Meta’s $18B Social Media Settlement, Nvidia’s Upcoming Results, Bill Gates’ Stark Warning About AI

0:13Welcome, everyone, to The Information's TI-TV. My name is Akash Pasricha. It is Wednesday, August 26th. Today on the show, Meta has settled a landmark social media addiction case against them. We'll dive into the details of that settlement. We'll also preview NVIDIA's quarterly results coming out later today with an analyst to get into how inference is changing the story of the chip sector right now. We'll then talk about the pushback against AI data centers with the CEO of Emerald AI, which just announced a new funding round that puts it at unicorn status. We're bringing on our Microsoft reporter to talk about Bill Gates' new warning about AI.

0:52Our editor-in-chief is coming on to share some of her thoughts on AI M &A. And finally, we have some exclusive details about some more OpenAI departures and also about Hugging Faces financials. It's going to be a busy show, so let's get right on into it. Meta has settled a landmark social media addiction case against them. The company said it will pay up to$18 billion and implement new protections for teenagers and stronger controls for parents. Our San Francisco Bureau Chief Jason Dean has been following the case. He joins us now with some takeaways. Jason, welcome back to the show. It's great to have you here.

1:28Good to be here. Thanks, Kosh. Is this the outcome that you expected, Jason? I don't think so. I mean, you can never rule out a settlement in high-stakes litigation, but I think it's fair to say this was a surprise. Okay. And why settle? I mean, what was the alternative here? This would have gone on for many more months, I guess, and there would have been many more headlines about it? I think, yeah. You know, from Meta's point of view, there was quite a bit of risk here. I mean, the amounts that have been talked about, even if they are exaggerated, were enormous. I think Meta at one point put the total amount that the states were seeking at over a trillion dollars.

2:14The state lawyers put it closer to 200 billion. But both of those numbers are far, far in excess of the number that they ended up at. And then you had just the reputational risk of the trial and the fact that Mark Zuckerberg and other senior executives were going to have to testify at this trial that began last week. And that was a trial involving a subset of 29 of the states that are involved in the broader lawsuit that was settled today. So that even wouldn't have been the end of it. so i mean i'm trying to figure out you know what the states uh the people issuing the lawsuits would have wanted here aside from the money you've been following here meta did agree to certain changes on its platform i mean was that the big win here uh or because 18 billion dollars is you know it's a fraction compared to what they seem to have been seeking yeah i think i mean you No, it's a large amount for a settlement like this.

3:16Nothing to sneeze at. But yeah, for a company that had$60 billion in net profit last year, this isn't like an existential threat. I think the remedies are at least as important for those on the plaintiff's side. They include a number of measures, default daily limits, nighttime blocks for teenage users, enhanced age assurance measures to prevent children from getting around those blocks, and new tools for parents to help protect their children. So, you know, clearly the state attorneys general involved here felt that those and other remedies involved beyond the money would help to address the concerns, the numerous concerns they had about how Meta's platforms have been affecting kids.

4:09And do you think that these measures will work? It's too early to say. I mean, Meta has announced countless at this point tools and changes that it says will help improve the welfare of its minor users. I'm not sure anybody who is a critic of the company is particularly persuaded that those have worked so far. We'll have to see how these are implemented, how they're enforced. I think it's safe to say that probably this is not the end of the controversy. Right, right. So there was an interesting part here of this morning's settlement announcement. So Meta is calling on YouTube and TikTok to sort of join in this, you know, push for more safe use of social media.

5:10So just remind us of the context here. So these cases that were filed, it wasn't just against Meta. It was against all these platforms. And so we only have a settlement from Meta so far. Is that was that the context here? Yeah, Meta is, I think this is something else that they kind of got out of this settlement. They're committed to pay about 70 % of that total over a decade in installments to the various states and territories that were party to the settlement. The remaining 30 % is contingent on the states getting TikTok and YouTube to agree to, A, make many of the same changes to their platforms that Meta has agreed to, and B, to pay an equivalent amount to that 30 % for each of those two companies.

5:58So, you know, they're trying to do some jujitsu here, I think, where they have been by far the biggest target of this sort of criticism. And they are now positioning themselves as the company that is calling on the industry to step up and make things better. So there's a benefit to them in terms of making a big portion of the settlement contingent on the actions of other companies and positioning them to sort of play a leadership role in this area. And do you think TikTok and YouTube and – I mean they didn't mention Snap, I guess, and if Snap was the other company, do you think they all settle as well?

