The AI Power Podcast
Understand everything that's going on in AI Policy and how AI impacts the world. Hosted by Gregory C. Allen.
If you work in AI policy — or you're just fascinated by it — this is the podcast for you. Every week, The AI Power Podcast unpacks the developments that actually mattered: AI regulation, safety, economic policy, US–China competition, semiconductor export controls, and national security. Think of it as drinks after work with the smart friend who tells you what's really going on, and what might actually work, in plain English.
Plus interview episodes — long-form conversations with the policymakers, builders, executives, and analysts shaping artificial intelligence and the global power competition built around it.
The AI Power Podcast
Licensing AI, the Memory Crunch, H200s to China & Is It a Bubble? — with Adam Goodwin
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
In the span of eighteen days this summer, the Commerce Department gave Anthropic ninety minutes to pull its two most powerful models offline — and then quietly let them back on. How did we end up with a de facto AI licensing regime without anyone actually passing one? For the first episode of The AI Power Podcast, Greg sits down with Adam Goodwin — former Senate staffer, Greg's old colleague from Blue Origin, and now president of Goodwin Strategies — for a wide-ranging tour of the AI policy stories that actually matter right now. They dig into the Mythos and Fable shutdown and Project Glasswing; whether OpenAI's cooperation (and rumored 5% government stake) means Washington is licensing AI in all but name; the DRAM shortage now pushing up the price of your next iPhone; China's memory champions and the fight over letting Apple buy CXMT chips; why Beijing suddenly wants those NVIDIA H200s after all; the morale crisis inside the Bureau of Industry and Security; the trillion-dollar bull-and- bear case on an AI bubble; how Chinese distillation of American models threatens the whole business model; and the coming "SaaSpocalypse."
Chapters:
00:00 – Meet Adam Goodwin
1:18 – 90 minutes to shut down: Mythos, Fable & Project Glasswing
06:33 – Is Washington licensing AI in all but name?
11:03 – The DRAM shortage and the $200 price hike
17:04 – China's memory champions: YMTC & CXMT
28:03 – Why Beijing reversed course on the H200
38:21 – Inside a struggling Bureau of Industry and Security
42:08 – The trillion-dollar question: is AI a bubble?
52:06 – Distillation, DeepSeek & the copying arms race
1:03:56 – The "SaaSpocalypse"
See Greg's full research archive, and the Decision Tree newsletter at decisiontreeresearch.com.
Welcome back to the AI Power Podcast. I'm Greg Allen, and today we have a treat. My friend and former colleague, Adam Goodwin, is with us to talk about all the big topics in AI policy that have been going on over the past few weeks. For those of you who don't know him, Adam Goodwin is a former congressional staffer in the Senate. He is my former colleague when we both worked at Blue Origin, the Jeff Bezos Space Company. And then he went into the private sector on the government relations side. He is now the president of Goodwin Strategies and has been a deep insider of Washington technology debates and all manner of debates for more than a decade now. So, Adam Goodwin, thanks so much for joining me on the AI Power Podcast.
SPEAKER_02Thanks for having me, Greg. I'm really excited to be here. And more than anything, really excited to have a forum where I can ask you questions about AI because there's a lot of confusing information out there that I would love to have you clarify.
SPEAKER_00And likewise, in terms of the DC behind the scenes that you are privy to that we can uh share with the audience. So, Adam, what are we talking about first?
SPEAKER_02So a lot of things going on in the world, but I think that the biggest one, the one that's on everybody's mind right now, on June 12th, the Commerce Department gave Anthropic 90 minutes to shut down its two most powerful AI models. And by June 30th, the ban was gone. Why did this happen and how did it happen? Let's start there.
SPEAKER_00Yeah, I think this is uh the sort of table stakes for understanding what is going on in the AI policy debate right now. And what's so interesting is you can kind of look at it and see so many different actors in this story reversing their position to a greater or lesser extent, right? You have the Trump administration, which starts off by saying that the Biden administration's AI regulations were choking the private sector, that they were going to let the private sector cook, and that they were going to have a light touch regulatory approach. Then you have anthropic saying that this technology is profoundly dangerous and that we need to have safety as an ultimate priority, and that that will require government regulation. Sam Altman testified the same thing before Congress several years ago, back in 2023. So you've got the government, the Trump government being anti-regulation, you've got the private sector to a certain extent being pro-regulation. Now, where are we? Well, Anthropic went to the US government about its mythos model and said, this is a game changer. This is so powerful that it actually is a meaningful step change in cybersecurity capabilities. And some of the stuff that they showed the government with mythos, at least according to Senator Mark Warner in the Senate hearings on this topic, was that it broke into some of the national security agencies' most sensitive intelligence servers in a matter of hours. Now, the New York Times has since done some reporting that claims that Warner's version of the story is a bit overstated and it was actually more like it had discovered vulnerabilities in NSA databases and critical things. But essentially, every major web browser, every major operating system, some systems that were viewed as some of the most secure in the world, Mythos was able to find and autonomously exploit in some cases these vulnerabilities. So, as we said, a game-changing cyber capability. Well, when Mythos and Anthropic went to the government, the government decided, yes, indeed, this is too dangerous to release the same way that AI companies have been releasing their commercial models. We need to put it behind special guardrails. And Anthropic instituted Project Glasswing, which in partnership with the government was to distribute Mythos and its cyber defense capabilities to critical infrastructure providers, banks, and other sort of key nodes in the US economy and cybersecurity infrastructure, and the same for allies in their cybersecurity infrastructure, and do that first. So that's really interesting, right? You have the Trump government acknowledging that some kind of regulation is needed. They passed an executive order for what this is going to look like on a voluntary basis. But then Anthropic releases Fable, which is a version of Mythos with additional guardrails that Anthropic put in there. And after reportedly an Amazon executive told the US government that the guardrails that had been installed on Fable were not enough, and that this was actually bringing a new degree of capability on the cyber offense side of the ledger, the United States government and Howard Lutnick gave Anthropic a call and used export control authorities to say that they're not allowed to distribute it to any foreign nationals, which means they basically can't even distribute it inside the United States. Anthropic shuts it down across the entire world. No mythos, no fable. And that is extraordinary, right? We now have a de facto licensing regime. So fast forward to today, and now Anthropic is being allowed to distribute Fable, which I gotta say, like as a user of Fable, it is a jaw-dropping degree of capability, especially in coding. If you haven't used it yet, like you really need to test this out, you will be astonished. I I play a lot of computer games, and I have been astonished that it is able to make like decent approximations of games that I really like, sometimes in one prompt or less, where you just sort of say, uh, create a version of Zelda. And it will not create exactly it, but something that is clearly a huge amount of labor, if humans were involved in creating it from scratch, would be an extraordinary lift. And uh Fable, which again is the nerfed version of Mythos, is able to do this. So a real whipsaw and Anthropic, uh ironically, which had been saying that guardrails are needed, a licensing regime might be needed, is the one who's saying the government has gone too far in these circumstances. Everybody's changing their position. It's kind of just a crazy moment.
