Google DeepMind's AI Autonomy, Anthropic IPO Voting Control & the New AI Safety Group

26:32Jing Yang:So we just reported that DeepSeek's annualized revenue run rate, ARR, just reached$1 billion. And this really shows revenue growth is accelerating because a few months ago in July, it was less than$500 million. So it has more than doubled.

26:54Juro Osawa:And that seems like a pretty impressive and fast growth rate. But how does this$1 billion ARR compare to revenues of ARR achieved by other Chinese AI labs? And also, how does it compare with the likes of OpenAI and Anthropic?

27:14Jing Yang:So this ARR, compared to other Chinese companies, especially the independent labs, it's kind of similar ballpark. it you know deep seek is kind of catching up and for example zai you know jupu so they said recently that their arr has reached 1.8 billion dollars so it's still bigger but minimax another listed company they said recently that arr is 800 million dollars so deep seek is current revenue is really catching up to some of those. But compared to US labs, it's still a tiny, tiny fraction because Anthropic OpenAI, their ARR numbers are in the tens of billions. So just orders of magnitude, much, much bigger.

28:06Jing Yang:So, yes.

28:08Juro Osawa:Yeah, I see. But what's very interesting in your story also is that you reported a couple of months ago that DeepSeek's gross margin from selling cloud-based access to its models were more than 80%, which sounds quite high and then even higher by OpenAI and Anthropik. So what is the reason behind DeepSeek's really high gross margin? Because we know that DeepSeek's token cost, its model still cost among the lowest, among the major language models.

28:47Jing Yang:That's right. Yeah. So DeepSeek's pricing is definitely, you know, it's one of the cheapest, you know, model providers. But what DeepSeek is very good at, according to, you know, all the people in the industry that we talk to is really their ability to keep the cost of running AI models very, very low so that, you know, the same amount of chips can run more or, you know, just to do the same inference. They require much fewer chips and including memory chips. You know, the cost of all these chips has really weighed on all of the major AI labs, but DeepSeq has kind of figured out how to do it, you know, very efficiently.

29:30Jing Yang:So that's what people cite, you know, including, you know, investors, their view of why DeepSeq has kept this margin very high. And 83 % margin is very, you know, indeed impressive considering, you know, their low prices. I see.

29:49Juro Osawa:That's very interesting. So these numbers were shared, you know, as you also reported, these numbers were shared by DeepSix CEO Liang Wenfeng in a recent closed-door meeting with investors as the company is, you know, looking to finalize its second funding round. What else did the CEO say to investors about Deep Seek's future strategy and its priorities?

30:15Jing Yang:So generating revenue, even though revenue is growing, you know, as we showed, but generating revenue is not the top priority. And the CEO, you know, kind of reiterated that their, you know, really priority is to train their new models, right? And I think they need more compute in order to, you know, train their new model. So DeepSeq is putting, still putting like the majority of their resources into the training of those models. And more than 70 % of their compute capacity goes into training. And then the rest goes into, you know, the running inference of the models. And one thing that, you know, we recently also reported from the same meeting, the CEO said, you know, their major priority is to, you know, train new models using Chinese domestic chips, you know, from Huawei.

31:13Jing Yang:And they've been working closely with Huawei, you know, in order to do more training with Huawei chips. And he said he expects Huawei to deliver those training chips, start delivering in the fourth quarter of this year. So this effort is going to accelerate. I see.

31:32Juro Osawa:I mean, isn't it quite refreshing to hear a company say that, you know, to hear a company whose revenue has been growing so quickly and at a very high margin to say that, well, revenue growth is not our focus.

31:47Jing Yang:It is interesting, especially for a company, who is in the middle of raising this second funding round. They're raising$7.5 billion, about$75 million US dollar valuation. And also they are preparing for an IPO down the line in Shanghai. So for a company like in that situation to say that, well, generating revenue is still not the priority. But I think people who know the founder, it has always been kind of his philosophy which you know he has talked about many times before that you know they their priorities still you know try to develop the models and deliver them you know make the model available to you know everybody and that's why the pricing is still low even after the price hike their pricing is still you know one of the lowest so that part But he seems to maintain that focus.

