Inside Anthropic, OpenAI and Google’s Secret AI Standards Talks, Anthropic’s $13.7B Compute Deal

0:13Welcome, everyone, to The Information's TI TV. My name is Akash Pasricha. It is Monday, September 14th. We have a ton of coverage for you today on Dario Amadei's explosive essay over the weekend. And many of the big AI labs have all chimed in and markets are moving. Before we get to it, I want to remind you that The Information is launching a new weekly show, AI Deep Dive Today, hosted by my colleague Rocket Drew. The first episode is going to be live at noon Pacific, 3 p.m. Eastern. It is Rocket's conversation with OpenAI researcher Noam Brown. You do not want to miss that. Today on this show, we're going to be talking about our exclusive reporting on what is going on behind the scenes of all of these pacing the frontier AI research essays that have been swirling around.

1:03We also have reporting on how worried Anthropic and OpenAI customers are about the labs using their IP. We have a scoop for you on a multi-billion dollar compute deal that Anthropic has signed. And to wrap the show, I'm bringing on the CEOs of Superhuman and Rillit for their thoughts on this crazy news cycle that we're in. It's going to be a great show, so let's get right on into it. It was a busy weekend in Silicon Valley, as yet another essay from Dario Amodei drew attention from all the big players in AI. The Anthropic CEO called for the AI industry to moderate the speed at which it develops cutting-edge tools.

1:41Sam Altman, Elon Musk, and Demis Asabas all largely agreed with him. Of course, it is important in these moments to talk not just about what leaders are saying, but also what they are doing behind the scenes. And that is why new exclusive reporting from our tech and politics reporter, Leo Schwartz, is so important right now in this moment. Leo caught wind of some of the conversations that are happening behind closed doors. And I want to bring on Leo to share more about what he found. Leo, what a weekend. What a weekend indeed. Not quiet. So I want to get to Dario's essay in a minute here. We'll unpack it all and get your thoughts on it.

2:19But what are you hearing is going on behind closed doors right now in these in these boardrooms? I think you have to wind the clock back a bit to July. That's when Demis Asabas, then CEO of Google DeepMind, published his own proposal, basically calling for a self-regulatory organization, a standards body among AI companies that would be lightly overseen by the government, similar to FINRA, the financial SRO. So what we reported is that after that proposal came out, there were actually behind the scenes talks, a working group below the CEO level between the three big labs. That would be DeepMind, Anthropic and OpenAI that continued for weeks.

3:01If you remember, a few weeks ago, I had reporting that there had actually been an executive order draft for this type of proposal that would create an SRO that stalled in the White House. And then, of course, you had yesterday's essay from Dario, which did mention the standards body, among other proposals. So our reporting showed that these discussions between the three large AI labs had been continuing with the discussion as recent as this past week about the idea of a standards body. And that bubbled up into Dario's essay. And so who exactly is in these meetings? Do we have any idea? We don't have specifics, but it's a working group.

3:41So it would be people working on policy at the three companies, not the CEO. So it wasn't Dario, Sam and Demis, but leaders at the three companies working on these ideas. And what do we know about what has sort of materialized in these discussions? I mean, how far have they gotten? Have there been drafts of, you know, shaft rules that that, you know, could have bubbled up into Dario's essays? Yeah, well, so clearly there there was progress enough that they had convinced the White House enough to have an executive order draft, at least. Yeah. Stalled, as we reported earlier. But the talks continued.

4:17And the idea, as I reported, is that you could actually do this, not necessarily with the government, but have your own SRO without the government. This sort of famously was raised by David Sachs, the former AI czar, in a tweet a few weeks ago where he said he didn't want to have a FINRA-type model. He wanted to have something more modeled after the Motion Picture Association, which was created entirely without the government. So even after this idea stalled in the White House and there wasn't government support, the three labs continued discussions about a standards body. I also reported that in a all-hands meeting last week, opening by CEO Sam Altman mentioned the fact that they wanted to continue forward with the plan, potentially without the government.

5:00And then in Dario's essay this weekend, he talked about the idea of working together with industry to create standards. He did say that it would likely need or be helpful with some sort of government mediation, but he didn't actually offer more specifics beyond that. Right, right. So I want to get to sort of all the different angles of the reaction to this essay that Dario put out, because, I mean, David Sachs sort of made this point, which I guess is consistent with what he's made previously here. He basically said, you know, you guys are out here calling for pacing the frontier. You are the ones with the power to pace the frontier.