SPEAKER_02Yeah, that's a kind of a very interesting model of uh be careful what you ask for. And inside of 18 days, all those things can happen to you. In interesting comparison, Greg, uh OpenAI's Chat GPT uh 5.6 went through the same government gate, but did so cooperatively, right? We can say a little bit more cooperatively. And now OpenAI is discussing giving the government a 5% stake of the company, I believe, if I'm not saying that wrong. Um, is Washington licensing AI in all but name now? We've seen this in a couple of other industries. There's been a lot of conversation about this, more on the national security side, but how how is this shaping up?
SPEAKER_00Yeah, I think this is a de facto licensing regime. And especially once you take into account the export controls side of the equation, it's an explicit licensing regime, right? When you say something is prohibited because of export controls, what you mean is it's illegal to sell it without a license. So this is not a model licensing regime of the kind that the Biden administration contemplated at various stages, but never actually implemented. It's also not the voluntary only regime that the Trump administration stated in an executive order that they wanted. And the reason why it's not voluntary is on the one hand, open AI is going to the government and the government saying, no, you can't release this uh yet. And so you could say that's cooperative, but the subtext of that whole thing is we can do that thing we just did to anthropic to you, right? If you don't do what we want voluntarily, we will just compel you to do it. And and Sam Altman's public statements made it very clear that even though they're they're acceding to the government's requests, um, they're not happy about the way that it went down. And they they feel like this whole sort of prior government review architecture is just not working. So I think there's two quotes here that are very telling. One is from uh Dean Ball, who used to be a senior advisor on AI at the Trump White House. Now he just actually joined OpenAI. But here's what he said on X quote, AI is licensed now, but the requirements change constantly and are always a secret, even to the administration itself. And then Microsoft president Brad Smith, um, he said on the record to fortune, quote, what we really have right now is regulation without transparent or complete rules. Without rules, businesses can't plan. And I think that's kind of where we are right now is on the one hand, everybody agrees that the rules should exist. Um, maybe not everybody, right? Maybe not David Sachs. Um, but nobody can agree on what those rules should be. And we're marching towards, right, the government's self-imposed 60-day deadline to come up with a voluntary prior review framework. What's going to include, what that's going to include, what it's going to look like on the unclassified side, what it's going to look like on the classified side, all of that, very uncertain right now. Um, and here's one other thing. Right now, the entire focus of the conversation is on cyber because that's where the capabilities are most mature, not super surprising. If you're really good at coding, you're probably going to be really good at cyber in the not too distant future. But the the one that's really coming around the corner is on the bio part of the story and whether or not, in the same way, this can uplift cyber offense or cyber defense capabilities. Could we see some uplift in biooffense capabilities? Which, if that's true, you definitely don't want AI helping anybody create the next pandemic. How do you create guardrails that can allow hundreds of millions or billions of global users to enjoy this extremely capable software while at the same time ensuring that not even one bad person uses it to create bioweapons? That's a really, really tough needle to thread.
SPEAKER_01Yeah, certainly. Certainly.
SPEAKER_02So uh shifting gears slightly, because what what we're talking about with anthropic and with open AI is a little bit more regulatory, sometimes theoretical. It doesn't necessarily impact people in their day-to-day lives as much right now.
SPEAKER_00It impacted me when I lost access to Fable Five. That was a real pain in the butt.
SPEAKER_02Well, this is fair. This is fair. When you have a tool that you're actively using and then it's just discontinued, that that impacts you. So yeah, one of the uh one of the exceptions would be you, Greg. Uh so point taken. Um, I think that one of the places that uh we tend to see people pay more attention more quickly is where the economics hit them directly and immediately. So uh there are a couple of corners of uh the AI world, and the infrastructure build out is one where uh because the infrastructure build out for these data farms and facilities is so large and so expansive that it actually has the ability to compete with uh other economic sectors, other parts of the market that just aren't AI involved at all. Uh, we're beginning to see this with memory chips, uh, specifically the dynamic random access memory, so uh DRAM. That is becoming slowly at first, rapidly now, one of the hottest conversations out there, both economically and politically. Lay out the basics for the rest of us on that particular issue. What does this look like right now and what are the implications moving forward?
SPEAKER_00Yeah, so this is a really cross-cutting topic, and it's for a very simple reason. On the let's start with the AI part of the story. The big AI companies, the big cloud hyperscalers are spending hundreds of billions of dollars this year on an AI infrastructure build out, right? They're they're doing the equivalent of what the Apollo Moon program spent over its entire 13-year lifespan. They're spending that same amount of money every 10 months. And actually, it may be even shorter than that, given the acceleration of spending that we've seen just in the past few months. So an extraordinary build out of data centers. Well, what's in those AI data centers? About half the cost of the data centers is the chips. And when people think AI chips, they think NVIDIA GPUs, they think Google TPUs, so on and so forth. That's about half the cost, right, of those data centers. Well, what's half the cost of those AI chips? It's actually memory chips. About half the bill of materials of an NVIDIA GPU is actually the high bandwidth memory inside it. And high bandwidth memory is essentially vertically stacked, advanced packaged DRAM. So DRAM is a commodity. It's in everything. It's in your laptop, it's in your smartphone, it's in your tablet, it's in data centers, it's everywhere across the economy. If it's using computation, there's a good chance it's using some kind of DRAM. And there's a big difference between the memory chip market and the logic chip market, which is in the logic chip market, old chips can still be useful, right? If you design the Boeing 737 variant in the late 1990s, well, if you build one next year, you might still want that same original logic chip because that's the one that went through FAA certification. Logic chips are sometimes a little bit more specialized, and the old stuff can still compete in the marketplace in various niches. Memory, by contrast, is much closer to a commodity. And because the chips are always getting better, the latest and greatest memory chips are frequently cheaper and higher performing than the old chips. So, because it is a commodity, there's a direct trade-off between that factory producing high bandwidth memory versus producing DRAM memory. If they want to make more high bandwidth memory, they can allocate more of the factory capacity to high bandwidth memory and less to DRAM. So even though the original shortage is high bandwidth memory, what that implies is also a shortage of DRAM. And Apple, very notably, very loudly and prominently, just increased the prices of a bunch of their devices, iPads, laptops, MacBooks, et cetera, by like $200. And they're solely attributing it to the DRAM shortage, which is coming from the AI build-out. And so we've talked about where the AI buildout can intersect affordability. Most of that conversation over the past year has focused on electricity and electricity bills. Now you're seeing it at this components level. Certainly the iPhone release every year gets a lot of attention, right? When the new iPhone comes out, presumably in September, if there's a $200 price increase, that's an easy thing to make a story on. That's an easy thing to make people talk about. And it is coming from this AI infrastructure build-out.
