Inside DeepSeek’s $7.4B Funding, Broadcom’s $35B Backstop, Nvidia’s Inference Rise

0:13Welcome, everyone, to The Information's TI-TV. My name is Akash Pasricha. It is Tuesday, June 16th. New this morning, SpaceX has finalized its$60 billion deal to acquire Cursor. It is expected to close next quarter. spacex shares of course continue to surge they are up more than 60 percent from their original ipo price the company is now valued at more than 2.5 trillion dollars as of this morning first up on today's show the information's asia bureau published exclusive reporting that deep seek has closed its first funding round will bring on our asia bureau chief to share with us how big the funding round was, how much the company was valued at, and why it was an uncommon deal structure.

1:01We also have an exclusive story that Polymarket and Calci are taking steps to battle fraud on their platforms. We'll then dive into new reporting on Broadcom's bold move to boost demand for its AI chips and NVIDIA's rising share of the AI inference chip market. and we'll wrap the show with some analysis on Fox's$22 billion acquisition of Roku. It's going to be a fun show, so let's get right on into it. DeepSeek has closed a big funding round, although the deal structure is slightly uncommon. Joining me now to discuss that is Jing Yang, our Asia Bureau Chief. Jing, welcome to the show. It's great to have you here.

1:40Thank you for having me back, Akash. Although I would have to say that it's more than slightly unusual. In my seven years... In my seven years covering VC deals, I've never seen a structure like this. Okay, well, so let's get the brass tacks here. So how much did they raise? What are they valued at? And let's get to the deal structure. Yeah, so they raised more than$7.4 billion equivalent. So the funding round was mostly denominated in yuan, RMB. So in that term, they raised 50 billion yuan. And I would categorize the investor's lineup into three types. First, well, four types. First is the DPC CEO Liang himself, who wrote the biggest check, 20 billion RMB.

2:27And then we have the second line-up, which is basically the who's who of China's tech giants. You have Tencent, JD.com, NetEase, and also EV battery giant, CATL, all writing checks of different sizes. Then we have the third, usual suspects of some of China's top VC firms. And then the fourth type is China's National AI Fund. Okay. Well, let's go to the deal structure. Now you've piqued my curiosity. Why is the deal structure so crazy? Yeah. And I want to set this up by the different categories of the investors because basically, we can put Liang himself as the CEO and a controlling shareholder in one pocket.

3:10We don't care how he wants to use his money, right? And for the remaining three categories, with the exception of China's National AI Fund, which is Step Act Money, the two others have to actually wire their capital into a limited partnership structure managed by DPC CEO Liang Wenfeng instead of putting that capital into the company directly. That's unusual feature, number one. And then number two, as a part of that limited structure arrangement, these investors have to subject a lockout period for five years, basically during the five years, with I think the only exception if the company goes public within that five -year period, then these investors cannot sell their shares to anyone else.

3:56This is unusual because you've seen in big VC deals like OpenAuthropics, et cetera, in the US or buy downs and others in China. Usually when you have a very big company that is raising billions and billions of dollars, what happens after the rounds are closed is that these shares get changed hands pretty actively on the secondary share market. market because some investors might want to exit before the IPO happens and then they have the avenue to do that. But then this would also mean that probably your shares will end up in investors that you may not even be able to fully trace, like who ends up really holding your shares.

4:38So DeepSafe wants to eliminate that. The CEO wants to eliminate that. So to that end, the structure they come up with is this five-year lockout. And then second, on the usual part would be, since you are actually investing in this limited partnership managed by the CEO, that also means that you will not have any voting rights at the company that is DeepSeek, which is essentially what you are funding. The investors get the usual VC privileges, such as financial information, priorities for future funding rounds, but not voting shares. So all of this comes down to control. What is it exactly that DeepSeek wants to do with the company?

5:25Is the vision that they are painting, is it an unusual sort of vision for the company? Do they have sort of a contentious ambition at all? Or is it just really that they're very protective of their shares? What's the rationale here, do you think? It's certainly not out of protectionism. So we reported many times that DeepSync CEO Liang is a technology co-idealist. He's a foreign believer of open source technology. He believes that, according to our reporting and people who are close to him, he believes that AI should be open source so that it can be made available to all and benefit all. And that vision obviously takes a long time to achieve.