5:42Effectively, what I understood him to be saying is quit talking about it and start doing it. And so it sounds like, I mean, that's still what he seems to be advocating for is, you know, leave the government aside right now. Stop talking about regulation. Just start doing what you say you're going to do. Is that sort of how you understood his reaction? Yeah, although I think a lot of these different ideas, proposals, and policies have gotten conflated somewhat. Dario's fault by conflating everything under this umbrella term pacing the frontier, which our colleague Corey, I think, memorably called an all-timer for corporate jargon.

6:16So there's different ideas here. What is this idea of having a standards body? So this would be some sort of independent organization where all of the AI labs, anybody developing models would basically be a part of. there would be independent evaluators doing audits and testing and creating a set of standards. That seems like the type of thing that companies could do on their own, wouldn't necessarily need the government. There's another idea here, which is actually slowing the pace of development and companies basically agreeing together that speed of development and training should only happen at a certain pace.

6:48That is where Dario at least said they would likely need some sort of waiver, like an antitrust waiver from the government to make sure that they're not later sued for antitrust collusion. That was where David Sack came in and said, you don't need that antitrust waiver. You can do this on your own. And I think that this is a sort of legal distinction that is still being worked out. Sam Altman, I think it was this morning, tweeted that they didn't need any sort of waiver or government support to do this. But again, there's a lot of different policies and ideas that are being conflated into one general argument right now.

7:21Right. Broadly speaking, what has been the reaction in D.C. or the reaction from lawmakers from this whole onslaught of messaging from these companies? I mean, it's sort of too bad that we live in this world where leaders are forced to give their real-time thoughts on things because then, like you said, things get conflated, people get mixed up. Do we have a sense for how D.C. is reacting to this latest weekend of thoughts from Silicon Valley? I mean, I would say we've seen the full spectrum of possible responses you could possibly imagine. Starting with Trump this morning, he said, we don't need any regulation at all.

8:04Yeah, I love he called Dario. He said Dario is pretending to be a perfect little angel in quotes, I think was the exact language. And then you have the other side, Bernie Sanders, the anthropic researcher who quit last week, and Steve Bannon apparently are doing an event tomorrow, which is more calling for a moratorium on any development. Together. Yeah. They're doing an event together. Together. So there's the whole spectrum. There's something that only AI can, I guess, do. This is the true beauty of AI. Right, right. So, but I mean, broadly speaking, so I'm sort of trying to look ahead to the election too.

8:44And, you know, some people have suggested online that some of this might have to do with that timing as well. I mean, what do you think might be the political agendas here of the big AI labs in this moment of vague regulation of an election coming up? And is this strategic, do you think? I think it's strategic in that it's certainly setting the table for legislation post-election. I think Dario, Sam, pretty much everybody is in agreement that at this point, mid-September, midterms in two months, a lot of disagreement within Congress and the administration how to act. It's very unlikely we're going to see any sort of legislation, let alone the type of comprehensive legislation that would actually address this.

9:29But when you look forward to the lame duck session, And then especially to the new Congress that will start in January, that's when the real stakes come. And I think they're really, you know, trying to set up Congress and the administration, the industry itself to figure out some sort of solution when January comes. Right. Great. Well, Leo, I want to thank you for coming on. I know it was a busy weekend for you, so we'll let you get back to it. That is Leo Schwartz, our tech and politics reporter here at The Information. The NASDAQ was down this morning, but surprisingly, software stocks like Salesforce and ServiceNow and Workday were all up this morning.

10:08For more on the impact of what this weekend's pacing the frontier conversation could have on big enterprise software companies, I want to bring on Aaron Holmes, our Microsoft reporter and Laura Bratton, author of our Applied AI newsletter. Welcome to you both. It's great to have you here. Aaron, I want to start with you. I mean, what do you make of this? We're going to slow the frontier and suddenly the SaaSpocalypse is over? What's going on? Yeah, I mean, like you said, we're seeing AI stocks down. And for some reason, software stocks are up. I can't really speak to why the market is reacting that way other than, you know, maybe some people think that a slowdown in AI progress could also mean a slowdown in these AI companies branching into other markets.