SPEAKER_02I mean, just to be specific, we're talking about a $200 price increase on units that are normally $1,200, $1,300. So when you're a little bit more than 10% of the actual baseline cost, that's significant. I mean, that's very significant.
SPEAKER_00It's huge. And here's the thing memory around the world, there's historically been three companies that dominate the sector. Two are Korean and one is American. So SK Heinex and Samsung of Korea are memory giants. And then Micron of the United States, uh, they're mostly based out of New York, is the other giant. So Micron's had a great year. I think their stock price over the past 12 months is up something like 700%. It was the best performing stock in the entire stock exchange uh last year. That's all because of this memory crunch, which is kind of amazing because a few years ago, before the AI boom, the post-Chat GBT AI boom, people were really worried about Micron and whether or not they were going to be able to continue competing against Samsung and SK Heinex.
SPEAKER_02So in that environment, and China is not going to take all of that line down, certainly with kind of limits to access in general. So what does Chinese industry look like in this context within this space?
SPEAKER_00Yeah, and this comes up. So as I said, in memory, it's harder to compete with older technology. In logic, it's easier to compete with older technology. So as China climbs the ladder of technical sophistication, it's easier for them to climb each rung on the logic side. Whereas in memory, it's hard to compete until you get to the very top. But actually, China has two domestic memory champions, um, both of which are interesting in their own ways. One is YMTC, which has historically been more focused on NAND flash memory. That's what goes into a flash drive, like a USB thumb drive or a laptop hard drive. And YMTC at one point was under consideration by Apple for using their memory in iPhones. YMTC uh legitimately has some interesting technologies, like some of their patents are genuine innovations in the space. They are not chumps. But they have been hamstrung by the entity list, which restricted their ability to buy many categories of advanced semiconductor manufacturing equipment. And that set YMTC back years. And when Apple was talking about buying YMTC memory, Senator Ted Cruz got furious at Apple and made a lot of public statements about how that was an unacceptable choice. Well, CXMT is China's champion in DRAM and HBM memory. And CXMT has actually been allowed to buy a lot of advanced semiconductor manufacturing equipment. Not the equipment that is banned on a countrywide basis to China. So no EUV machines, for example. But as an entity, you know, CXMT was not entity listed in the way that YMTC was. And so as a result, CXMT has been buying an extraordinary amount of semiconductor manufacturing equipment from the United States companies, from Dutch companies, from Japanese companies. And they are growing in sophistication, especially in DRAM. And they obviously have a technical roadmap to try and be competitive in HBM. So now bringing it back to Apple, well, Apple has a long history of trying to identify where are their choke points in its supply chain and how can they use China to introduce new entrants into the market and make there be more competition so that they can drive down the profit margins of their suppliers, drive down their costs, and increase their own profits. You've seen this in display technology, where Apple was instrumental in uh helping Chinese companies ramp up and be more competitive against Korean companies. And now here it is coming again, potentially in the memory space. So, according to the best reporting on the topic, Apple is trying to get assurances from the US government that if they start buying DRAM from CXMT, that they're not going to have a repeat of the YMTC story. That essentially CXMT would be entity listed shortly after Apple shows an interest in it.
SPEAKER_01And so exactly.
SPEAKER_00So from Apple's perspective, they're saying, look, DRAM, there's a shortage worldwide. This is a commodity. Why does it matter if China makes this? Just let us buy it from them, and you won't have bad headlines, Trump administration, about growing prices in the iPhone.
SPEAKER_02Yeah, and just as a matter of perspective, how many iPhones are out there right now in the world?
SPEAKER_00So Apple's market share by units, so like by count of active devices sold, is around 19 to 20% of the world, but their share of revenue and profit is much, much higher, right? Like they make 50% of the revenue in the whole global industry. They make 80% of the profit in the whole industry. So it's a huge, huge driver of American competitiveness, America's tech industry. You can understand why they want low prices, but I still don't think this is in America's national interest.
SPEAKER_02One of the points of contention that you can see coming, especially in a congressional debate on this particular issue, is There's a direct line to treat CXMT the same way that other Chinese companies in the space have been treated. But when you're talking about just the percentage increase in cost of ubiquitous technology, kind of like what Apple has, that's very difficult to argue for when you're in the middle of an affordability crisis, which is really what's confronting us right now, especially leading up to the midterms. So I think it becomes a contentious space to be able to try to argue that uh as we're moving through the year.