32:47Jing Yang:Of course, they do need to raise money and they want to go public. But that part seems to remain consistent.

32:56Juro Osawa:And we'll see how public investors in China would be responding to this kind of ethos when JPC goes public. This is a very interesting story that we will continue to report. Thank you, Jiu-Ru. Thank you.

33:15Stephanie Palazzolo:A new op-ed published in The Information suggests that America's AI ambitions in some ways aren't as well positioned for success as China's AI ambitions. The piece was written by Ryan Cunningham, founder of Edgerunner Ventures, and Christy Loke, an expert on all things AI in China and a MATS research fellow. I want to bring on Christy to share more about her perspective. Christy, welcome to the show. It's great to have you here. Thank you so much for having me. So I want to start with this line from your piece. You wrote, the country, the US, is financing an industrial revolution most citizens currently want no part of.

33:54Stephanie Palazzolo:Can you unpack a little bit of what you meant by that? Yeah. So currently there is increasing, we're seeing early signs of public backlash against AI, whether it's against the dangerous elements of AI, like the existential risk kind, or the fact that it seems like some of the AI data centers are being built in people's backyard and it's not officially appealing. So there is also, of course, this growing sense of AI fatigue that a lot of us feel. And that is not going to get solved easily, I would argue, unless if something changed in terms of the way that the U.S. government or the U.S. industry leaders go about articulating the vision that they have for global or inclusive success.

34:38Right.

34:39Stephanie Palazzolo:And I start with that because that is very much the context with which we are going into this conversation about what America's AI strategy is and what China's is. One of the things you do really well in your piece is you sort of break down what you view each of these strategies to be. And so let's start with the U.S. How would you define the U.S. ambition for AI right now in your words? Yes, I think a lot of it has to do with frontier AI maximizing and chasing. So the massive data center build out is a piece of this. And in the past, you know, a lot of the US strategy is about containing access to, you know, superior chips by the Chinese.

35:23So that's another kind of prong of that. The strategy is very much about accelerating at full speed, whilst making sure that the US continues to lead at the frontier.

35:34Stephanie Palazzolo:okay and but going back to what you were talking about i mean there is sort of another way of looking at this ambition which is that we just don't want to lose to china basically there's a lot of uh at least rhetorically there's a lot of the uh but china right like we want to do x y and z but china is uh um it's a looming threat you know if it gets there first then everything is ruined. I think to quote Basson very loosely, as was said recently. Right. And I mean, I want to ask you, I was at a dinner the other night and we were talking about this topic. And one of the comments that came up is, well, the US has not always talked about their technology ambitions in relation to what China is doing in the sense that the fear that China is developing things, we need to beat them.

36:31Stephanie Palazzolo:That narrative wasn't as strong or prevalent or driving the ambition 10 years ago than it is now. Do you agree with that? And what do you think has changed here? Yeah, I think for a long time, the US essentially was the technological hegemon, not just political hegemon. And so it's a very comfortable part of the identity as well. And China's rise technologically does challenge that. And I think a lot of folks in DC, a lot of folks in the tech world, but less so than let's say DC, still are a little bit delayed in their response to China's technological rise. So there's a lot of emphasis on distillation and a lot of doubt even post-deep seek that, oh, China can't innovate.

37:20You know, this can't be true, which is one way to stay comfortable. But it doesn't really chime with reality as we see it.

37:31Stephanie Palazzolo:Right. Okay. So what about China's ambition in AI? How do you view that and how is it distinct and different from the way that you think America is approaching us in a maximalist and sort of, in some cases, fear-based ambition? Yeah, so the way that China sees AI is, I would say, more multidimensional. And the reason for that is because of how they look at technology. I think the party state very much thinks, when they look at technology, they're trying to find solutions to existing problems, like essentially cures for existing problems. And for them for a long time, including now, the priority problem is how do they exit the middle income trap?