6:15And so now that he's raising money, he wants to make sure that the investors are joining him in the long run. And he wants to weed out people who are just here for quick gain, quick exits, so that he can sort of get the benefits of both worlds. On one hand, he has a fresh war chest to invest in compute so Dipsy can continue to innovate with its models. On the other hand, you know, Dipsy's previous success was largely credited for the company never raising any outside money so that it's free from commercial pressure that most startups are subject to. And then in this way, he can also make sure that the company can continue to run, hopefully freely, of that external commercial pressure.

7:03It is kind of interesting to me that the company's mission is so centered on open source technology, and yet the deal structure seems to be, by what I'm understanding, one of the most closed deal structures that you say you've seen in so many years covering startups and venture capital. Kind of an interesting dichotomy. Jing, I want to ask you about where DeepSeek's product roadmap is at and specifically where that roadmap is at in the context of Anthropic coming out with these models that people are scared of being so good. I mean, is DeepSeek still the threat to North American AI models that we once thought it was?

7:53First, I would maybe try to clarify the sentiment that I gather from talking to AI developers and protect practitioners in China. They don't necessarily revolve their lives and the railway innovation around America. They don't necessarily say we have a goalpost, which is the US or Antarctica or whoever, and we want to outcompete them. Most of the metrics are very realistic and pragmatic. they are mostly driven by what they believe is the way that AI should be. That's number one. And then so the threat, I think narrative is more maybe from the bystanders, for the lack of a better word. And then DeepSeq, as we reported before as well, they took like over more than a year, year and a half to come up with a next generation model.

8:46So then when that model dropped, it kind of underwhelmed a little bit and then they were also considered for being late to the coding frenzy. They joined that game relatively late if you see what others have done in China and the US. However, I would not underwrite DeepSeek just yet. I believe that with the way we are living in such an unprecedented era in so many ways that as a journalist I would just not write off any possibility at all so I'm very curious to find out what Dipsy is going to do next with model updates and in terms of a roadmap so yes, the company as we reported has been taking baby steps in trying to commercialize its technology, whether it's in the way of having an app.

9:47Their app, by the way, maintains at a level of 40 million daily active user bases, which is not huge, but also not insignificant. And also maybe through enterprise customer sales like API access, they are trying that. But then again, that's not their goal. Their goal is not to turn a profit or to increase the revenue at any time soon. They want to maintain that sort of idealistic, you know, ethos, if you will. Right. And look, I mean, by the way, I tend to agree with you. I mean, look, I think we are just one model release away from the entire AI game as we know it flipping on its head. And we saw the first DeepSeek moment, and it seems, you know, inevitable that there will be more wrinkles in the plot with respect to which model is better, which one is more efficient.

10:44We know that DeepSeek is open source, which means that people can actually work with it a little more than Anthropic or OpenAI for that matter. And on the show, we have covered the potential of open source models. It is certainly a class of models that people are getting more excited by. And so I think you're exactly right that we cannot write anything off. And by the way, this funding, I mean, now they can go out and hire talent, right? I mean, the talent war is something you touched on in the story. Yes. I mean, they have been losing a lot of key talents. These researchers went and joined the most established tech names in China or in the U.S.

11:23is certainly something that has prompted one of the key reasons that prompted them to change the stance on external money. And then we were to see that what they can do to, you know, if they can hire more people or maybe even bring back some of the people that have left. And then circling back to this unusual nature of the structure, I think to play the devil's advocate here, if maybe this is like just a necessary evil, like a necessary means to achieve a grander, a greater good end. A similar parallel I would draw is OpenAI and ThoughtBik, especially OpenAI believing that, we want to develop where we want to get there first to be full bad actors do, but that we also cannot afford to have such powerful technologies be open source and available to everyone.

12:13So that's why we have to keep it proprietary, right? You know, everyone can have an opinion about whether the end justifies the means. Same thing with OpenAID, same thing with Anthropics and Visual Deal Structure. But then again, anything can be beyond our imagination as things go forward. Great. Well, Jing, I want to thank you for coming on. I know this is a story that you have reported on very closely. It is great to see that the funding round has finally come to fruition. And Jing is our Asia Bureau Chief here at The Information. Thank you for joining us. We will see you very soon. Polymarket and Kalshi are taking extra measures to address fraud that is occurring on its platform.