SPEAKER_00Um Yeah. So I I think Apple is they're poking at issues that, as you said, like have political resonance. Let me just lay out why I think this would be a terrible idea from a national security and an economic security perspective. The first thing to note, and my friend Jimmy Goodrich has made this point uh publicly, is that the memory industry has historically been a boom and bust type industry. Essentially, it takes like two years to build one of these fabs to make new memory chips. So you have a shortage because nobody wants to build fabs. And then when there's a shortage, a lot of people want to build fabs, and then two years later, all of these new fabs come online and suddenly there's a glut of supply. So the price crashes, and then nobody wants to build fabs for two years, then there's a shortage, then there's a boom, then boom, bust, boom, bust. That's the history of the memory industry in so many words. And what's different is that the Chinese companies are viewed as national champions and have enormous domestic subsidies. And so what that means is if you let these companies enjoy the boom times, when the bust comes, the government is gonna make sure that they grow and gain market share and they will have a number of unfair advantages over all the other companies when the bust inevitably comes, as it always does in the memory industry. The second thing is, I do not think there is a way to buy Chinese memory chips that is not a de facto tech transfer policy, right? As I said, it's really hard to compete with memory chips that are worse than the state of the art. So, what that means is it's going to be in Apple's interest to help Chinese memory companies like CXMT produce at or close to the state of the art. And that is not going to stay contained magically in China's memory chip sector. That is going to transfer over to their logic chip sector, which is going to strengthen their ability to compete with NVIDIA in the GPU market, which, as we all know, is very strategically relevant for the overall AI race, which this Trump administration has compared to the space race. And who else thinks that this is a good comparison to the space race or the nuclear arms race of the Cold War? Chinese cyber companies. Um, the CEO of 360 Security Technology, which is arguably like the top Chinese cybersecurity company, you can think of them as analogous to a Palo Alto Networks or a Mandiant or one of those types of firms. He said that mythos was a cyber nuclear weapon and that it was extremely concerning that the United States had monopolistic access to this technology and didn't give it to China. Well, why China doesn't have its own cyber mythos like right now? It's because in December of 2024, the Biden administration restricted China's access to high bandwidth memory. And CXMT has a history of basically telling their suppliers on the equipment side of the equation, the semiconductor equipment side, hey, you just help us make these friendly DRAM chips, and that's all we're gonna use your equipment for. Meanwhile, they go present at these conferences around the world and say, oh yeah, we're absolutely trying to make HBM, we're absolutely trying to break into the AI memory chip market. Here's our technological roadmap. And so the basic point is if you let Apple buy chips from this company, they're gonna do what Apple has done for every other tech company of China, which is try and help them make world-class products at the at the scale and at the quality required to be relevant to an iPhone, which is usually the most sophisticated integrated product in the world when it comes out. And that skill diffusion, that technology diffusion, that equipment diffusion, it is not going to stay constrained in memory. It's not even it's not even going to stay constrained within DRAM. And so the point here is that this would be a grave strategic error. You are not just shipping a lot of money that's going to help China's uh semiconductor sector and AI sector, you're shipping a lot of know-how and technological capex that's going to help uh China's sector. I mean, I'm for going in the other direction, right? Rather than, hey, how can we help CXMT more than we are now? I think we should help you helping CXMT less than we are now. I would love to tighten the export controls. Like there are export controls that affect CXMT. You're not allowed to sell through silicon via uh semiconductor manufacturing equipment, which is super important in trying to make HBM uh, for example. But this this is this just doesn't make any sense, and we definitely should not do it. Now, does that stink? Does that stink uh because we have this uh memory shortage? Yes. But like again, memory is a boom and bust industry. This is not going to last forever. In you know, two years, it's totally plausible that we'll be talking about the memory glut. And when we are talking about the memory glut and Micron is going to be laying off people in New York, do you want them to be laying off people even further because CXMT is taking all of their jobs? I think that's just a huge unforced error that we don't need to make.
SPEAKER_02No, that's a that's a very fair point, Greg. And I want to kind of pull the thread uh as a as a segue here because you spoke a little bit about NVIDIA and some of the policy implications uh of the regulatory actions that we've taken on uh H-200s in particular, right? Uh so recently Beijing has decided to reverse course and allow the purchase of H-200s uh into China. Uh, can you talk about how we got there and obviously a little bit of background on what that regulatory framework had been uh kind of prior to this decision would certainly be helpful.
SPEAKER_00Yeah, so we started off by talking about how the US government and the US companies keep going back and forth on their position around AI regulation. It's kind of the same between the US government and the Chinese government on allowing exports uh or imports, if you're China, of AI chips to China. Um, first you have the US government doesn't want to allow it. Then, when the Trump administration reverses course and says we want to allow some chips, then China is saying we don't want to allow the purchase of these chips. Um, here's what I think is going on. The first is it's helpful to not view China as a monolith in this story. There are different factions within China that reflect the different interests of people who are at different roles in the AI value chain. If you are Huawei and you want to design your own chips that are competing with NVIDIA, if you are SMIC and you want to manufacture Huawei's chips so that you can compete with TSMC of Taiwan, you've loved the NVIDIA ban from day one, right? Like that is awesome. And in fact, here's the thing: even before America banned the export of NVIDIA chips, Huawei was benefiting from China's plans to prohibit their purchase. This is the 3-2-1 plan that China had in 2020, which basically says all you know, Chinese government ministries, all state enterprises, you only have three years to get off of American semiconductor technology. So even before the October 2022 export controls, China is saying we need a pathway to not buying NVIDIA chips anymore. Who was the obvious choice to make those chips as an alternative? It was Huawei, right, which was introducing AI GPUs. Um, in fact, they were they were on track to manufacture at TSMC at the 7 nanometer node, even before NVIDIA. So uh that's one faction within China. The other faction is the cloud hyperscalers, and they uh are profit-oriented, they compete viciously with each other. And so what they want is the best chips at the best price. And for a long time, that has meant NVIDIA chips from their perspective. And so even as um even as these export controls went into effect, they were deeply, deeply concerned. And when the Chinese government said that they didn't want to allow purchases, they opposed that internally to the extent that you can oppose a Chinese policy. And so what you're hearing now is that they are saying that the the downsides of not having access to NVIDIA GPUs, even last generation NVIDIA GPUs, such as the H200, has been painful enough that they want this to be allowed. Now, let's go back to something that I actually read in an earlier podcast with Chris McGuire. This is from the 10 cent investor uh relations call. So this is um uh a key Chinese tech company talking to their investors. Here's what uh one of their executives said on the record quote, and in terms of your questions about the various bottlenecks between GPU, CPU, networking, and so forth, to recap that the reason why there's been a GPU bottleneck that's been much more pronounced in China than elsewhere is a combination of the policy restrictions on certain foreign design GPUs being brought into China, and then the China design GPUs facing limited fab capacity within China. And as a result, the country has been really short of GPU or ASIC capacity. And that's now being addressed because the China designed ASICs are seeing more supply from fabs within China, as well as more supply from fabs in neighboring countries. Now, here's the thing. Why is China seeing more supply from neighboring countries? I think that that is tragically confirmation of the fears of some that when the Bureau of Industry and Security basically, as I discussed with Chris McGuire on a previous episode of the podcast, when they failed to clarify the regulations in a timely fashion that no, you're not allowed to sell to Chinese subsidiaries outside of China, there is a second failure mode, which is that like the fab uh rule, you're not allowed, you're not allowed to manufacture chips on behalf of Chinese designers or Chinese subsidiaries outside of China chip designers, which basically means Chinese companies getting access to either TSMC manufacturing capacity or to Samsung outsourced manufacturing capacity. That's the the most plausible interpretation of this claim that there is more chipmaking capacity for AI chips that is accessible to Chinese companies outside of China.
