The Pomp Podcast - China Is About To Catch The US In AI (Here's The Timeline) | Justin McAfee

Episode Date: September 8, 2026

Justin McAfee is an AI researcher at Delphi Digital. In this conversation, we break down why Chinese AI labs are closing the gap with America's frontier models, the innovations forced by US export controls, and the pricing war now squeezing OpenAI, Anthropic, Meta, and Google. We also discuss the race in humanoid robotics, whether AI can ever solve financial markets, and how the US should respond to stay ahead.======================Arch Public is an agentic trading platform that automates investment strategies across Stocks, Commodities, ETFs and Crypto. Whether you’re rotating into AI & Gold, allocating to the S&P 500, or accumulating Bitcoin, Arch Public executes your plan 24/7 without ever taking custody of your assets or funds. Sign up today at https://www.archpublic.com, and start your FREE automated trading strategy! ======================Simple Mining makes Bitcoin mining simple and accessible for everyone. We offer a premium white glove hosting service, helping you maximize the profitability of Bitcoin mining. For more information on Simple Mining or to get started mining Bitcoin, visit https://www.simplemining.io/pomp======================0:00 - Intro1:10 - Export controls & how constraints forced Chinese AI innovation4:01 - Could China's AI eventually eclipse the US?7:25 - Chinese AI companies & infrastructure plays to watch12:11 - Meta and Grok's price war, and the squeeze on OpenAI/Anthropic15:07 - How should OpenAI & Anthropic respond to the competition?16:33 - Physical AI & humanoid robots: China's lead18:39 - Can AI ever fully "solve" financial markets?21:11 - Geopolitics: how the US should respond to China24:43 - Biggest concerns: staying in the frontier26:56 - What's happening with AI in Europe?27:36 - Delphi Digital's research & where to find it

Okay, when I sell my business, I want the best tax and investment advice. I want to help my kids, and I want to give back to the community. Ooh, then it's the vacation of a lifetime. I wonder if my out of office has a forever setting. An IG Private Wealth advisor creates the clarity you need with plans that harmonize your business, your family, and your dreams. Get financial advice that puts you at the center. Find your advisor at IG Private Wealth.com.

China produces 2x the electricity of us. They're adding grid capacity six times faster than us. Their AI energy usage demand is like 1 to 5% of the capacity they've added in the past five years, whereas in the U.S. is like 50 to 70% of the capacity we've added. What's going on, guys? Today we have a great conversation with Justin. He's one of the AI researchers at Delphi Digital. In this conversation, we're going to get deep into the weeds.

What is going on at the American closed-source AI model labs? Why is China putting so much pressure on these labs? How's it changing? What you and I pay when we actually access these models? On top of that, what are the constraints that the Chinese models are facing? And how is that constraint leading to innovations that are actually making the U.S. labs a little nervous? On top of that, we talk about some of the American challengers and how it's changing the entire industry, both for software and hardware, and ultimately how maybe you and I are going to end up being the big winners

because competition breeds excellence and a better product at better pricing for everybody. This conversation is fascinating. I think you're going to get a ton out of it. If you're into Bitcoin, AI, or any sort of frontier technology, you're going to want to be paying attention to what we talk about in the next 30 minutes or so. Here's my conversation with Justin. All right, Justin, everyone's been talking about the American closed-sourced AI labs. They've been doing a fantastic job of bringing products to market.

But all of a sudden, there's a challenger that's come on the horizon in the Chinese AI labs. One of the things I find interesting is they have certain constraints that maybe the AI labs don't have, but constraints can breed innovation. It looks like that's what's happening with the Chinese AI models. You've recently written a great research report explaining what some of these innovations are. Can you just walk us through what the constraints the Chinese labs have faced and then what innovations that's now leading them to have on a breakthrough front? Yeah, for sure.

So China, with the US, we started expert controls back in October of 2022, banning the A100 and H100 class GPUs. From there, we did some additional tightening as well. And what this essentially did is put China in a situation where they had limited compute, limited memory, and like half the chip-to-chip bandwidth that US labs faced. So this resulted in a situation where they needed to innovate across like different levels of the stack in order to successfully serve their models on their cheaper chips. So in the report, I discussed, there's like three main innovation themes.

