US AI Policy Debate Reignited by Kimi K3, Where Anthropic is Lagging, AMD’s New Helios AI System

0:13Welcome everyone to the information's TITV. My name is Akash Pasricha. It is Tuesday, July 21st. Today on the show, we are unpacking AMD's new rack scale system, Helios. We'll then get into the policy debate around Kimi-K3 and open weight models. that is sweeping through Silicon Valley and DC. We've got two experts coming on who are going to help us understand both sides of that conversation. We'll also get into our exclusive reporting on human data labeling company Mercor and how concentrated its revenue is. And to close out the show, we'll take a look at why Anthropic is lagging behind rival AI model makers when it comes to voice agents.

0:54It's going to be a great show, so let's get right on into it. AMD is preparing to ship its first rack-scale AI system, Helios, taking direct aim at NVIDIA. I want to bring on Austin Lyons. He is a senior analyst at Creative Strategies. Austin, welcome back to the show. It's great to have you here. Hey, thanks for having me, Akash. Always good to be here. Okay, so Helios, break it down for us. What is it that AMD is launching here? Sure. So Helios, this is AMD's first rack for AI inference and also for AI training. So previously, AMD has had GPUs, they've had CPUs, they've had networking, but they've only been like eight GPUs connected together in a server.

1:34Now, AMD is launching an entire rack full of 72 GPUs, all of it interconnected, all of them acting as one system. And they're really targeting frontier model inference. Okay, so they've never had any kind of a rack system. This is something that only NVIDIA has had before. What's the competitive landscape here? Yeah, yeah. You could technically get racks of AMD stuff, but really the size of system was like eight GPUs in a server. And you could put these servers in a rack, but they weren't all acting as one big brain, if you will. And so this is now NVIDIA's had NBL 72 with the Vera Rubin. And this is AMD's opportunity to finally have a rack scale system that all behaves as one and therefore can really squeeze out as many frontier tokens as possible out of one system.

2:29In other words, so when you bought an AMD server, you could build a rack on your own, but that was all up to the customer to do. That was all integration they had to basically invest in themselves. AMD is saying we'll sell it as a whole product together for the first time. I wonder, what does that mean for, like, pricing? I mean, does AMD, you know, do you think that they offer, like, a discount on, like, if I were to get eight servers, put them together in a rack, like, is it just eight times one cost of one server? Or is there discount pricing? How does it work? Yeah, yeah, yeah, good question.

3:06And so they do have partners that would build these racks for their customers, like a Dell or a Super Micro. But they were all connected on what's called a scale-up fabric, which is just like all these GPUs talking really quickly together. And so now that they're all a rack and they can all talk really quickly together, they can act as one system. So what does that mean from pricing? I mean, you know, at the end of the day, AMD has always been a little bit behind NVIDIA. NVIDIA always has premium pricing. AMD is trying to come and offer competitive performance, but also generally a little bit lower price.

3:44And so I think when you think about pricing, you can think about NVIDIA type prices, but just a little bit less so that they can be as competitive and obviously have a way to get into the door with customers. So how competitive is this with NVIDIA? Is it as good as what NVIDIA offers at the end of the day with their racks? Yeah, good question. I mean, it obviously depends on who you ask. If you ask NVIDIA, they'll say no. If you ask AMD, they'll say yes. I'm asking Austin Lions, Austin Lions. Yeah, totally. No, I do think it's competitive. You know, I think there's this concept of Pareto Frontier.

4:19And at the end of the day, you're trying to squeeze as many tokens out as possible. And you want those tokens to be as fast as possible. And usually there's a trade-off. The faster the system responds, kind of the fewer tokens that it'll generate at a time. and nvidia and amd are both uh you know pushing the frontier here and and amd is is going to be competitive at certain spots on that on that frontier you can get that same level of performance from from amd okay what does this tell us about amd's strategy broadly speaking in its next phase of products hmm so amd strategy so amd has always been a high performance computing platform company They have a portfolio.

5:01They have a portfolio of CPUs that they want to offer. They have a portfolio of GPUs. And here, what's really interesting is right now, the data center is sort of the computer, if you will. Meaning, you know, even for inference, you want racks of GPUs and high-performance CPUs. And especially today with agentic AI, you've got more and more demand for CPUs. So literally, you need extra racks of CPUs just to run. Like when I'm using Cloud Code or Codex, and I spin up, you know, I'm trying to write a bunch of software. You've got all these agents using tools, writing software, testing things. And actually a lot of that happens on CPUs.

