services. right? Google document, Gmail, or Slack. All the data is stored on the cloud. We use similar security models, encrypt the data. But also for enterprises, we have created a lot of data retention mechanism. You know, enterprises can configure the system to remove audio, let's say after three days or remove the transcript after 30 days and just to keep the summary. And also in terms of access, we have a very sophisticated access control system. You can specify who have access to certain content. You can create, similar to Slack, you can create private channels so that only a limited number of people can access certain meetings.

44:39Sam Liang:So there's a suite of access control and security mechanisms. We have SOC 2, Type 2 compliance, we have HIPAA compliance. We passed the security review of a lot of financial companies who have the most stringent security review system. Great.

45:03Gil Luria:Well, Sam, I want to thank you for coming on. That is Sam Liang, the co-founder and CEO of Otter AI here on TI TV. That does it for today's show. A reminder, we are on this stream Monday through Friday at 10 a.m. Pacific, 1 p.m. Eastern. I want to thank you all for tuning in. We really do appreciate your viewership. Make sure you subscribe to the information on YouTube and follow us on X, Instagram, TikTok, and check us out wherever you get your podcasts. I'm already excited for our next show tomorrow have a great rest of your monday bye-bye for now