12:58My colleague Michael Rodden wrote an in-depth story outlining the current state of those measures. I want to bring him on to talk about what he found. Michael, welcome back to the show. It's great to have you here. Hi, gosh. Michael, how much fraud is happening on Polymarket and Kalshi? quite a bit apparently um this is quite interesting because these companies are growing really fast you know you can bet on anything from the weather to the war in iran and polymarket and calci have this fierce rivalry they want to get as many customers in the front door as they can um obviously that's great for revenue you know they get trading fees um but it also comes with a problem uh and that's that they didn't really have the right controls in place to see whether these customers were real or whether they were you know part of criminal fraud rings trying to test credit cards uh in order to get you know free sign-up deposits or liquidity rewards offered by the programs um and that's a problem because one it's expensive um you know when you have fraudulent charges coming through your systems you're you're on the hook to pay those back so it it means that like the revenue that these companies making it's a a fraction of that has to be refunded.

14:10So it's a business cost. And then on top of that, it really irks their payment providers. So companies like checkout.com, their banking providers, they get really sort of antsy about this because the card networks, Visa and MasterCard, they really don't want to have to deal with all this fraud coming through their platforms. So what I found was that Polymarket and Cauchy have been hiring external consultants to come in and help them like bring that cost way down earlier this year it jumped above the the rate that normally companies want to stick by which is about one to 1.5 percent of their total volume being fraud and so they've been trying to bring that down but yeah it's just another challenge that these companies are facing you know they've got insider trading problems.

15:04They've got legal problems with the state lawmakers that want to ban them. And then on top of this, they've got really boring fraud problems that a lot of fintech companies face, but they just are learning to become more mature companies as they grow up. Right. So there's a lot to unpack there. So let's talk about the checkout.com and Visa angle to this. So is that the dominant way through which Polymarket and Calci are detecting fraud right now? Is they're getting flagged from the payment companies or are they detecting any of it themselves? Yeah, so this is an interesting question. Basically, the companies would have realized that they'd got all these fraudulent chargebacks, but they didn't really know why or where it was coming from.

15:50um basically how this works was early this year visa was pretty upset because the the rate of like disputed charges fraud backs people depositing money trading instantly with it but then the funds weren't clearing um that creates a lot of problems as well uh visa would go to checkout.com which is the payments provider a startup based in london um which both polymarket and kausha use and it said hey, can you do something about this? It's really expensive. And if this keeps happening, you get financial penalties. If it becomes really bad, we'll just cut you off. So the company's decided to hire an external fraud monitoring company called Riskified to come in.

16:37And basically, a company like Riskified, it has a long track record of being able to detect how customers interact with the platforms uh you know where certain card numbers might be coming from that might be illegitimate um it basically knows how to monitor for this stuff and because polymark and kalshi they're they're quite young companies in the grand scheme of things they don't really have this long detailed history of understanding how customers interact with the platforms you know it's like a little deposit here and there normal is this big rush of um you know similar card numbers from one bank um is that normal for that to be rushing the program and then pulling back out they just didn't really know so it's like they've brought in external um expertise to try and clamp that down right and and and so i just want to make sure i understand the train here.

17:34So Visa tells Checkout.com that... Well, Calci uses Checkout.com to check out. Checkout.com uses Visa and Visa inevitably has to get paid back. So Visa doesn't get paid back. Visa tells Checkout.com we're not getting paid back. Checkout.com says to Calci, this is all fraudulent? Is that the train here? Not all fraudulent. It's not just Calci. It's Polymarket as well. But generally... I use them interchangeably. They're all the same to me. You're hitting onto this problem that is very common in fintech circles. And it's like Polymark and Kalshi, they are gambling platforms, but they're also fintech companies.

18:15And there's a lot of different counterparties in all of these relationships. You've got a bank. You've got a payments processor. You've got the card networks. You've got the actual platforms themselves. There are other payments processors involved in it. there's this long line of responsibility that connects me the customer to calci so it's like i have to go through all these intermediaries and there are problems through like the whole chain um and if there's one sort of weak link in that chain it creates all these problems um right and so there is this this difficulty in all these programs that you need to have the right controls Otherwise, it causes one expensive problems, seeding into revenue.

19:00And two, you know, the risk that you'll just get cut off and you have to, you know, find a new payments processor. You know, earlier this year, Polymarket lost one of its banking providers, BMO. And so it's had to like find replacements for that. So it's like it's a problem. And, you know, there's there is a lot of fintech companies out there willing to do this. but it's a big challenge to try and off-board from one payments process and re-board onto another one. At the same time, you just don't want to irk the payments. Right. Okay, so now one of the key points of your story was that Polymarket and Calci, they have started to take measures in place to try to address this.