6:43Do they have track records of settling these types of disputes? I think we'll have to see. I mean, they could face some of the same pressures that Meta was under, but neither of them has had the degree of scrutiny and pressure, the sheer number of lawsuits that Meta has been facing. I mean, let's just talk about this broadly, Jason. Like, if you think about Meta and its product focus right now with AI, that's certainly what investors are evaluating the company on right now. Its CapEx is extraordinary. They want to charge for these consumer products now at some point down the line. I mean, help us connect the dots here.

7:30How do you see these two as related? They're sort of on this paradigm shift, I guess, from being a social media company to an AI company. Does this settlement have any impact, do you think? or were they thinking about the AI push as they decided to settle this? I mean, I think this, if the companies are wise, they will see this as a precedent for what could happen with AI. Many of the same issues of platform responsibility and potential harm to not just children but adults as well exist with these new AI, you know, with AI chatbots and other AI products. not just exist, but in some cases, arguably, are much more acute.

8:15And, you know, Bill Gates is saying as much this morning. You know, he's the loudest this morning about it. Talking about, you know, the stunting of sort of emotional development of children from overuse of AI chatbots. So much hangs in the balance, and the companies have a choice of either trying to get ahead of that and do the right thing early, potentially at the risk to growth, which is obviously the thing that Silicon Valley has been obsessed with for a long time, or not, and kind of compete in a lowest common denominator way, as arguably has been done in social media. And then we end up with a similar set of lawsuits and regulatory actions, probably in the not distant future there.

9:02We already have some, Right, right. Let's go back to what you started this conversation with. You said that this likely will not be the end of the controversy, even though it's a settlement. So are there still lawsuits hanging in the balance? Or do some states still have them outstanding? Where will this story go as it relates to the social media addiction issue? I mean, there are many more lawsuits from individuals, groups against Meta and other companies around this issue. This settlement covers 52 states and territories, including the District of Columbia. There was already a legal battle that Meta lost with the state of New Mexico.

9:49So, you know, the state attorneys general and territorial attorneys general are largely covered by this, but there are many more lawsuits still to be fought. And I think that was also part of the incentive for Meta to, you know, contain the risk with the settlement in this one. Right. Great. Well, Jason, I want to thank you for coming on. That is Jason Dean, our San Francisco Bureau Chief here at The Information. information. NVIDIA reports its quarterly results tonight, and there is a lot to look out for, but as big an event as it is, the chip story has so much else to it right now with the memory crunch, inference chips, OpenAI developing its own chip.

10:29I want to bring on our partner UBS for a conversation about all this. Tim Arcuri is head of AI Semiconductor Research at UBS. Tim, welcome to the show. It's great to have you here. Thank you. Okay, so let's start with With the news at hand here, NVIDIA has got results tonight. What are you going to be looking for? Well, I expect a very strong quarter. Backlog and demand signals are very strong. This year, our model implies they're going to ship 10 to 11 gigawatts. You saw from the deal that they have with SpaceX, they have basically 100 % share there. SpaceX has said that they're going to add six gigawatts next year.

11:06That's a huge incremental number on what they're doing this year. So obviously, that speaks to how strong demand is. I think the multiple on the stock has gone down so much. I don't think anything they can say really is going to change people's mind about the AI narrative. To me, it's more about the numbers than about the narrative. I think people kind of have their mind made up about the financing and all the questions around that. I'm not sure anything he can say about that is going to really change people's mind. It's more to just prove to people how strong his demand is and how big his revenue can be over the next few years.

11:42Right. Well, I mean, let's talk about what he could say on the call. You know, there's a lot to get to. There's certainly the open weight model push the company's making. There's the whole, you know, expert control issue. It'll be interesting to see if he addresses that, if at all. I mean, what I'm sort of looking for is a little more color on the inference play that they've made with Grok. And we had Nebius on yesterday to talk about their adoption of the Grok chip that NVIDIA has been working on. I wondered, as you think about inference broadly, have we seen inference show up in the financial results for these chip companies in a meaningful way?