SPEAKER_02Well, to be clear, we're not actually even talking about smuggling. This is really just a loophole that's being exploited right now, right?
SPEAKER_00Yeah, yeah. It's a terrible, terrible, terrible loophole. Um, so hopefully the Trump administration is going to close that immediately. And I don't think it's necessarily a coincidence that right around the same time we're closing these two loopholes, China is like, well, maybe we should buy those H200s after all, right? So their whole thing about how we don't need H200s, was that a flex about, you know, the state of internal Chinese chipmaking capacity? No, I think it was a flex about the efficacy of smuggling and the efficacy of loopholes to exploit, you know, gaps in the existing rural architecture. As I often have to say, uh, and as this 10 cent quote makes clear, the export controls have had a significant effect, but boy, has the implementation been a complete mess. And we are only achieving a fraction of what could have been achieved with a much wiser implementation. And now that we're talking about, you know, sending all of these H-200s to China, I think they're gonna buy them. And I think the arrival of a Chinese mythos tier capability can't be too far behind once they start getting unrestricted access to unlimited numbers of chips. I guess I should say though that the Trump administration has suggested at various times that maybe it wouldn't be an unlimited number of chips. Maybe there would be some kind of cap per company, cap per something other. I don't know. Whatever it is, it's higher than zero, which is what the right choice would be.
SPEAKER_02Yeah, and as you mentioned, uh those Chinese companies can make up a lot of ground, specifically with logic chips, by using less capable older chips just in different ways or stacking them together, right?
SPEAKER_00So yeah, let me let me um let me say something on that score because that's a that's a great point. Um there's there's two ways to think about it. Um, one is the older chips are what you need, which is like my 737 example. Like you do you don't just need an old uh old chip process for manufacturing, you need an old chip design because this this design was approved for something or other. Um, or just because it's dirt cheap. Like you're if you're making an um if you're making a uh electronic watch for one dollar, you don't need a $500 fancy chip. You can use a chip from the Stone Age and it'll work just fine. A second way of thinking about it is with each new generation of chips, some share of the performance improvement is because of the improvement in manufacturing technology. So we can put more transistors per per square millimeter on a chip. And then some share of the performance improvement relates to superior design. And so China has been cut off from a lot of the advances in manufacturing technology, but they still have smart chip designers. Um, this is why the Huawei Kieran series uh back in 2023 was able to use older manufacturing nodes, but to have smart chip designs uh that had ballpark comparable performance to newer chips from Qualcomm or whoever you have, who have you when it comes to design. And in the example that I just said, it was because they had their application processor and their modem co-located on the same die, which Apple was not able to do because Apple designed the application processor, Qualcomm designed the modem. They were they were separate chips. So there were sort of sources of performance improvement separate from the manufacturing node uh decision that had to do with the design decisions, and Chinese companies can still exploit that kind of thing. Now, I still think that on an apples to apples basis, Chinese chips are worse in terms of reliability, they're worse in terms of absolute computational performance, and they're worse in terms of their ability to collaborate together as a team, and they're worse on software in terms of like what soft what does the software ecosystem look like uh when we're talking about, you know, using these chips. On all of those things, they're way worse than NVIDIA, they're way worse than NVIDIA's last generation, which is why the Chinese cloud hyperscalers are clamoring for NVIDIA chips and have been for so long.
SPEAKER_01Understand. Understand.
SPEAKER_02Uh, I did want to circle back to something you uh mentioned earlier about the involvement of the Bureau of Industry and Security in this larger export restriction fight. Uh, there was a recent political report that said the BIS, the Bureau of Industry and Security, is having a hard time and kind of grinding to a halt. That that obviously piqued my interest as a government guy. Um, what is what does that look like right now? What do we know?
SPEAKER_00Yeah, the Bureau of Industry and Security has been under-resourced for over a decade, right? Um, it's it's really unfortunate because they they just don't create a lot of jobs. It's not like building a new fighter jet where there's a bunch of congressional districts who are going to lobby to buy more of those fighter jets. The BIS, all those people either work in in the Washington metro area or they work abroad to a first approximation. And so there's not a huge constituency, you know, in favor of strengthening uh the BIS. They had historically been under-resourced. A separate problem has happened under the Trump administration, which has been more willing to contemplate increased resources, at least at various points in its budget request history. Um, but the morale, at least according to this political story, and I'll just say, according to what I've heard from folks inside BIS, morale is at a pretty low point right now. And you can understand why. They're getting whipsawed messages from the very top leadership of the administration. The Trump administration comes in saying we're going to be way tighter on export controls. And then suddenly they're allowing the exports of NVIDIA chips to China. So it's like, what are we, what are we fighting for? What are we trying to accomplish here? Like, what is our work actually for? If you go work at the Bureau of Industry and Security, it's because you really care about US national security. And you've been making an argument for years about how restricting AI chips is a national security matter. And then suddenly you find that this stuff that you've always said for decades, the Bureau of Industry and Security has said that national security is not up for debate in trade negotiations. And then it looks like the Trump administration is very willing to negotiate this in trade debates with China and so on. So a lot of folks have left, even though the budget to hire more people has gone up. There's actually still a lot of unfilled seats. The process for all of this, um, I think it's fair to say, just feels more political than it had in the past when it was more of an administrative, process-driven uh type of agency. And you have these big, big misses by the organization in terms of the fab loophole that we've talked about, in terms of the selling chips to Chinese subsidiaries, loophole we've talked about. And that's that's because this is a super important mission that is just not being taken seriously to the extent that it needs to be. And I think it's really unfortunate. I mean, I've been I've been rooting for this agency to reach its full potential for a long time. And at least according to the reporting uh uh in Politico, it's far, far from that.