But first is the training and systems efficiencies, and this is dealing with that lower chip-to-chip bandwidth. And then there's the architectural efficiencies that they've done, which is how essentially you like route around the lower memory constraints. And then like one of the large areas, I would say, is on the post-training efficiency, which is taking a lot of the compute-intensive pre-training phase and shifting those capability games into, to the post-training phase.

So yeah, they've had to successfully, they've had to deal with these constraints overall. And it's actually what's pushed them to be so innovative and so cost effective to be able to compete against our frontier labs today. When you look at these kind of different levels of innovation, is there one that you point to where you're like, this is the one that really matters?

Or is it actually the fact that they've done it on all three layers that is leading to the performance gains that they've had? Yeah, I think. Certain innovations are definitely like more interesting. Like one of the big ones that I think is particularly cool is they've been going below the abstraction layer and hardware and doing like co-design with their own companies like with Huawei. But yeah, I would say generally it's across the board. Like each of these different layers has contributed to them being able to compete on costs, specifically with serving efficiency. And now it's a matter of will they be capable of reaching a frontier,

pre-training on domestic hardware. And my expectation is that happens probably somewhere within the next two years or so. So if they basically started significantly behind the U.S., and we're effectively working against them, right, with all these obstacles or friction points, but you still think that they can get to that frontier kind of positioning in the next two years, does that mean that their trajectory would put them on pace to eclipse the U.S. labs, maybe in the next five years or so? Do you expect the Chinese models to eventually be better?

I think the U.S., we have a pretty significant lead on the capability on the Frontier side. And when you look at the hardware supply chain in China, I think in like the more bullish scenarios, they're able to serve maybe like 85% of their compute by like 2028. So they get there, but like I would say that we have such a significant lead that is difficult to say that China, Chinese models like compete one for one at parity with Frontier U.S. models, but I think the case is more interesting that if we have a model that's 80% as capable of a frontier model, but one-tenth the cost, does that warrant switching for enterprises and developers?

And so far from some of the data we see from things like OpenRouter, like this seems to be the case. Now, why do you say that that's still a question? Like, to me, if you can get, you know, a significant cost savings, you know, one-tenth the cost would definitely qualify there. and you're close enough. One of the things we talk about internally is what is obvious to the naked eye? If you have something that ranks on a scale of 1 to 100, 100 being the most intelligent,

one being the stupidest, something's 100 and something's one, very obvious that there's a difference between those two things. What is the difference between like 85 and 90? Probably not that obvious to the naked eye, right, to the average person using the product. And so like how far away can you be and still fall within the, you know, kind of understanding of the consumer, oh, this is the same thing. And I don't have an answer, right? We constantly talk about just like there is some gap that is almost permissible or is like an error band where the end user does not know. And so if you get the cost savings with, you know,

kind of lower intelligence, I guess, you know, there would be a pretty strong argument for these companies to switch. Yeah. Yeah, I'm in the same boat. Like, I don't know where necessarily that capability difference is. But I think we're getting close. And like as the frontier pushes up more, I see a lot of people with the most recent Open AI Astero, at least that just came out. You know, there's people that are skeptical of benchmarks. And I think part of the reason for that is because your average user isn't necessarily using these frontier models on the things that they're becoming better at, like frontier math and frontier science work.

And so we're probably reaching the point where 80% capable of a frontier model would be useful for a majority of work that people are putting AI to use for today. And then, like, those higher sensitivity workloads and, like, the real frontier work probably still sits with, like, open AI and anthropic models at the top. Well, I also would say, like, 80% of a frontier model today is, like, more than 100% of a frontier model two years ago. Yeah.

Yeah. Right? So, like, as the frontier pushes further, also, it's pulling the, you know, cheaper kind of open weight type models. They're actually surpassing the old frontier models. And so in a weird way, we're like moving the goalpost while we're playing the game. Yeah, exactly, exactly. Okay.