12:24Is it all just commentary right now? I mean, you know, we certainly have the companies that are specifically focused on only inference chips, but how are you seeing it show up in the financial results? I mean, look, training is just the R &D phase. And we've trained all these models. And now all these models are now really, I mean, we're seeing massive increases in inference tokens and inference equals revenue. And this is the real scaling of AI. So inference is what is driving revenue for all of these companies. It is what is driving revenue for the hyperscalers. It's what's driving revenue for the neoclouds.

13:00It's what's driving revenue for NVIDIA. It's what's driving revenue for the memory companies. So inference is the story moving forward because that is the production of AI. It's like, I always use the analogy, you go into the kitchen when you were a kid and you learn pretty quickly not to touch a hot stove. Well, that was the training that basically made that rule in your brain. And then every time you walk into a room that has a hot stove, you create an inference token knowing not to touch that hot stove. So that's the value that gets created to you from the rule that you train. So we have shifted now from creating these models to now inferencing tokens and generating revenue.

13:44And so that's become the big driver for all of these companies. So, I mean, whereas some people might go into the results looking for specific numbers tied to sales of these inference-focused chips, you're sort of making the argument that it doesn't really matter if they tell us how much of the Grok chip they sold or even maybe what Cerebrus' results tell us about inference sales altogether. You're saying this is driving demand across their entire product stack right now. Yeah, I mean, what you're talking about is fast inference. So there are different parts of the market. So inference is what is driving everything.

14:26Fast inference is where the debate is. It's what is the threshold for people to pay more for tokens? And we're still sort of sussing all that out. It's like, what are the markets that people will pay more for tokens? Coding is definitely one of those markets. And so that's the market where, you know, OpenAI went out and signed this deal with cerebras uh you know on you know hotly on the on the you know tales of that nvidia bought grok uh i think that jensen very quickly saw that there was a pretty big market and that deal you know convinced him that there is a big market for fast inference now he's said different things at different times whether it's 10 or you know you know uh you know 20 uh you know percent of the market uh it's still a big piece of the market and it's and and so uh i think that's what you're talking about is the fast inference market.

15:15So yes, I would like to hear something from him on how that market breaks down and how big the fast inference market truly is. Yeah. We saw this week that OpenAI was making some noise about its inference focus chip, Jalapeno. Do you think that that is a threat at all to any of these inference focus chips or fast inference uh i mean sure they're you know every single um uh hyperscaler and also the large language model providers want to make their own chips and in many cases uh this is going to be to serve the fast inference market so i think i think that the stratification of the inference market is still being worked out uh and i and i don't think that anyone really knows how how how the market will bear you know paying what price for what inference token so i think that there's just a rush to have a solution that can serve every conceivable part of the infants market out there right i want to ask you about the m a landscape right now so grok certainly came in as you said to nvidia by way of a uh licensing deal or sort of pseudo acquisition you could say uh amd has made some big acquisitions i mean they're coming fast and furious um is there a Are there particular pockets of M &A you think are going to be very active here from the big public semiconductor companies?

16:38What are they going to try to buy here? Yeah, I mean, for the big companies like NVIDIA and AMD, it's been mostly on the software modeling and the networking side. There's a lot of activities in areas of technologies to string GPUs together and racks together. stuff like advanced packaging, new chip architectures, and new networking and connectivity technologies. I guess if I had to generalize though, the M &A deals in semis, the bigger ones anyway, have really just been to build scale. It's really no different than what we typically see as markets evolve. I mean, if you just look at the size of the chip business, we're going to be a$1.7 trillion business this year, scaling to about the mid twos next year.

17:22So now a lot of that's, you know, memory, but even if you exclude memory, we're growing at 30 % this year, if not more. And we're, you know, pushing a, you know,

17:36trillion dollar business soon. And so I think there's really a lot of opportunity for and, you know, a lot of desire for these companies to build scale as the market gotten to be that large. Right. And sorry, you said$1.7 billion,$2 billion. That was on which metric was that based on? Semiconductor revenue, the entire size of the semiconductor industry, semiconductor revenue will be$1.7 trillion. Trillion. Okay, okay. Right, right. $1.7 trillion. $2.5 trillion next year. So I'm just pointing out the size and the scale of the revenue opportunity. And if you are a company in this industry, what you want to do is you want to build scale to capture that massive revenue opportunity.