38:17How do they build an innovation ecosystem that allow them to produce some of these frontier or frontier adjacent productivity enhancing technologies? And so that's really their preoccupation. And the reason why I know this is because I actually found a document from 2006 called Short Medium Long Term Planning for S &T Innovation. And back then they were already talking about how to use like emerging technologies and different types of technologies to build a more robust foundation for development in the long term. And so when they found AI Soda in 2012 and Xi Jinping became very interested in the technology, It really very much is to solve the economic problems at home, which are, of course, linked to geopolitics and security and things like that.

39:07Stephanie Palazzolo:Do they care about pacing? What's interesting is that the strategic framework that they use when they think about S &T priorities and AI priorities has four prongs, right? And then, in fact, the fourth prong is all about governance. It draws the red line at, you know, if the technology that China's developing is no longer meeting, like making people's lives better, then the purpose of the technology ceases to exist. Which is really interesting. And because they draw that line there and because they have always been governance pilled, I think since 2018, when global discussions around AI focuses on both risk and opportunities.

39:55Yeah, I think very much they are incentivized to care. And also from a political perspective, disorder is no good, right? Domestic disorder is no good. Chinese open models wreaking havoc in the world is also very, very bad for PR. And so for all those reasons, they're very invested in essentially catching up to the U.S. safety frontier beyond just, you know, keeping up with the developmental frontier.

40:21Stephanie Palazzolo:Really quickly, you mentioned that that's the fourth prong. what are the first three prongs? Just lay that out very quickly for us. Yeah, of course. We talk about those three prongs in our piece as well. And so the first prong has to do with being close enough to the frontier. And the reason why that exists is because for a long time, China's really good at one to a hundred types of technology. So from tech to adoption, they're very good at that, but less good at basic technologies, right? And so adding that there is to incentivize like scientists to aim high, touch high is how they say it. The second component is actually about productivity gains from technology.

41:01And they often talk about it as the most important prong out of the four. Because if technology, according to them, can generate productivity gains, the productivity gains can in turn lead to more technological gains. And so it's a very positive loop. And the third component is really a sign of the times. Because when this strategy came out, when this four prongs came out. It was two years, two years-ish after the first export controls against like Huawei and CTE in China. And so they did a lot of soul searching and they figure out that in order to survive, you know, this sort of challenge, they really have to add technological security, supply chain security around core technologies to the third prong.

41:46And so that is the third prong. And the fourth one is about making sure that people are okay. You know, people's like livelihood, life and health are okay.

41:54Stephanie Palazzolo:So let me ask you this. And so if those are the way that China is approaching this, what do you think the U.S. can learn from that approach? And what does China's ambition or the way that they laid it out, what does it enable them to do better than what the U.S. has positioned itself to do right now? Yeah, by just having multi-prongs and having a diversified approach to AI development as early as 2020 when this came out, I think that set the Chinese political leaders, but also people that they then give signals to, you know, decide on how to implement and kind of, you know, actually implement, execute and get them to these goals.

42:45Um, so yeah, so that was very, very important. Um, and in terms of how the U S can learn from this, I think this really inspires in the sense of, oh, you can actually look beyond frontier AI chasing and you can also balance the governance goal, the risk management goals that you might have as risk emerge. Um, and you can also do this at the same time when you have to deal with security risk, Right. For example, over-earth is a part of this for the US side. And think about how to distribute the gains, how to make sure that people feel like they're on board with this fission for AI. They feel like they're actively benefiting from it.

43:28not because it's either damnation or you know like abundance but because maybe there's a middle path you know like in industrial revolution style but like a slow path where everyone can become more productive but also see you know incremental gains um and everyone can access the technology

43:49Stephanie Palazzolo:for example um through it's funny it's funny because you know i i think the researchers are all doing the same work on the ground um and it seems to me like this is all just a comms challenge really you know from the from the government's perspective it's like you know that's what we're talking about here is like how to really socialize uh the the work that your nation is doing and you know to what extent that incentivizes the uh positioning of these products and stuff like that last question before i let you go how much of the difference in these strategies bubbled up in this week's meetings between china and the u.s and the country's leaders i mean uh they were supposed to talk about ai i don't know that they really got anywhere materially but did we see the differences in strategies play out yeah so that's a really interesting.