When you look at these companies, there's a handful of Chinese companies that I think people know, the Huawei's and kind of hardware manufacturers, those types. Are there other companies that you think people should be paying attention to, whether they're public or private, that you think a lot of this innovation and kind of value capture is happening in China? Yeah, I think when you look at the innovations being done, there's like a couple areas that benefit certain companies. So for example, on the hardware co-designed side of what the Chinese labs are doing,

they're essentially reformatting how the model bit rates are in order to work directly with these labs or with the hardware companies so that they can successfully serve on them. So you have like CamberCon and more threads, Highgon, MetaX, a lot of these Then additionally, you have, like in the, on the memory side of the equation, you have companies there like CXMT and SMIC. And I think like similarly to what we see in the US with a lot of the AI investment thesis, you can essentially invest in some of the infrastructure, which is significantly cheaper than the labs themselves.

I think most recently the Chinese AI labs have, like, like a 65x to 100x multiple on revenues. So it's huge. Their values significantly more than the US labs. And so these infrastructural plays in the Chinese domestic supply chain become a lot more interesting, I would say, than the labs themselves. Now, when you're looking at the labs,

most of the companies that I have seen, they are releasing these open weight models. I don't know how many of them are closed source, but most of them seem to be open weight. Part of that seems to be part of this price competition. Are there other reasons why they are embracing the open weight model so much more than maybe American companies? I think it's mainly distribution. And if I was like China's leadership, I would be very open to having Chinese labs essentially be subsidized to drop these model weights out into the market and slow down some of the revenue and growth of the U.S. labs.

So there's kind of like probably some geopolitical strategy going on as well at the national layer. But for the labs themselves, I think it's predominantly distribution. And the price competition on this side has been extremely intense. This, like the past few months, actually, we're starting to see a little bit of a sign of them testing out if they actually have pricing power. So you have, for example, DeepSeek, like raising some of their pricing at the lower end. You had Moonshot pricing K3, the new community model, around, like Anthropics Sonet level pricing. And then like they're experimenting around the licensing itself.

So for example, they're doing things like if you're a model service provider, like an open router, for example, or one of these neoclods, you and you generate X amount of revenue, then you need to direct license with us. So they're definitely experimenting around this side. But yeah, it's hard to say, like to your point earlier, how much can you raise prices and be at 80% of the frontier without a consumer just wanting to switch back to the frontier, right? So I think they're in a tough spot here on monetization, but yeah, the open weights themselves is just primarily, I think, a distribution mechanism for them.

Today's episode is brought to you by Arch Public. Arch Public has just expanded its agentic trading platform beyond crypto, so pay attention. This is a big one. Now they are automating strategies across stocks, commodities and ETFs, and I think that this is going to be huge. You can now automatically take profits when one market hits new all-time highs and rotate that capital into other markets showing more opportunity. Whether you're rotating capital into AI stocks, gold, if you're investing in the S&P 500, or you're accumulating Bitcoin, Archpublic brings real discipline and automation to your investment strategy. Additionally, they've launched a powerful new tax loss harvesting tool.

With crypto being so volatile and its exemption from the wash sale rule, arch public can offset gains with losses without, compromising your long-term positions. It's exactly what every serious investor does. Institutional-grade automation that works across every major asset class. There's no more emotional trading, no more missing tax opportunities, just smarter, hands-free execution of your preferred strategies. Go to archpublic.com right now.

Connect with their team, set up a time, bring your account if you'd like, and then you can learn what automated trading can do for you. Archpublic.com. Now, we have seen both Meta and GROC come out with, you know, kind of price sensitive or price-aware messaging. Does that mean that the closed-sourced American models like the Open AIs, the Anthropics, it's almost like they're getting attacked from two different angles. They're getting attacked from China, but they're also now getting attacked from maybe the competitors people thought were behind.