5:38So what's interesting is AMD is catching their sort of inflection point right now in the back half of 2026 with very performant GPUs. They already has the history of having really performant CPUs and it's just perfect timing with them with the agentic inflection, if you will, because AMD is showing that they've got the GPUs to do the inference. They've got the CPUs to help with the agentic AI. And really, they're just trying to, you know, come and compete on all dimensions. What's the reception being from customers to this Helios rack system? Yeah, customers are excited. AMD historically has lagged from a software perspective, and they've been a little bit behind from sort of having the rack scale offering.

6:21Finally, AMD software is definitely catching up and it's robust. The racks are going to be there. And at the end of the day, customers, Meta, Microsoft, OpenAI, they just need compute. And they are willing to deal with standing up NVIDIA systems, standing up AMD systems, having to manage multiple vendors and having to write software that works across the different systems. Customers are saying, that's fine. We will do that because we just need more compute. And so they're excited to have more compute from AMD. And then also, it's always good to have competition. It brings out the best. It also gives you pricing opportunities when you can buy from different vendors.

7:02And so generally, customers are happy to have more competition in the market. But I mean, on the topic of competition, is this just NVIDIA's roadmap like a year or two later? What do you think? You know, that is fair pushback. And you could argue that it is. It's AMD offering a similar solution, just a little bit behind. On the other hand, I think what AMD brings to the table now in terms of agentic AI is a broader CPU portfolio than NVIDIA. So if I'm AMD, I would be counter-positioning and saying, hey, yes, we're bringing you these rack scale inference systems that you need, but we also have many different CPU SKUs that can help depending on the type of agentic AI inference that you're trying to do.

7:51Because at the end of the day, it's the whole system. It's about the GPUs, but it is also about the CPUs. And so AMD's CPU business compared to NVIDIA's Grace CPU business, AMD's CPU business is a lot larger? Yeah, so great question. That's a very good question. So AMD, their server CPU line is called Epic. And they have many different, when I say SKU, SKU, it's like many different flavors of this, different core counts and different performance. NVIDIA just has one CPU. And that CPU was always built in purpose of serving the GPU. And so this is how AMD can differentiate, is they can say, hey, guess what?

8:30You know, NVIDIA has these really high-performance CPUs. It's really good for feeding a GPU. and that's awesome. But they don't, for example, have very high core count CPUs, which if you're running lots and lots of agents, depending on the type of workload, sometimes you want an agent that can just talk really fast to the GPU, but other times you're okay with a little bit of delay there if you can have hundreds or thousands of agents running just on one CPU that has, say, 256 cores. And so depending on your needs, AMD has a much broader port, much more diverse CPU portfolio. So basically what I hear you saying is for people who argue that this is an NVIDIA-esque system a couple years after the fact, your argument is yes, and also because they have different flavors of CPUs at AMD, there are actually workloads and scenarios where it would be more appropriate to use the AMD rack system given that they have all these different CPUs if you're running more agentic workloads, for example.

9:39That's kind of the distinction that I'm getting from you. Yes, yes. You put it very well, Kash. Thank you. So let me ask you one more question about the pros and cons of going with a rack system, because NVIDIA has gone with this approach. It can be frustrating sometimes for customers too, to say, well, I just want to buy the chip. I don't want to buy the entire rack. I mean, is AMD saying that we are migrating entirely to a rack-based system, or is this just sort of an extra offering that they're giving, saying if you want a rack, we can sell you the whole thing? Yeah. Okay. So that's actually a really good question because hyperscalers, they want racks, they want data centers full of this, but not everyone does.

10:21If you think about lots of enterprises, just your typical Fortune 500 companies, think of, you know, like Fidelity or some large healthcare, they don't necessarily want to buy a full data center full of racks. So AMD actually, within this 450 series family, they've got like the high performance MI455X, but they also have something called MI440X and MI430X, which are different flavors and you don't have to get a full rack of them. You can get just back to that server that has eight GPUs in it. So they also have more of a portfolio Leo, of different offerings, depending on if you're an enterprise customer or if you're the government, you know, you might want, you might only want to spend a million bucks and not 10 million bucks, for example.