19:41Riskified is one of the platforms that you said that they have gone with. I was surprised, though, to hear from your reporting, you found that they are finding ways to work around the around riskified using a platform like Aeropay explain that well it's not exactly working around riskified but there's like I said there's a million different ways to upload money onto Polymark and Kalshi Aeropay is another one of their providers which is a very complex thing called ACH transfers so that sort of goes around the card networks uh and helps you just move money from your bank straight to uh polymarket or koushi um but on the on the same side you know polymarket has this big uh cryptocurrency based uh business uh which you know doesn't really interact with these things either um i'm not sure sure they're so concerned about that um because it is a bit of a wild west and i'm sure they're happy to to keep the money flowing on that side.

20:43Koushi also has some sort of tokenized business that helps get around that as well. But for the majority of the users, if you or I wanna place a bet on what Elon Musk is tweeting, we're just gonna get our debit or credit card, go onto the site, pop it in, and then start gambling straight away. And that's how regular users will approach these companies. So it's important for Polymarket and Koushi to actually make sure that they're really keeping a lid on chargebacks, fraudulent charges, disputed transactions, things like that. Right, right. Great. Well, Michael, I want to thank you for coming on.

21:21That is Michael Brodin, our fintech and finance reporter here at The Information. Last week, Broadcom announced an AI infrastructure partnership with Apollo and Blackstone anchored by a$35 billion financing commitment. The first project will support one gigawatt of compute capacity for Anthropic with the broader partnership expected to enable more than 20 gigawatts of AI capacity through 2028. But new reporting from the information shows the deal represents a risky move by Broadcom to boost demand for its chips. My colleagues Dakin Campbell and Valida Powell wrote the story. Dakin joins me now to share what he learned.

22:02Dakin, welcome back to the show. It is great to have you here. Walk us through this. What exactly was the agreement that Broadcom agreed to? Yeah, so this is a$35 billion purchase of chips. Anthropic is going to rent them. And Broadcom basically to get this deal done provided a backstop, a guarantee on some portion of the chips. And so really what it means is if at the end of their useful life or at the end of the lease, if the chips are worth less than what everybody thought they might be worth, Broadcom will have to step in and provide some money or otherwise guarantee on those chips. So it's an interesting, it's a backstop, what we've been calling it.

22:59It basically requires Broadcom to step in if the value of the chips declines faster than everybody expects. So is this effective like co-signing a loan or is it slightly different? Good question. A little bit, I guess. I wouldn't say, I'm thinking about like my student loans. I don't think Broadcom is - A little more of the student loan, even though those are quite - Yes, right. Thankfully, college didn't cost that much, although it's working on it. Yeah, I mean, Broadcom is basically saying, you know, if Anthropic defaults on its lease and the chips can't be resold for a certain price, that they'll step in.

23:45So, yeah, I mean, they're co-signing it to some degree. okay so um you know you also talked about uh the transaction involving a special purpose vehicle spv to buy and lease these chips can you walk us through how that fits into this equation yeah so uh an spv is basically like a box um uh it sells debt uh to take get money into the box and then it basically uses that money to buy the chips. So then the SPV, the box, basically owns the chips, and then Anthropic agrees with the SPV to then lease the chips. So it's a pretty common structure in the financial markets. It's certainly more complicated than if Anthropic were to buy the chips themselves or if they were to just lease them from some corporate provider.

24:44but it is somewhat traditional. We are seeing more use of SPVs in the AI build-out. We saw a big one used to build a meta data center down in Louisiana, and this is yet another example of that. Okay, and now what about the different tranches of debt? How does that fit into the equation here? Right. So this has three tranches of debt, A1, A2, and B. The first two are the A1 and the A2 are what benefit from Broadcom's backstop. Because of that backstop, they're rated investment grade. The B tranche does not benefit from Broadcom's backstop, and so they are non-investment grade. So it's actually pretty interesting.

25:36If you're holding the B-tronch, you don't have Broadcom standing behind you. And so if Anthropic were to default on their lease, and if the chips were to not be worth what you thought they would be, your credit exposure is actually too Anthropic. And so a couple of people we talked to made the point that if you're a bond investor, there aren't many opportunities for you to get exposure to Anthropic, this big high-flying AI lab. And so this is one way for you to do that. So it comes in a non-investment grade option, but this way you're getting a healthy interest rate and you're taking some credit exposure to Anthropic.