10:53although not sure that's going to happen. You know, Dario Amadei was talking about trying to stop the risks of potentially AI wiping out mankind, but we are still seeing Anthropic start to move into new markets like financial services, as some of my colleagues reported last week. So I'm not really sure if I personally would read anything into SaaS stocks stemming from this slowdown. Laura, what are your thoughts here? because, I mean, every enterprise software company in the world has touted the power of AI as, you know, I think as Bill McDermott calls it, you know, getting rid of the soul-crushing work that, you know, knowledge workers worldwide have got.

11:35I mean, this was the whole pitch, was that AI was going to save us productivity. And now, I mean, the only way I could think about this maybe is that maybe the models were good enough to do that. And so maybe the enterprise software work is fine and it's the really, you know, math Olympiad solving capabilities need to be tempered. I mean, is that one way of thinking about it? I mean, like Erin said, I'm not going to bend over backwards to try to understand an irrational market, but I guess I think that this is so simple where, you know, investors are, who are maybe like not reading so much into the fine print of Dario's essay or thinking, oh, if the pace of AI model development slows down, then that will limit enterprise software customers' ability to vibe code their own apps and maybe slow the so-called death of these software companies, which I think we can all agree that they're not necessarily going anywhere anyway.

12:38But yeah, I... Head scratcher, in other words. It's a head scratcher. I can see it in your face. You're like, I don't know why people are reacting this way. I do want to ask you, Aaron. So Satya Nadella did have a reaction over the weekend. What did he say over this discussion? Yeah, Satya called for pacing of AI, sort of joining the chorus of the other tech leaders that we've seen join that call. And I also interviewed Mustafa Suleiman, the CEO of Microsoft's AI unit yesterday. And, you know, he similarly has put out this sort of statement saying that Microsoft's development of AI needs to make sure that it won't harm humanity and that Microsoft is also open to potentially having, you know, third party auditors come in and look at their own development of AI to make sure it's adhering to those safety standards, which is something that Dario proposed on Saturday.

13:36So we're seeing Microsoft, even though it's not quite at the same frontier of AI development does OpenAI and Anthropic essentially throw its weight behind this broader push to coordinate some sort of industry-wide pacing of AI. Laura, I want to pivot for a moment here to a story that you and Aaron published. It was today. The days are all blending together. I think it published this morning. So your story looked at how the customers of the big AI labs have concerns effectively over how the AI labs are using their data. This is a fear that we've been talking about a little bit in the show, but you went deep on your reporting.

14:16What did you guys find? Yeah, so Aaron and I talked to customers of AI model providers, and something that I focused on, I was particularly looking at Anthropics data retention policies, and I found that big Anthropics customers, you know, including participants in project glass swimming, were restricting how they use Anthropics commercially available version of Mythos, Fable, the Fable models, they're restricting how they use the Fable models because they're worried about Anthropic's ability to see their data, about how they're using that model, because Anthropic doesn't have zero data retention in place for the Fable models.

14:58And so this was the zero data retention. How does that relate to the June policy change at Anthropic that kind of sparked all this? Yeah. So that's a great question. In June, Anthropic said that it was going to retain data about how enterprises and other users were using its Fable models. Previously, Anthropic allowed enterprise customers to negotiate zero data retention contracts for the other models that it was using. But what I heard from customers is that even if they had these contracts in place with Anthropic saying that it can't retain data about how they're using other Anthropic models, that didn't apply to Fable.

15:45So it raised all these other concerns. Now what we've seen is Anthropic has introduced this new program called Enterprise Frontier safeguards that allows eligible customers to use Fable without that data retention. And they can, you know, keep all the data about how they're using Fable on their own servers. So structurally, you know, it's not going to Anthropic, but that's not going to be rolling out until this fall. So, you know, that's impacted how massive customers like Palantir, NVIDIA, government contractor, Booz Allen, a large utility that I spoke to. It's preventing them from using Fable in a more widespread way across their organization.

16:26But so when that change comes into effect later on this fall, what is your understanding from these customers? Do they think that that goes far enough to sort of address their concerns? So for some, some companies thought that it went far enough, but some companies that I talked to were worried about the fact that within this program, there's a provision that says Anthropic can unilaterally revoke zero data retention or basically pull back and change the rules whenever it wants. And that's the issue with having a program like this rather than a contract like companies previously had with Anthropic that said, hey, we want zero data retention for all of Anthropic's models going forward.