11:05Right. Last question for you, Austin. What else are you watching for this week from AMD's customer conference, which is set to kick off, I think, tomorrow? Mm-hmm. Yes. So I'm hoping to hear more about the rest of their Inference Roadmap, so their MI500 series. I'd like to hear more about them from networking, because at the end of the day, it's not just what do you have for a CPU and GPU? It's how well can you help them communicate? And so that's networking. They've got something called Pisando Chip. Hopefully they talk more about that. And then obviously at the end of the day, everyone's just hoping to hear more from customers.

11:38You'd love to see customers get up on stage and say, we're excited to deploy this so that it's not just AMD talking its book, but letting its customers sell AMD. Right, right. Well, Austin, I want to thank you for coming on. That is Austin Lyons, a senior analyst at Creative Strategies here on TI-TV. The impressive early reviews to Moonshot AI's Kimi K3 model have sparked debate around how U.S. government policy should consider AI models developed overseas and whether any policy is actually appropriate at all. For more on that, I want to bring on Neil Chilson, head of AI policy at the Abundance Institute, and Ryan Fadasiak, a fellow at the American Enterprise Institute.

12:19Welcome to you both. It's great to have you here. Thanks for having me. So Ryan, I want to start with you. Frame this debate up for us. KimmyG3 comes out. We get a lot of posts on X. We get op-eds as well about what America should do with this model. Just walk us through the different sides and what the reaction generally has been. Sure. Well, I think the headline is that it's an impressive capability and we shouldn't be complacent. On the one hand, it seems pretty clear that the American moat in building frontier AI software might be softer than people had thought. And now, you know, we should anticipate some equilibrium in the race to build better and better AI systems.

13:01It's clear that Chinese labs will be able to distill and openly release a version of the American frontier in a matter of weeks after U.S. labs. On the other hand, the race really does seem to be on to build and install vast quantities of computational power. Kimmy's a very large model. Moonshot had serious problems maintaining high uptime when it first came out. And now we're going to continue to have this fight to get AI into customers' hands. Okay. Neil, does that encapsulate the debate as you understand it? What's been your reaction? Yeah, there's a lot of dimensions to this debate, but I think Ryan covered the key business points.

13:43On the policy side, the question is, what could the federal government do to improve the U.S. position vis-a-vis China on open source AI models? And is it interested in doing so? Right now, most of the tools that especially the White House could deploy would lead to restrictions on how U.S. customers can use these tools. That seems to put us in the backseat, not the driver's seat. And so the question is, how can we build an ecosystem in the U.S. on open weights or open source models that can keep up to where the Chinese are, even if their open weight models are behind where our closed weight models are?

14:30So, Neil, I mean, what policy do you think is appropriate? Where do you stand on this? Well, I think there's a lot of things that the government could do. To the biggest one, Ryan pointed to the right bottleneck right now. It's less about incentivizing the building of models, and it's about the bottleneck around compute. We have a lot of barriers to building big projects in the States and in the U.S. China, especially when it comes to energy, which is a critical input, China, for some reasons that are good and bad, has fewer of those barriers. And so I think there's lots of things we could do to streamline the building of infrastructure.

15:12On the policy side, I think the risk here is that we overreact to Chinese open models and we try to clamp down an open source overall, which is a key input to competition, but also to research. And it's part of the ecosystem that has kept the U.S. great in the software development business. I think clamping down an open source would be a bad alternative. We should focus on building infrastructure faster. Okay. So Ryan, Neil says clamping down would be a bad alternative. What do you think? I think it depends on the way it's done. You know, there are different problems we're trying to solve for.

15:50And some of the concerns we have are because the models are built by China. And some of the concerns we have are because the models themselves are open source and ungovernable. You can think of a few, I think, baskets of concern that the federal government might have. One is about data poisoning and indirect control in a three trillion parameter system that's very difficult to audit. An entirely separate problem is capability uplift for malicious actors. The fact that open source can serve as an always available tutor for would-be bioterrorists or cyber criminals. And these are vastly different baskets of risk that the government has different tools it's capable of wielding.