44:48That's a really interesting question because the way I see it is although, you know, I was emphasizing how the way that the US talks about AI and the way that it prioritizes different parts of AI is very much like pro Silicon Valley, pro this AGI dream, pro American leadership. But at the same time, because of events of the past couple of months, you know, Trump is like legitimately freaked out, right? A lot of people in the administration are freaked out about the cybersecurity implications of a mythos-like model and the next iterations of it. And so I actually am on the more optimistic side where I feel like the two countries, because China is getting very close to the frontier, they're starting to see increasingly so, but not yet, similar-ish risks.

45:34And so the stakes of managing these risks are really high and are increasingly shared because of it, which is why, you know, Bessin and Halifong were able to have that meeting and come to a similar understanding of potentially the need to have an incident communication channel or mechanism. So that if, say, you know, a U.S. or Chinese model did something bad, you know, unintentionally, the other country's leaders and technical people will be able to say, OK, that was unintentional. So that level of the de-escalation is actually really important. And I see it at the summit as well. And in terms of what was kind of striking for me was actually C's subtle, not so subtle response to people who are advocating for pacing at the frontier or pacing the frontier.

46:23So the way that he framed it is because the US and China are major AI powers, they actually have the ability to, but also have the responsibility to manage these risks and make sure the technology goes well. So it's very much like a can-do spirit to the pacing question. And so what I do expect is that China's going to continue to invest in AI safety, going to look at the US very closely in terms of what the risk debates are, and try to get ready for worst case scenarios. And for Trump, because although he's very pro-leadership, he's also pretty scared of some of the less predictable AI security risks.

47:07And so they're going to continue to be able to work together in some narrow capacity. but that doesn't solve the domestic problems, right? The domestic problem for the U.S. is still how do you gain trust and have the public agree with this vision and feel like they can gain from it? And that is -

47:28Stephanie Palazzolo:Brings us right back to where we started the discussion, which was people don't have a lot of trust in it right now. So Christy, it was a great piece, a thoughtful piece. And I want to thank you for coming on and sharing with us more about your perspective That is Christy Loke, a research fellow of Matt's Research here on TITV. Earlier this week, The Information launched our brand new AI Inflection Index. It is a new data product for pro subscribers that brings insights into key thematic opportunities in the AI area. The project was published by Eric Bellamo, our head of research at The Information Pro.

48:10Stephanie Palazzolo:He joins us now to break it all down. Eric, welcome back to the show. It's great to have you here. Thanks for having me. I should say, welcome back. It's your first of many. We're going to have you on more and more, though. I mean, this is, you know, the Information Pro is, as a former Information Pro reporter myself, there is so much to dig into, and I'm excited to. So tell us about the AI inflection index. What does this thing actually show? Yeah. Well, again, thank you for having me. Indeed, the first of many. Happy Friday. Yes. So the inflection index aims to basically present quantified thematic insights based on our reporting and our corpus.

48:57So if you think about it, even over the last five years or so since the chat GPT moment, we've had about 10 ,000 documents, both articles and briefings that go in detail on a vast array of topics. So this index quantifies and maps over 80 themes on AI alone. And you could imagine as a subscriber or an investor that you might read one piece, go about your day, your job, make an investment, what have you, and then find yourself with a deluge of content, right? There's more content moving faster than ever. And this index aims to basically track those themes and track these developments over time.

49:42Stephanie Palazzolo:And give us a taste of some of the themes. Because one of the interesting parts here is you can see what themes are becoming more prevalent, which themes are sort of declining in terms of relevance to the moment. So give us a snapshot right now, what's trending up and what's trending down. Yeah, exactly. Again, this is a multifaceted development. There's a lot happening in real time. One that I actually like to highlight is agentic commerce. And I'm a former consumer analyst. So before joining the information, I covered consumer technology in general. And the idea of agentic commerce was one of the most visible kind of low-hanging fruits, if you will, in the world of AI.