And these guys are not only releasing new models and trying to catch out from a technical standpoint, but they're also trying to undercut on pricing. As much as I would think, like, okay, that puts price pressure on Anthropic and Open AI. Anthropic just announced like they're increasing prices on certain models and tokens. And so it almost feels like they're like immune or they don't care and they think they've got the position to do that. Yeah. Yeah, it's definitely two pronged for the US labs, like the major US labs. You have Nvidia coming out with Nematron 3, which is like super cheap and like co-designed specifically for the hardware.

which is kind of what the Chinese labs were doing as well. Google has Gemma 4, Meta came out with Muse Spark. So there's definitely some pressure here. I think one of the big caveats here, though, is that US open source models are significantly smaller than what the Chinese model releases are. So like Kimmy and them, they get up into the trillions of parameters where the US labs are pretty much under 130 billion parameters,

with the exception of a couple like Nevatron and Inclan, which gets a little bit higher, but it's definitely like we have no U.S. labs currently shipping frontier adjacent open weight models today like we see in China. And why is that? Like is that just a training dataset, you know, kind of limitation? Is that a capital constraint? Like, what would prevent the U.S. labs who want to go this route from just making bigger models? Yeah. I mean, for the for the major labs, like Anthropic and Open AI, they're generating a significant amount of revenue from the API and subscription services that they offer.

So there's not really a benefit for them to go down into open source outside of just like some narrative positioning. For the other providers like the MetaZ and XAIs and in videos of the world, like they have a benefit of doing this on the open source side. But yeah, I'm not entirely sure why they haven't pursued the strategy, like more so today. Because like we said, it's a really good distribution technique. It's really just a, it's like in crypto classically, you know, you open source a lot of the stuff and it becomes a little bit harder to figure out the monetization on top. So like companies like Nvidia with acquiring hugging face and releasing an open source model, now it's like, you know, use Nvidia stack underneath.

You kind of like bake that in. So there's there's a model there for them, but some of the smaller independent labs, they still have to figure out some type of monetization. That's not so directly tied to easy revenues like that. If you are in charge of Open A or Anthropic, how would you respond to, you know, the multitude of threats and kind of the siege that is underway from all these different competitors? Yeah. You know, if I was them, there's like, I wouldn't be so much worried about like the smaller open weight models that are shipping.

I think some of this story is interesting, but, you know, it's still like at the end the day, like 90-some percent of tokens are being done through anthropic, open AI, Google models. So they're not, it's a little bit overstated some of the pressure that's coming from the Chinese side of things. But if I was then, one of the things I would strongly consider is actually trying to work more closely with the US government's investment like grid capacity and things like this. Because we run out of like grid and energy long before we run out of the hardware side of

things. And so I would just continue to. really push on the frontier side, maybe even consider doing things like taking a percentage of revenues from like IPs and discoveries that are made with their frontier models, specifically in like scientific research. But yeah, I wouldn't be too parsed about the smaller open weight models today. That could change quickly, you know, if meta actually reaches like frontier at J.CENC. But as of now, I think they're in a relatively good spot still. How does this change when we start thinking about physical AI? You know, it seems

seems like maybe Tesla as an example is so far out and ahead of everyone when it comes to self-driving cars and its ability to not only manufacture the car, but also with the models and full self-driving. I've got to imagine that there are thousands of use cases for the crossover or the implementation of AI into the physical economy. And a lot of what we are talking about right now, and I think most people in the industry are talking about is like the software AI industry. Is any of this change when it comes to physical AI? Is China ahead or behind or is it pretty much the exact same thing, just whether people

want to take these models and put them into hardware? Yeah. I've been looking at the humanoid side of things a lot recently. So on the human side, China is also like absolutely trouncing the US, to be honest. So I think it's like 97% of humanoid shipments were from China versus the US. They have Chinese factories now producing it. a humanoid robot roughly every 30 minutes. They have the market leaders like Unitary, UBTech, AGIBOT.

All of these ones are in China. And they're able to kind of use a lot of their commodity and industrial pipelines to make these like really low costs. But I would say with a caveat that we do have some interests in the U.S. on pursuing more of these strategy. I think Elon Musk and Tesla, what they're doing is particularly interesting since acquiring or like merging together with XAI. Now you have the software brain of the robot tied directly with the production and manufacturing in the robot.