16:29I frankly agree with Neil that I think it's wrong to crack down on open source as a category of software. But I was heartened to see, for example, Secretary Besant say today that the United States will take action to protect intellectual property of frontier labs. I think many people have this kind of misguided view of China as a sort of Robin Hood, stealing from rich frontier labs in the United States and giving to poor software developers. but this misconstrues the way we handle intellectual property in this country. And it confuses the way that China has approached IP and many other fields.

17:05Right. So Neil, I mean, both of you seem to be sort of on similar sides here. And so I just want to sort of play this out a little bit. I mean, one question I've had is what policy could even look like the people who are advocating for a clampdown on these models. I mean, I'm trying understand what legal mechanisms they would have to lean on what that policy could actually look like um again knowing your stance and and what it is what do those folks argue how novel a theory would this have to be in terms of policy well i i think what the most likely thing we would see here would be something that looks uh much more soft law than it does hard law uh software is very difficult to constrain.

17:53And so a lot of it would be essentially trying to find choke points. Maybe that's data centers, maybe that's the hyperscalers, and putting pressure on them to not allow or to slow the deployment of these types of models. We talk about them being open weights, but they still need to run on big computers for the most part, most of these large models especially. And so finding that as a choke point might be one policy lever that they might pull. Congress could possibly do stuff in this space. It doesn't seem like a huge priority in Congress right now yet. I think the urgency is more in the executive branch and their tools are more limited.

18:40And I think they would ultimately have to look more like soft power, persuasive, cajoling than than putting on export bands or import controls or other things like that. So basically what this would look like is like, instead of a model being hosted in a data center in Asia, for example, you would have to have the model hosted in the US, basically. That's one way you could see this playing out. Well, it could be that, or it could just be that the pressure is because we have a lot of the world's compute, local to the US, within the US, the pressure could be on those US companies to have standards for what types of models they host and offer to their customers.

19:28That could, because of the nature of the Internet, that could drive people to maybe move compute overseas. But given the strong advantages that we currently have, I think those types of rules on the margin could be effective to slow open source possibly. Ryan, what do you think? Yeah, please do. Yeah, let me weigh in here. I think there are kind of two directions the policy could go. One would, in an extreme version, try to restrict the availability of Chinese open-weight models. And there, like Neil is saying, it's a really hard problem set without infringing on free speech. We've ruled in this country that code is speech.

20:06And so it's a difficult value proposition to tell a cloud provider or to tell a download platform like Hugging Face that they're not able to host Chinese open-weight models for download. An alternative approach, and the one that I think the administration would be more likely to go in, is to enforce these kind of soft requirements around labeling the provenance of models. People should be aware when they're using a product that is using a Chinese open-weight system so they can evaluate for themselves whether they're okay with the provenance of that information, whether they're okay with their data transacting with servers in China, which, as we know, are made available to China's intelligence services.

20:47I think, frankly, increasing transparency and awareness around the use of Chinese systems is the more likely and durable framework the administration will gravitate toward. Neil, do you think that KimiK3 actually changed the conversation here? Did it just bring opinions to the surface that were already there, or did it change people's minds, given how good the model And I ask this because, I mean, two years ago, I mean, we had DeepSeek. Maybe it was two. I think it was two, one year ago. I can't even remember. It feels like a decade. Yeah. You know, we had DeepSeek. We had the market crash. There was not even that big a market swing here with Kimi K3.

21:28But it does feel like people were talking about the policy elements more. So do you think anything changed or was it always there? I think it's just bringing people who hadn't thought about these issues into the conversation more than change the actual debates. I mean, the open source, is open source more secure? Is it less secure? Is a question that's as old as open source? AI brings some new dynamics to that, but I don't think the debate has changed fundamentally with the release of Kimmy 3. I think there's just more people talking about the same issues. Okay. Okay. Ryan, do you see one side winning here at all right now in the conversation based on what you're reading?

22:12Here's the thing. I consider myself a China hawk. I'm concerned about security issues vis-a-vis China. But I also think it's a fool's errand to try to control the development and proliferation of software. I just don't think that's a fight. The United States is capable of winning. We tried with encryption and we've tried in other areas. And so really, I think it's about making the United States safe for the development of open source technologies, regardless of where they're developed. I believe the best defense is a good offense and that the United States should seek to lead the world in the development of open source AI technology.