26:30How does this deal compare and contrast to the deals that NVIDIA has struck? So it's similar in many ways. I think what is different for Broadcom is Broadcom is not NVIDIA. They are not as financially sound as NVIDIA. I mean, we just saw this week NVIDIA is selling debt for, I think, the first time in five years. Broadcom, on the other hand, has$65 billion in debt. They've got less than$20 billion in cash on their balance sheet. And so they are in, I don't want to say precarious, but in relation to NVIDIA, they are not in as strong financial shape. So by lending their backstop in this way, they're taking on a little bit of additional debt-like obligations, and it affects their credit rating a little bit.

27:29Do you expect that we'll see more deals like this to come, NVIDIA, Broadcom, other chip companies? I think so. I mean, I mentioned earlier that we've seen a few SPVs now set up. You know, the people we spoke to in reporting this story think that this is going to be a model going forward. It's the first time this has been done in this size. You know, I think meta was 30 billion. This is 35 billion. And we've also seen Broadcom talking about, you know, building this platform with Apollo and Blackstone that could see them, you know, enable 20 gigawatts of compute. So if you think this was just for one gigawatt, we've got another 19 to do.

28:20And so I think it's very real possibility that this transaction serves as a model going forward for Broadcom's larger platforms. Do you think this ultimately was a good decision by Broadcom to pursue this approach? I mean, they would know better than me. But, you know, we get to assess things on our end, too. They did the deal. It doesn't mean it's the right deal, but... Yes. I mean, you know, we do make the point in the story that Broadcom was at risk of being left behind. I mean, NVIDIA has done all these different ways of enabling and supporting purchases of its GPUs. This is Broadcom's big sort of entry into that.

29:10And so I think if they were unwilling to do that, it would make it harder for them to sell additional chips. And that means that it might make it harder for them to compete against NVIDIA. So I think, you know, from what we understand, this was a hard decision internally at Broadcom. but they feel they got to the right solution. And, you know, it certainly keeps them in the game for a while longer and allows them to enable more of these chip purchases. Great. Well, Dakin, I want to thank you for coming on. That is Dakin Campbell, our AI finance reporter, here at The Information. As cloud giants like AWS and Google make headway on their own chips, You might think that NVIDIA's market share would be coming down slowly.

30:01My colleague Phoebe Liu wrote this week in our AI infrastructure column that may not actually be the case. I want to bring her on to share with us more about what she learned and about her estimates. Phoebe, welcome to the show. It's great to have you here. Thanks so much for having me. So you crunched some great data that I want to get into. What was the headline of the data? Let's start there. Yeah, so the impetus behind this was that there are lots of kind of critics and skeptics saying that NVIDIA might be losing market share in inference because they've historically dominated training with like, I don't know, 90 plus percent of the market, which is kind of crazy.

30:37So Jensen Huang, NVIDIA's CEO, was kind of defending against this at NVIDIA's last earnings call for the quarter ending at the end of April, basically saying NVIDIA is actually gaining market share inference very, very fast. And basically we were like, okay, he didn't share specific numbers as to how fast. Very vague, vague claim. Very vague, very confident claim. So I was like, okay, can we figure out whether this is actually true or not? I don't think he would say something false on the earnings call, but I was curious to see what kind of napkin math we could do. And we actually found that he was right, I guess depending on what definition of fast you have, but we crunched some numbers from talking to cell-side analysts and customers that run AI workloads.

31:20and basically found that NVIDIA's market share and inference has grown from around 66 % to 74 % over the past year, which is, I guess, pretty marked improvement. And this is significant because as a whole, AI workloads are moving from training to inference. I think our back of the neck and our back of the envelope math found that about 60 % of AI workloads are from inference. So revenue generated from running AI rather than training it. And that's up pretty significantly from, I think, using Deloitte's numbers, 50 % in 2025 and around a third in 2023. And that's, I think, set to grow to as much as 80 % over the next two to three years.

32:06As Jensen said at the Computex conference in Taiwan earlier this month, inference is money. So having market share of inference would kind of determine who owns the next era of the AI boom. So there's a lot to unpack there. So the first thing I want to get into is the definition of the data that you focused on. You were focused on inference data here. So talk to me about what is included in these workloads and then what is excluded, because NVIDIA Chiff can do a lot. Yeah, for sure. And the first thing to note there is that we're kind of looking at a proportion of hardware sales that go to running AI workloads, so inference rather than training.