And I think we'll see more of this. Even in China, they have a deep seek took like a couple percentage point investment into Unitree's IPO. So I think there's probably a space where we have more. more of this collaboration being done between the labs and the actual hardware companies themselves. One of the things I find most interesting is like in these benchmark tests, there's all kinds of different things people try to get them to do. There's some that are created specifically for AI. There's other things like, you know, Play Go, the game, and at what point can you be the human

or whatever? Financial markets is the greatest game on Earth. You have the smartest people, you've got the most technology. You've obviously had this, you know, very clear economic reward for being able to solve the market's puzzle. At what point should we expect AI to really take over? And is that kind of the end state? Is that the AI bots are just making money for everybody and they're all just sitting on their couch? That's the dream. That's the dream. I mean, yeah, it'll be really cool to see if we do get pretty significant gains in AI as being good at managing a lot of the financial layer. Right now, I think it's still challenging in these types of financial environments to be successful.

It's not so much like hallucinations nowadays, but it's more so around like consensus. LLMs aren't necessarily trained to be good at like out of distribution thinking, which is like what makes you successful in the financial markets. But I feel like with a sufficient harness on top of a on top of a model, you could probably get a long way there, especially I'm sure you saw the Open AI aster release yesterday had beat the RKGI 3 benchmark, which means we have some degree of fluid intelligence inside of a harness, which would make it good potentially at doing more of the financial management layer, but TBD on it. Yeah, it is pretty interesting. We have some engineers that are building the CFO-Silvia product we have,

and they love using Sylvia to try to find different investment opportunities. And, you know, you're always kind of like bias because you're like, hey, our product is amazing, right? But when I start seeing this, I'm like, wait a second, of course people are going to be using these tools to trade on a daily basis, but the human is still in the loop. At some point, people will take the human out and just let it ride, just autonomously, like, go make me money. And maybe that is like the Tesla fleets of cybercabs. You know, same thing. Like, hey, just let it ride, right? Go make me money.

Bots in the market, go make me money. And I don't know. Like, bleed is a human emotion for a reason. Yeah. I imagine a lot of people probably use AI under the hood for their workflows nowadays when it comes to training. So, yeah, yeah, it's going to be crazy. I'm very excited for my AI to make money at scale. Let's talk a little about the geopolitics of all of this.

Obviously, China, you know, they have kind of an economic interest, but they also have the geopolitical interest. Is this something that the United States government should get involved with or prevent or like, how do you look at, you know, the way the U.S. maybe should respond to the current. pieces on the chess board, if you will. Yeah. So there's a lot of interesting angles here. For instance, we started out the conversation talking about the export control side of things. Since the Trump administration's come in, we've reduced our export controls. But interestingly, China is not taking our shipments of H-200s now.

They are essentially pushing a lot of their companies and data centers to move off of any U.S. or global stack and back onto domestic chips. From the US side, I would try, I would say that it would be most efficient to try to get Chinese labs back onto Nvidia to slow down any kind of sovereignty push. Because once you have the ability to pre-train a frontier adjacent model domestically in China, then we actually have an AI race. I think the AI race isn't quite there today. It's more a playing fast catch-up.

once they're able to actually pre-trained, then it's a whole different story. And so we would prefer, I think, U.S. companies to be kind of the infrastructure for that so that we have some degree of control over that relationship. The other thing I would consider from the U.S. side is how we position like power and grid investments. I think we need to treat our grid investments similar to how we treat chips. Insofar as like China produces 2x the electricity, of us. They're adding grid capacity six times faster than us. Their AI energy usage demand is like one to five percent of the capacity they've added in the past five years, whereas in the U.S. is like 50 to 70 percent of the capacity we've added. So yeah, we like I said earlier, we run

out a chip or we run out of grids and we have to lead before we run out of chips. And for that reason, it's like a strategic investment that we should be making that China's already doing. Today's episode is brought to you by Simple Mining. Bitcoin Mining has a reputation for being complicated, risky, and hard to evaluate as a real investment. If you're considering mining in 2026, what actually matters is in headline profitability. It's uptime, repairs,