SPEAKER_02Yeah, I from personal experience I can tell you that government agencies, especially um outside of the direct national security space, tend to be under-resourced. So the thing that keeps you there is the strength of your mission, commitment to mission. There aren't a lot of people who just stumble and fall into being export lawyers, right? You have to actually care about that stuff. Um so you can deal with being under-resourced, but when the morale starts to plummet and you're really under-resourced as far as people nearby, not just financially, that takes a toll. That does take a toll. Um kind of want to switch gears a little bit and talk about uh some of the larger, uh, some of the larger issues we're looking at out there. Uh, in particular, there are ongoing conversations about reasonable valuations for a lot of kind of frontier AI companies moving forward. Uh, we're we're preparing to see IPOs from anthropic and open AI. Uh, there is that uh devious word out there, uh bubble that is starting to come up. It seems like it comes up kind of every generation with a new set of technology. Um let's examine the two ends of this, right? What's the bull case? What's the bear case? Because it certainly does have a bearing on a lot of the other things that we've talked about, right? So memory infrastructure investment for companies trying to produce it, uh, regulatory efforts for how valuable these things are, right? How much China really wants to compete in these industries, all really is based on are we about to confront a bubble or not? Is is the investment worth it? Is it real? So what are we actually looking at?
SPEAKER_00Well, I I think I think you you asked it the right way, which is like, what is the bull case? So let's start with. I mean, there's there's three IPOs on the AI model vendor side of the equation um that have either happened or or expected to happen in the not too distant future that are sucking up a lot of the oxygen. There is SpaceX, which now, despite the name, includes the XAI subsidiary. So the former Twitter part of the Elon Musk empire that's building gigawatt data centers and trying to train frontier AI models like Grok. That was a part of this IPO, and it was most of the valuation of this IPO, which I think was originally listed at 1.75 and is now worth more than uh. Uh, two trillion dollars as just one firm. So crazy debut on the stock.
SPEAKER_02Other AI companies that have recently been purchased by that entity, right?
SPEAKER_00Yeah, you're talking about like Cursor?
SPEAKER_02Yeah, yeah, yeah, yeah. So yeah, there's the drug that we know of, and then OpenAI. Right, yeah.
SPEAKER_00I mean, they they haven't completed that acquisition, but uh it seems likely to go through if I had to guess. Um so then you have anthropic and open AI, both of whom are expecting to IPO sometime in the next 12 months. And they're also targeting valuations in the ballpark of $1 trillion. So let's start with the bull case uh here. Anthropic alone, their revenue run rate, which a run rate for those who don't know, is like what if you took how much money they made in one particular month and you assume that they made that for the entire year. Their revenue run rate reportedly went from less than $10 billion in December of 2025 to $47 billion in May of 2026. When your revenue is already in the billions, 5x your revenue in five months. That's pretty good if you can make it happen, right? This is the peak of companies basically saying they're gonna rate their software engineers based on how many tokens they consume per month, right? That is an insane pace of growth. And if you believe that in the same way that uh software developers have gone from AI is not helpful to me in my job circa 2023, to my job is fundamentally different. All I do all day, every day is interact with AI agents and tell them what to do and oversee their work and plan their work, et cetera. If that's coming for lawyers, if that's coming for doctors, if that's coming for all white-collar work, you can understand how this revenue can continue exploding at this absolutely insane pace and how also it can keep diversifying across the economy until these companies are ultimately worth a trillion dollars. Okay, now the bear case. The first thing to say is that I went to business school, and there's supposed to be like a mathematical relationship between the amount of profits a company makes and the value of the stock price. Um, and and you can actually run an equation that tells you like what is the equivalent between a bond that produces such and such amount of cash, and then how much cash would a stock have to produce in order to be valued the same way. So let's talk about a very simple kind of bond, which is an annuity. An annuity is just a bond that that kicks out a payment uh every single year of a certain amount. And so you can buy, you can buy an annuity today with today's money, and then you get future money as the bond pays out. Well, to justify uh a valuation of one trillion dollars, if you were to value this as an annuity, um at today's like risk-free interest rate, like what we sell risk-free US treasuries for, that is equivalent to a $45 billion risk-free, aka guaranteed cash payment every year forever, forever, for all time, you're gonna get $45 billion a year guaranteed. That's what uh an annuity that costs $1 trillion is supposed to be worth. How much money are these AI companies kicking out every year? Negative, negative money, right? They're not kicking out $45 billion a year. And and companies that do kick out $45 billion a year or higher are companies like Apple, right? That have 80% profit share of an entire industry that's one of the largest industries in the world. Now, here's one thing I said risk-free, right? These are risky. Like, like, like this, this could all not, this could all fail, right? You can't have 10 companies who make AI models who all have monopolies. Um, if monopoly means only one of them survives. So if you want to justify a $1 trillion valuation and you're risky, not risk-free, it's not $45 billion a year equivalent, it's $90 to $100 billion a year every year guaranteed. So this is like crazy wrath of God money. And how could that like that that like number one in the bear case is this does not resemble their current financials at all. Their current financials are they could make money, but to make money, they would have to stop investing. And if they stop investing, they're gonna be immediately overtaken by their competitors who will have a better model. Um, so while they have profitable gross margins, they're deeply unprofitable on a net profit basis. And how and when that will stop, and they will start generating incredible net profits, is only gonna happen on the other side of hundreds of billions of dollars of infrastructure capex, which is kind of a scary thing to do. Um okay, that's one part of the the bear case for these companies. The other part of the bear case uh is not just that they're risky, not just that they might not be the winner and what could be a winner-take-all industry. It's also that the the Chinese open source models. And the the key here is that Chinese open source models are extremely cheap when they're provided through cloud services and when they're just downloaded as uh models that you run locally, they're essentially free. So the if if the choice is between Claude Mythos, which is this absolutely exquisite capability or GPT 5.6, the latest and greatest open AI model, which is another exquisite capability, but it costs an awful lot of money. And Chinese open source models, which are approximately free, what share of the market potential is that going to eat up? What share of these use cases is is good enough enough, right? Like maybe if I'm developing some mission critical product, I want mythos, I want Fable 5, I'm willing to pay what it takes for my mission critical assignment. But if I'm like just creating a chat bot for my website so that users can ask questions that really they should just be reading in my FAQ, um and all the LLM has to do is go fetch from the FAQ and deliver it in a conversational interface, then yeah, maybe the the Chinese open source is good enough for that. And the question is if if both open source and closed source keep getting better, what percentage of the addressable market looks like that sort of sufficiency threshold? That right now the sufficiency threshold is for FAQ chatbots for Airbnb, for example. They've said for a lot of their use cases the Chinese models are good enough. But if the models keep getting better and better and better, maybe more of the world uh will look like that. And I think that the challenge here really comes to um distillation and the fact that right now the defenses, there's an offensive and a defensive game when it comes to Chinese companies trying to distill American frontier models. And I think it's fair to say that right now the advantage lies with the distillers. The technology for harvesting the intelligence of these models and replicating it is superior to the technology that exists against prev for preventing that kind of out.