17:14And my understanding is Anthropic is not doing that. They're not giving the zero data retention as a blanket. They're sort of doing a little more case-by-case sounds like. So it's always been case-by-case, but with this new program, even that eligible customers that do effectively get zero data retention, there's like a line that says we can change the rules. We can pull back whenever we want. We can take away zero data retention. Got it. So Aaron, what about OpenAI? How is OpenAI dealing with this and how are their customers reacting? So OpenAI's data practices have been in the spotlight since last week when these researchers working on the Navier-Stokes problem basically accused OpenAI of spying on their use of codecs and other OpenAI tools.

18:09Since then, OpenAI has come out and said pretty definitively that they did not use any data from those researchers for the past two years and that, you know, their models solved Navier Stokes independently. But more broadly, we are seeing, you know, competitors like Microsoft and also other competitors like NVIDIA starting to try to peel away customers of both OpenAI and Anthropic, essentially by, you know, pointing at this idea that when you're feeding data into a large language model, even if the company say that they are not using that data for anything, you know, maybe you never know. And we do know that some large customers have had this very cautious approach where they won't put any, you know, proprietary data into a large language model, period, unless that model is running locally on their own hardware, which is, you know, very much in response to the same sort of concern.

19:00So, Laura, I mean, how much of a threat is this for the businesses of these big AI labs? I mean, if we sort of assume that these concerns may become more widespread, maybe the concern sort of gets wider in radius, is this a threat for the labs? Or do they have an easy out here? And is there a way to address the concerns very simply? I think that the programs they're rolling out to let customers, you know, run these, retain data about how they're using these models on their own servers is a step in the right direction. And if they can roll that out to more customers, that would be great. But when you think about it, at the end of the day, these model providers depend on improving their frontier models at a rapid pace.

19:49And improving those models depends on getting crucial data. and so you could argue that their business incentives are sometimes at odds with their enterprise customers. I think that this bolsters the argument for large enterprises using open source models, which I do think is a very real threat to the growth of the AI labs. However, you can argue that they're growing so rapidly and AI is still early in adoption that this isn't going to significantly hinder them in the near term. Right, right. Great. Well, I want to thank you both for coming on. That is Laura Bratton, author of our Apply Day newsletter, and Aaron Holmes, our Microsoft reporter, here at The Information.

20:36We are continuing our Anthropic coverage with some exclusive reporting that the company has signed a$13.7 billion six-year compute deal with Rum Group, a tech company with ties to the Trump administration. Our colleagues Phoebe Liu and Valida Powell reported this story over the weekend. I want to bring on Phoebe now to discuss more of her reporting. Phoebe, welcome back to the show. It's great to have you here. Great to be here. Tell me about this deal that has come together. Yeah, so the deal is that Anthropic, as you said, signed a$13.7 billion agreement to rent compute capacity from Run Group, which is a pretty new, young cloud as far as things are considered.

21:22You might even call it a Neo cloud. Yes. Yeah, it started out as kind of a YouTube alternative for people who wanted kind of lighter content moderation and acquired a Neo cloud company, Northern Data, in June of this year. And a lot of that capacity has turned into opportunities for big deals for companies such as Anthropic, which is very interesting because it goes to show kind of how desperate at Anthropic is to sign additional compute deals with kind of anyone and everyone, which maybe is an interesting parallel to the essay that its CEO, Dario Amadei, published over the weekend saying that we should kind of pace the progress, the speed at which AI is progressing and perhaps even looking at limiting the amount of compute that is used to train and evaluate and use AI.

22:16This seems like it's kind of counter to that narrative, although maybe it's still early on and how things might change. Right, right. And, well, I mean, who knows, right? This deal could have been signed a couple days prior, right? The blog post comes out. I mean, maybe this deal— Yeah, I think it was before the essay. Yeah, yeah, yeah, yeah. And hearing just from anecdotal conversations that they're doing a lot of these, like, slightly smaller deals, My colleague Valida reported last week that Anthropic has more than$500 billion in total compute deals, mostly over the next few years, which is a staggering figure.

22:55Right. But I mean, it certainly does speak to the tension that you're pointing to, which is that while Anthropic and we just had Leo on before you, I mean, while Anthropic and OpenAI and Google are all having these discussions about pacing the frontier or moderating AI development behind the scenes, they still are all trying to get their hands on all the compute that they can because the competitive threat is very much here and now. tell me a little bit about the the rum group i mean this is not a neoclatter company we've talked about at all on this show uh what are the origins of this company and and why how did they grow to be a company that anthropic looks to yeah totally so i teased that this a little bit a couple minutes ago but it was founded in 2013 as kind of a youtube alternative with kind of a free speech focus, which meant that it gained early traction among conservatives and its early investors include Peter Thiel and former Vice President J.D.