and whether the operation is run like a real business. That's why I've been using Simple Mining. They're based in Cedar Falls, Iowa, and they run a white glove hosting operation where you own your miners. You choose your own pool, and you have Bitcoin sent directly to your wallet. They were featured on the Inc. 5,000,

as the fastest growing company in Iowa with over 40,000 machines under management. What stands out to me is execution. They have the number one rated ASIC repair center, and for the first 12 months, repairs are included. If mining margins get tight, you can pause with no penalties. And if you want to resize or upgrade your fleet, there's a marketplace to resell equipment instead of being stuck. To help people think it through whether mining actually makes sense right now, they put

together a short resource called the 2026 Bitcoin Mining Blueprint. It walks through the five mistakes investors make when Alexixt. to mining, and they also explain how to avoid them before deploying capital. If it sounds interesting to you, you can get it for free at simplemining.io slash pomp. That's simplemining.io slash pomp. Go check it out today and see if you should get into the mining game. What are you most concerned about when it comes to all of this? Like, do you have specific concerns?

Are you just more observing, you know, kind of what's happening? I've been more observational. I mean, as a U.S. citizen, I want us to stay in the frontier. And I do think, like I said, we have a huge capability lead right now. I worry when I look at like politics, specifically like Bernie Sanders letter that came out yesterday in some of these like slowdown efforts. Any type of slowdown in my mind benefits China because it gives them more time to build out the hardware infrastructure they need to actually compete.

So I worry about this insofar as like I want the U.S. to be the world. winner of this race. And the only way we do that is like push forward, make more investments, push harder than what we're currently doing. I worry that we reach point of stagnation and allows for these other countries that come catch up. It does feel like we're reaching this point where everything can kind of take off from here. And so, you know, you move from kind of wet cement to the cement being hardened and we may not be able to kind of go back. We may not be able to make some of these changes. But I do think competition is really, really effective. And it's maybe why I'm so interested in, like, there's new chip manufacturers that are trying to come to market. There's

obviously a lot of people who are making a big bet on specialized workflows over general intelligence, but maybe the market is just bigger than we all think it is. And there's going to be so many different winners and so many different applications of this technology that people have to get rid of the, like, black and white, you know, or kind of single winner viewpoint and understand that you're serving all of humanity, right? Both software and hardware, like it's a pretty damn big market. Yeah. And, you know, the benefits of AI need to be diffuse. So if it's multiple global hubs that are pushing forward on the frontier, I think that's better for everyone. It's just, if you have any of those concerns about runaway AI or anything like that, then obviously

you want one direct leader probably in the West to lead forth. But assuming that scenario doesn't happen is more about just like proliferation AI. I think that's good for everybody. Just to get the the Europeans all worked up, I don't hear anything about European AI. I hear about the United States and I hear about China. Is there anything going on in Europe or anywhere else that you've been paying attention to? So I mean, they have some companies like Mistral. Mistral is hot for a while on the open weight side. But yeah, I mean, they're just too much of like a regulatory hub, I feel like to really, really innovate at the speed that Washington and Beijing are doing.

Yeah, it's just the government. They got their finger on the scale over there. By the way, all my European brothers and sisters, they don't like it, right? They wish it was not that way, but such as life. All right. And then the other thing, I guess, is as we're watching all of this play out, you guys are continuing to pump out a lot of great research. I mean, I've been involved with Delphi for a long time.

And maybe you can talk a little bit about the research you guys are doing and where people can go read that. Yeah, yeah, for sure. So Delphi Digital. I host all of our research. We cover a wide spectrum. We started predominantly as a crypto research firm. We've been branching out into other areas like AI, robotics, drones, NeuroTech. So we're trying to be leaders in the research for deep tech generally.

So yeah, check out our reports. I think you guys have done a fantastic job. Nobody really gives you guys the credit you deserve because they don't know about the Delphi Ventures or the research, any of this stuff. You guys have crushed it. So keep up the great work and we'll, you know, pump out a couple more good reports. Teach me something on the reports and then you'll come back and we'll do it again. Let's do it.

Let's do it. All right. Thanks, Justin. I appreciate it. We'll talk soon. Awesome.