50:27We could all imagine these personal assistants. But our index actually, I think, presents both a contrarian perspective and an accurate perspective. So we see the theme of agentic commerce emerging and gaining momentum in our coverage a little bit later than consensus, actually. But I think that's a more accurate read because we've had several checkpoints. We've had a wave of on-site shopping assistance. We've had the open claw moment transition to Muse. And even this week, we had headlines around Amazon and Meta and managing that traffic. And I think this is a more developed time to be talking about things like platform disintermediation, retail media networks, and so on.

51:10So that's one example.

51:12Stephanie Palazzolo:There's another one that really caught my attention. custom inference silicone? What are we seeing there? Yeah. So this is the idea of basically custom chip development. So OpenAI's jalapeno chip is a good example of this. And the idea being that large language models and generative AI becomes more specialized or more compute intensive, that we might need different chips developed for special purposes. And the idea that a frontier lab would be developing a chip in-house was a big deal and has significant downstream implications on the semiconductor trade. Also, I come from the world of private markets and the venture capital backing Challenger semiconductor startups, chips startups would have been hard to imagine not that long ago.

52:05And this year we've had significant financing going into some challenger startups. So, yeah, the overall idea being custom chip development beyond, you know, a typical offering, if you will, quote unquote, typical from NVIDIA.

52:21Stephanie Palazzolo:Another one for you that I saw, what's happening with data center build outs in terms of what you're seeing on the index? Yeah. So taking kind of one step back, right? Like the idea of data center and power for AI is a ubiquitous concept. Now we almost can't talk about it, can't not talk about it in a given day. But the idea part of behind this index is at one time, that was a percolating concept. That was at one time, just an idea people were getting their heads around before this became a macro trade, if you will. So we're seeing, number one, more momentum behind the coverage and the dialogue.

53:05But even taking a broader, a wider lens to this, it's interesting to think about. This is having significant implications on global treasury issuance, right? Long-dated treasury yields. these ideas were, this would have been hard to conceptualize not that long ago. And our index flags it as a watch list item, actually at the end of 2021. And by 2025, we have PJM, which is the largest grid operator in the US, estimating that 75 % of the region's increased capacity payments come from data center demand. So again, just to bring it back to the index, the idea is We flag an emerging concept early, and we see it incrementally priced by financial markets.

53:52And the idea is to give our readers a leg up on those developments. Right.

53:57Stephanie Palazzolo:And last question for you, Eric. Are there any terms that haven't yet peaked right now, but ones that are on your watch list, ones that we should be watching for? Yeah, we might need a whole other segment to talk about those. I am biased, as I mentioned earlier, in the direction of some of the consumer developments. We kind of talked about consumer developments and commerce. I think bio and life sciences is a big one. I think we all can sense that there's going to be some breakthrough. I think another way to think about it is we might have to have that breakthrough. The idea that to kind of get the public back on board in some respects with the benefits of the tech, you know, maybe the data center buildouts are worth it if it means curing diseases and stuff like that.

54:52Yeah, we might need to cure those diseases to build that trust. But the idea, again, it's relatively small, relatively nascent in our coverage. It is generating outsized attention relative to the share of coverage. But I think you're starting to see some of these developments in the market materialize too. So Isomorphic Labs becoming Dennis Sebastis' primary focus. Some of the DeepMind leadership focusing there. Anthropic has made some purchases, M &A deals in the space as well. So I think we're all looking for that, both as long-term opportunity and also kind of an area of urgent need as well.

55:36Right. Great.

55:37Stephanie Palazzolo:Well, Eric, I want to thank you for coming on. It's a great project. That is Eric Bellamo, Head of Research at The Information, and leading our Information Pro product here at The Information. Pleasure's gone. 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. If you can't make it then, episodes are available on theinformation.com, on our YouTube channel, or wherever you get your podcasts. Make sure to follow us on social media on X, Instagram, TikTok, and LinkedIn. I am already excited for our next show tomorrow. Have a great rest of your Friday.

56:12Stephanie Palazzolo:Bye-bye for now.