23:59Vance. I would say that Rumble has aligned itself with the Trump administration pretty strongly. Over the past few years, it hosts the Truth Social platform on its cloud services when the Rumble video platform is not being used. But that business model hasn't exactly allowed Rumble to take off. Financially, it's a public company. So in June, it kind of pivoted to being a little bit more of a NeoCloud. The Rumble video business still exists. It's kind of two separate tracks. But it bought this German NeoCloud company called Northern Data for about 800 million. once that merger closed the company renamed itself the rum group and with that acquisition it now has a couple hundred megawatts of very importantly grid power which is very very hard to come by these days like so hard that elon musk is trying to build his own turbine blade factory as my colleague ann reported on a few weeks ago which is crazy and because it has that grid power I think people are kind of clamoring to get in on what capacity they can rent from that because so many of the other new builds these days are still awaiting power that can be turned on and connected immediately.

25:26Right. And we should say, I mean, it is a public company, as you pointed out, and shares did jump pretty substantially following the report here, which very much reminds me of any time any of the big AI labs issue a press release, you know, you can count on a stock pop. And so your report is certainly something to that effect. I want to ask you, Phoebe, so whatever the ties are with the Trump administration, do you think that this is at all a political decision for Anthropic? Or is this really just Anthropic trying to get compute from whoever can supply it? I think it's very much the latter. And I think the Rum Group would also even say that they're not politically aligned one way or another.

26:14It just happens to be that their user base tends to swing one way. Right. Yeah, I think it's definitely... And that too, only for one part of its business, basically. Right, exactly. Yeah, I think it's definitely Anthropic being very desperate to get any compute that it can. It wants to have as much as its competitors, even though they may be talking about pacing of the AI frontier. That doesn't mean it's necessarily going to slow down new compute deals. So that's going to be one key indicator to look at going forward to see kind of whether they're putting their money where their mouth is. Right.

26:50Great. Well, Phoebe, it was a great scoop. I want to thank you for coming on and sharing with us. That is Phoebe Liu, our NVIDIA reporter here at The Information. OK, shifting gears slightly. Superhuman, the company formerly known as Grammarly, is making its biggest acquisition yet. The company is buying Fathom, an AI note-taking company. Superhuman has been making a number of acquisitions lately. I want to bring on CEO Shishir Mehrotra for a conversation about his strategy. Shishir, welcome back to the show. It's great to have you here. Thanks for having me. So walk me through the acquisition here.

27:25What is Fathom? Why did you decide to buy them? Yeah, so Fathom is one of the market's leading note-taking products. And it's a category that I spent a lot of time in. And it's one that our users ask for. Basically, usually all our products are asking for this. And I spend a lot of time looking at the category, looked at all the different companies, really fell in love with the Fathom team and with the Fathom product and came to a real alignment on what we want to do together. And so we agreed to acquire the company and announce it this morning. How much did you buy them for? You know, one of the pleasures of being private, private transactions is that we don't have to talk about that.

28:03You said that last time too. when I asked you about GPT-Zero, you said, hey, you know, I don't need to tell you anything. I think if you're going to be a private company, you take full advantage of it. I think both sides were quite happy with the deal epidemic. So I think, more importantly, we're going to build something really great together. And so the vision here, I mean, help me understand here, the vision here, AI note-taking, I mean, how do you make it competitive with all the other AI note-taking apps that are installed in just about every enterprise software platform now? I mean, even Zoom calls now.

28:39I mean, it gives you the notes right away. How do you plan to make it competitive? You know, I think there's a couple different things we can uniquely do together. I think that if you think about the way Superdum in all of our products, including Grammarly, including Mail, including Docs, and now G50, and so on, our goal is to give everyone an AI-native productivity suite that works where you work, how you work. And we think we can bring that to note-taking as well. If you think about note-takers today, many of them require a real change in your behavior, whether that's pre, during, or post-meeting.

29:13What you need to do to actually take advantage of those tools is locked into a particular surface or a particular platform. We believe we can take the power of what we're doing with Superhuman Go and bring note taking to all your services. And the other way around, you can make a note taker much smarter. So imagine having a note taker that really feels like a virtual assistant, but actually knows everything about you and knows everything about your past meetings, knows about what's happening in your company, knows the actual names of the projects and the tasks and the people, and can really inform you as you work your way through that note taking journey.