SPEAKER_02Yeah, and I mean to be specific, we're talking about the type of distillation that we saw with the development of Deep Seek, where certain results from the model itself spat back out saying that that was essentially information, it was it used the same training information as the frontier models in the United States.
SPEAKER_00Yeah, no, the the the story, the story you're um referring to is is really funny. Um in when when Deep Seek was really hitting uh hitting the public consciousness back in January of 2025, um you could ask DeepSeek, what AI model are you? And it would say, Hi, I'm ChatGPT. It's great to help, right? And the reason for that, like what I guess we should take a step back. Like, what is distillation and why might you want to use it? I think it's helpful to use the analogy of let's say you're trying to create an AI system that is like the world's best chess player. Okay. The way you would do that using neural network deep learning is you would go get a library of chess games, a huge training data set, you would feed that to your chess algorithm, and then it would uh spit out an AI model that has learned from those chess games and is a really, really good chess player. Well, but then you can have that AI model play against itself and create additional data of chess games, and you can feed that back into the training algorithm. And so now you have the original training corpus, which is like a bunch of chess games you found on the internet. You have this additional training corpus, which is your AI model playing against itself, and after a ton of investment in data acquisition, and after a ton of investment in computing resources, you have this ultra powerful AI model, call it the chess bot 5000, and then you release that on the internet and you say, Look, guys, I've made the world's best AI chess player. Well, the Chinese companies they can just go play games against the Chess Bot 5000 and have it play against itself. And so now they can acquire a training data set of chess games, but they start at the top. They don't start at the bottom. So they're all of the games in their library are Olympic gold medal quality chess games. And they skipped all the computation of that early reinforcement learning, they skipped all that data acquisition of like finding old chess games from the 1700s that were really high quality. Um, and so they they only incur a fraction of the cost to extract most or all of the benefit. And you can do the same thing with chess games, you can do that with computer programming, you can do that with professional white-collar work type questions, you can do that with math. And so the point is as long as Chinese companies are able to access American AI models and do so in a way where they're allowed to download the results of a chat conversation, especially chats that involve stuff like programming, they can get most or all of the benefit for a fraction of the cost. And then with open source, they essentially give it away for free. So it's not just that they're stealing the intelligence of the AI model, they're also harming the ability of American companies to earn revenue from that because they're giving it away from free. I mean, the practical implementation is different, but the strategic outcome is no different than if Chinese companies just bought American pharmaceutical pills, did some chemistry to figure out what the drugs were, and then they released generic versions of those drugs to the entire world, right? The US companies they had to search through every potential drug that might be useful. They had to run clinical trials. All the Chinese had to do is copy the right answer and then give it away for free. And that that's why distillation is a real, real threat to the long-term story, right? The Chinese um AI companies have really brilliant engineers, right? They absolutely do. Certainly. Um but that is not why these Chinese open source models are like such a strategic threat to the American AI companies. They're such a strategic threat because distillation works. We don't have great defenses against it, and because we don't have a great mechanism for telling companies not to use open source, right? We can say, no, you're not allowed to sell Chinese pill X on the American market if we know that that company is infringing upon American patents. But for some reason, we have chosen not to, or we don't believe that we can credibly enforce equivalent bans on Chinese software.
SPEAKER_02Yeah. Uh so just to kind of follow through on one point, um, and it I'll I'll give you kind of an example to begin with, because I really want to hammer this home. In other parts of the national security industry, there are types of technology called anti-tamper technology, right? So if you have a box that's valuable, has very valuable national security stuff, uh, it's designed so that if it falls into the hands of the other guy, when the other guy tries to pry it open, it essentially blows up and it does so in a way that you can't reverse engineer the stuff inside of set box, right? So, general theory for anti-tamper technology, and it's almost rudimentary for how we design certain things to make sure that the IP that we're talking about and really kind of the capability that we're talking about doesn't fall into the hands of the other guy. Um, it doesn't seem as though there's really a version of uh AT anti-tamper technology out there for this particular technology set. So just to hammer that one home, does is there thought to that? Does that exist?