22:48And on the other hand, it's probably wise to make sure we do enforce some transparency measures to make sure that consumers are educated about where the information they're sending to be processed is being processed when they're using a Chinese AI model so they can make that decision for themselves. Ryan, when you look at the stances people take on this on the left and the right, how does that shake out? Are there people, does it generally swing depending on party lines or are there people on both sides that kind of believe both things? I don't know that this is a partisan an issue. I think some of the concerns people have about Chinese open source are well-founded and they transcend party lines.

23:30Like I said, it's about capability uplift for malicious actors. It's about data poisoning and indirect control. It's about the potential for propaganda and political censorship that are embedded in Chinese systems. And finally, the concern people have is about economic viability of American frontier labs and permitting widespread competition. Where people fall on each of these four issues doesn't tend to be partisan, but it is, I think, an extremely fraught set of issues that can be tackled with different policy tools. Like I said, I think it's probably the wrong approach to crack down on open sources, a basket of technologies.

24:09It's just not how technology works. Neil, Ryan mentioned the labs here. Do you have any insight into how active the labs are behind the scenes here in D.C.? I think all the major labs have certainly upped their game since this has become a major issue of state and federal litigation, and especially since the White House has gotten involved. On the open source issues, the labs, especially the labs with closed models, they have a very particular business interest. These are in some ways competitors to them. And so I think that while it's very hard to be against open source, generally speaking, I think that is something that also cuts across all party lines.

24:56People tend to think of open source as a huge benefit to the U.S. technological innovation and economy, and I think that's true. But I think the labs, when it comes to open weights, they do have business concerns and are probably more skeptical in many ways of open weights than many of their customers who are trying to minimize their token spend might be. And so we need to keep that in mind as we evaluate how the labs are talking about these issues. Neil, last question for you. I mean, how quickly do you think this ultimately moves? We have an election coming up again. The more immediate deadline, I guess, is we have the policy framework that is due and people are waiting to see what's in that.

25:48So how quickly do you think we actually get clarity here on this issue? I don't think clarity is coming fast. I think there may be policy actions, but this is an area of deep nuance and complexity. So I wouldn't expect clarity around how we're going to treat Chinese open models or open models overall anytime soon. I think there will be continued discussion, and that will probably give you and your listeners lots more to talk about. So like you're saying, we can talk about this for a year and still have runway. I suspect that the back and forth between China and the US on open models will continue.

26:28And we will not have solved it definitively. And overall, I think that's probably a good thing. And these complicated ecosystems need innovation. They're very dynamic now. That makes them hard to regulate, but it also means there's huge opportunity for US entrepreneurs and customers. Great. Well, I want to thank you both for coming on. That is Neil Chilson and Ryan Fadaziak here on TITV. The information has exclusive reporting on human data labeling company Mercor's latest financials. This is a booming category of AI. I want to bring on Corey Weinberg, our senior reporter covering Anthropik, to talk about what he and our colleague Julia Hornstein found.

27:12Corey, welcome to the show. It's great to have you back. Good morning, Akash. Okay, so how much money is Mercor generating? What did you find? So we found, I mean, look, these data labeling companies are always some of the sort of most surprisingly large recipients of this AI boom. So Mercore, we found on a very top line gross revenue basis. So all the money basically flowing into the company, that's between$610 and$620 million in just the first half of this year. Mercore keeps roughly between 30 % and 35 % of that revenue after a lot of it is paid out to their contractors. uh and so this is a company that sort of has an annualized run rate of over 2 billion it expects to be at 2.8 billion by the end of this year uh and so the numbers numbers are getting pretty pretty large for for mercoir wow and so i i just want to break this down a bit for people so i mean they're doing about 600 plus million in the first half of the year uh your findings suggest that that has been accelerating throughout the course of that first half.

28:29And because it's accelerating, that could get that. I mean, it's more than two times 600 is 2.8 for the annualized rate, which is pretty staggering acceleration, I should say. Yeah, it is. I think it speaks to on a positive end, we were I think we continue to be somewhat surprised by the resilience of the human data market. I think for as long as I've been writing about these companies, you know, going back to scale AI and sort of their rise, I think there's always been sort of a question of, won't these AI labs just start using synthetic data, or won't we be able to cut humans out of the process?