18:24Right. So on that note, they're on acquisitions then looking at scale. I mean, is there any way you think that they might try to address the acute issues of, for example, the memory chip supply chain crunch? I mean, is that something they could sort of address with acquisitions? Might they try to address some of these other bottlenecks down the line with acquisitions? I think it's pretty hard. I mean, it's pretty hard to go out and buy a memory company. I don't think acquisition is the solution. I think what you're seeing is you're being forced to make trade-offs in your product roadmap to try to arrest some of these bottlenecks.

19:09And you're making trade-offs in terms of cost and performance to try to address the fact that memory is in such tight supply. And one thing you're seeing actually is rather than M &A, what they're doing is they're going out and they're signing these long-term agreements with, for example, memory companies to basically commit to a certain volume over time. And they're committing to pricing as well, because everybody wants to have some control on their input cost on the memory side because it has become such a big part of their input costs. Right. Well, Tim, there's a lot to watch for and certainly a lot to watch for tonight.

19:50I want to thank you for coming on. That is Tim Arcuri, head of AI Semiconductor Research at UBS here on TI TV. Emerald AI raised$150 million at a valuation of$1.05 billion. The company is developing software that helps data center companies optimize its power use it is backed by a number of big names including nvidia emerson collective salesforce ventures and john dore of klander perkins i want to bring on founder and ceo varan shivaram for a conversation varam welcome back to the show it's great to have you here okay thanks for having me congrats on the funding round and notching the unicorn status so last we had you on we were talking about your customers you mentioned that Oracle was a big name that you were working with.

20:38That was quite a few months ago. Tell us about some more of the customers that you signed on for your product. Well, look, we're just delighted to be on. Thanks for having me on. We're delighted that Emerald AI transforms AI data centers into flexible allies to the power grid, capable of reducing their power use when the grid is stressed while protecting the performance of these critical AI workloads. We've now demonstrated the technology in five different data centers around the world, And we've entered our commercial scaling phase. We're deployed at full data center, multi-megawatt scale. And with NVIDIA, we have the world's first power-flexible AI factory or AI data center coming online this year with Digital Realty in Virginia and Manassas later this year.

21:22So we're delighted that we count among our customers some of the world's leading AI firms, global data center platforms, as well as leading American power utilities and utilities around the world globally. Right. So, I mean, I see some of these logos on your website. So Nebius, who is a frequent guest on this show, is that another company that you're working with, with their data centers? We did a fantastic showcase of the technology with the Nebius AI Factory or data center in London. Here's what we did. The National Grid, one of our investors in the utility in the added kingdom, needed the data center to reduce its power consumption during the moment of peak grid stress.

22:02just so happened, Akash, this moment of peak stress was at halftime of a soccer game. Everybody in England watches soccer. They turn on their tea kettles. And when the tea kettles spike, launching a gigawatt of load onto the grid, that AI data center was able to flexibly reduce its demand. They launch their tea kettles during halftime? That's, I didn't know that. And at full time, they stop watching their TVs. They go into the kitchen, they turn on the tea kettles. So as it so turns out, every jurisdiction is going to be different. Hot Phoenix day, all the air conditioners turn on. And that's when we made a Phoenix data center, an Oracle data center, reduce its power consumption.

22:36California, when all the solar goes down in the evening, there's a need for demand reduction on the grid to help us balance supply and demand. The point is AI data centers can now be these flexible allies to the grid. And Emerald's built the technology with NVIDIA to make that possible. So have you done any data collection then? I mean, you've done some work in Arizona, Illinois, Virginia, a number of states here in the US. Have you been able to collect any data on how much your software has saved maybe the local counties in terms of electricity costs or affordability? because this is the heart of the issue, right?

23:18Well, it's a multi-layered issue, I should say. It's one facet of the issue here, the pushback against data centers. Have you collected any data that suggests that it's bringing down costs for those counties? Exactly, Akash. This is the heart of the issue. And so, you know, I was proud that in the New York Times yesterday, we were profiled as this AI startup wants to reverse the backlash against AI data centers. There are two key problems that we solve. The first one is we make it possible to solve America's power crunch for AI, right? You can't fit enough data centers on our power grid today unless you have Emerald AI.

23:53With flexible data centers, we can unlock 100 gigawatts of power overnight, and we can fit more data centers onto the existing grid because they're flexible. And the second problem we solve is with Emerald AI, as you said, Akash, you're actually able to reduce power bills or restrain the rise of power bills. because a flexible data - How much? How much? By how much? Every time you increase the utilization of a local power system by 10%, you reduce bills by three to 4%. The point here is we have a big grid already in America. There's 100 gigawatts that isn't being used today. If you can better utilize the existing grid, an AI data center that joins that grid will reduce community bills.

24:34That's how you help to solve the affordability crisis and the reliability crisis because data centers can be unpredictable users. By making them controllable, flexible users, you actually prevent blackouts. So communities should look at these AI data centers that are flexible with Emerald as potential good grid citizens, good citizens for the community. And they, of course, benefit because they get faster access to power. So you are talking day and night to these data center developers. As you look at the pushback against the data centers, which, as you said, is at the heart of the issue right now, What kind of regulation do you think should be implemented here in terms of data center development?

25:16We need to recast the contract that AI and data centers have with power grids and with local communities. And that new contract is what Emerald AI and our software platform, the Emerald Conductor platform, operationalizes by making these data centers flexible. That new contract says, look, if you are a utility, you're a regulator, you're a governor, you should be inviting these flexible, good citizen data centers into your community. You should have lots of them. They drive economic development. They drive jobs, innovation. But they also, because they're flexible, they better use your existing system instead of forcing you to expensively upgrade it and raise everybody's costs.

25:54And they keep you safe, your energy secure and reliable. That's the compact. If a data center is willing to bend over backwards and be flexible, and our software ensures that the critical AI workloads continue to run at peak performance, right? We are orchestrating AI workloads with AI agents within an AI data center in order to precisely control the power while keeping the AI performance acceptable for the customer. That's a compact that I think we can get behind because today's compact doesn't work. Governors need a new option, a third option where they say, hey, come to my state, come to my utility, but make sure you're helping my community out.

26:30Last question for you, Varun. You've done a lot of work in the U.S., also in London. You said I'm sure you're focused globally. Is there a country that is doing this really well right now? I'm thinking about all the energy advancements happening in China. are they sort of the leader here in terms of keeping electricity bills in check as they build out data centers? Or where are you looking for your own innovation advice, I guess? You know, the funny thing is China doesn't need this, right? China built 500 gigawatts or more power generation last year, 10 times what America did. And between now and 2030, we'll have a 50 to the 80 gigawatt gap in what we'd love to build for data centers and what the grid can actually plug in.

27:20That's why it's so important that overnight, we unlock this 100 gigawatts that data center flexibility does. Now, you mentioned the global opportunity. Look, we're not just confined to the US. America and all of our allies need this in the race for AI innovation. That's why we're proud that in this Series A round, we've raised$150 million, and we now count 12 Fortune 500 global companies on our cap table, not just GE, Vinova, and NVIDIA, and Salesforce, but also Siemens in Europe, and Samsung in Korea, and several Japanese entities, Turkish entity, et cetera, et cetera. We are going to take this technology, bring it global anywhere that you have a grid that can't plug in enough data centers quickly enough, and anywhere where rates are rising for customers, we're going to help to make AI a good grid citizen, keep rates affordable, and get way more data centers connected.

28:09You can have them all at the same time with flexible AI data centers. Great. Well, Varun, I want to thank you for coming on. That is Varun Shiveram, founder and CEO of Emerald AI here on TITV. Thanks for having me. Bill Gates issued a warning today of the threat of AI. In a blog post, he wrote that AI will either be the greatest equalizer ever invented or the worst source of injustice. I want to bring on the information's Microsoft reporter, Aaron Holmes, for more on this. Aaron, welcome back to the show. It's great to have you here. Happy to be here, Akash. Okay, so this is the co-founder of Microsoft saying that AI is going to cause some real pain.

28:51Did this surprise you coming from Bill Gates? You know, it is somewhat interesting because this is kind of a pivot from where Bill Gates was on AI just a few years ago. We actually know in 2019, you know, he was advising Microsoft executives that he didn't see the hype in OpenAI and he wasn't sure if it was worth investing in them. Then by 2023, he had come around and he was extremely optimistic about AI and wrote an essay that year saying that AI was going to potentially lead to abundance and solve world hunger and solve things like cancer. Today's essay or last night's essay is a lot more dire.