SPEAKER_00Yeah. And um this is this is the right question. And it's it does exist. The problem is that it's not effective enough. So um with the uh anti, you can call it anti-tamper, uh, let's just say it's anti-distillation, right? With the anti-distillation techniques that exist today, you can make because the companies distill their own models for the same reason, right? That they can create models that are smaller, that are um cheaper to serve to users, that are you know useful for subsets of applications, right? Like distillation in some sense is just like a law of computation and mathematics that this should work. So if you want to stop it, you can introduce a bunch of mechanisms that are designed to make it harder. A very simple one is just like IP address monitoring. Um, and you can also like look at the types of questions a user is asking and be like, okay, this does not look like a normal user. This looks like a user who is like going from A to Z and asking a question about every letter in the encyclopedia, like because they're trying to download the whole corpus, whatever the computer software equivalent of that analogy is. Um and and you can try and like shut down their access as soon as it looks like they're doing um distillation type stuff. You can also try and spike the data set. Um Cloudflare has a thing called AI Labyrinth, which is basically trying to monitor accessing computer websites. And if it decides that you are an AI model and you are uh scraping that website in violation of that website's terms of service or robots.txt file, it will just start create using generative AI actually to create new BS web pages that are full of BS information. So literally they're poisoning the data set to try and ruin the training. Um, you can do the equivalent of that AI labyrinth thing. So instead of cutting off the access of the people who are trying to distill your model, you say, Oh great, I found a distillation farm. Time to poison it, right? So these countermeasures do exist. The problem is that from what I understand, um, and this is this isn't this is uh an arms race, right, between the distillers and the counter distillers. So it's gonna evolve and change over time. At least right now, it looks like distillation countermeasures can make distillation maybe 10 times less effective than it would otherwise be, right? I said the the the companies distill their own models for a lot of uh reasons. So if that's like 100% efficacy, you can make distillation that's adversarial, like what the Chinese companies are trying to do, be only 10% as effective as that. The problem is that distillation, compared to building your own gigawatt data center, distillation is like a thousand or ten thousand times cheaper as an option. So even if distillation is only 10% of adversarial distillation is only reaching 10% of the potential of cooperative distillation, it's still way, way, way more attractive than building a gigawatt data center, than harvesting the entire internet as a training data corpus, and on and on and on. And so that's really the the question is like, how do we make adversarial distillation return on investment unattractive compared to having to do your own thing? And keep in mind, we're trying to make it difficult or impossible for China to do their own thing with export controls, with other measures. And so that's kind of the outstanding question is just how good can adversarial distillation countermeasures get? Um, if there's no progress over a five to 10 year time frame in blocking adversarial distillation, and if there's no barrier to companies around the United States and around the world using open source models that were created as a result of adversarial distillation, then these trillion dollar valuations don't make sense, full stop, right? Um last thing I will say here is if open source continues, and let's just say, like, not even in a um adversarial distillation type context, just in any type context, if if open source continues, but it lags the frontier. I think what you would have to believe for the bull case to be true is essentially that most of the revenue and profits in the industry will be made with frontier performance. So this would basically be a replication of the Apple story, but in AI, right? So Apple only sells 20% of the smartphones by unit count, but they make like 80% of the global profits. So it could be that the vast majority of users are using open source, but they're they're they're not paying anything. They're getting it for free. So the Chinese companies aren't making any revenue from that. Meanwhile, for the people who need the mythos of today and the mythos equivalent that's way better five years from now, if they're actually willing to pay for that superior performance because they're doing hardcore software stuff or whatever may have you, right? It could be that there's a huge disparity between the number of users and the revenue of profit of the industry. And I'm sure that's what the companies are hoping for as they think through trillion dollar valuations and whatever they're hoping with their valuation will be five, 10 years from now.
SPEAKER_01Certainly, certainly.
SPEAKER_02Um so I wanted to ask a clarification question. I think it's probably going to be a detailed one here. Uh, you you did mention kind of the economic challenges that are following a lot of the software as a service companies out there, right? Kind of in light of all these advancements. And the term that I've heard is SaaSpocalypse, if we can put a specific one on it. Um so can we talk a little bit about uh specifically what is the SaaSpocalypse and also kind of the economic implications around the changes to those companies?
SPEAKER_00The valuation of a lot of software as a service, SaaS companies, is plummeting on Wall Street, even beyond what I think is appropriate. But if you think about like a company like Shopify, which makes it easier to make your own e-commerce website among other functions. Well, what if I just ask Claude to make me an e-commerce website, et cetera? So the the bear case for SaaS companies is just that like whatever their boutique software, their exquisite software is, whether that's Salesforce for customer relations management, whether that's Microsoft for Office productivity type software, whether that's Shopify. The point is like in the future, I won't buy Adobe Photoshop. I'll just say, Claude, make me an image editor, or I'll just say, Claude, edit my images, right? And I'll I'll interact with it in that way. Um so this is really scary if you're you're SaaS companies and you rely upon this. And I think the the the bear case, even if it's right on the extent, I think it's off on the timing. I I think these companies have more runway than is is commonly assumed. Um but I think Microsoft's CEO, uh Satya Nadella, put it really well, which is like you have to earn the right to exist. Right. Your customers have to be glad that you exist. And in that regard, I think the the the present competition of AI and the potential future greater competitive pressure from AI is making all of these SaaS companies ask themselves how are we going to be sure that our users are not making a mistake when they choose us versus just choosing some like homebrewed AI custom software? And I think there was a great um example that was given uh with Ben Thompson on Stratchari um interviewing an analyst from Moffat Nathanson. And he said that like not all of Shopify's benefits come from Shopify as a software platform. Some of them come from having really large economies of scale so that they can negotiate attractive terms. So for example, like when you think about like what it costs to process a payment online, a lot of that goes through Spipe. Shopify gets a huge discount because there's such a like such a huge share of online payments go through the Shopify platform to Stripe, that Shopify gets really attractive payment terms that they then pass on the savings to their customers. So what that means is even if Claude replicates one-to-one everything on a software basis that Shopify does, they can't replicate those economies of scale and those attractive payment terms that Shopify can argue for on behalf of its customers. So that's like an extreme case, but I think it's it's it's relevant to the point, right? That like all of these companies have to not only have an awesome software as a service capability, but they have to have other reasons that they can justify their existence as AI software generation gets better and better and better. Which, as I said, right, if you've if you've tried out Fable Five, it is pretty amazing to just wish for software and then have it appear. Um, it's it's pretty it's pretty entrancing.
SPEAKER_02Yeah, yeah. Uh I was uh speaking to my father this weekend. He was telling a story uh about my uh great-grandfather. Uh in something that you said there just kind of struck me on this. My great-grandfather was uh the national sales director for a company called National Cash Register. Uh, he lived in northern New Jersey.
SPEAKER_00Like the OG computer company. Oh my god, this is amazing. Yeah.
SPEAKER_02Yeah, he actually knew Mr. Watson that founded IBM. Uh, and there's an interesting family story about why I'm not worth billions of dollars because uh my great-grandfather was actually offered initial equity in IBM, but thought it would have been a conflict of interest to accept it and refuse to do it because he was a man of great character.
SPEAKER_00Um but looking back at ethics, grandpa, I would have preferred the billions.
SPEAKER_02It's uh just definitely bringing those ethics with me. It's great. It's it's working out well. Um But the job that he did, the job that he did once upon a time, uh, was literally selling cash registers, right? And that business still exists. There are point of sale systems at a lot of brick and mortar stores kind of uh throughout the country, but it's much more of a software-driven, data-driven, data collection business than anything that existed before. So the type of business that he did once upon a time just cannot, it does not and cannot exist anymore. But there certainly is a stream of that that exists today just in a much, much, much different way.
SPEAKER_00Yeah, like if you're if you're a lumber mill, you you can you can keep running the same business for hundreds of years. But for almost every other industry, basically a lot is gonna change. A lot has changed. And if your theory of the case is I'm gonna keep doing what I'm doing, and that's gonna be how I stay in business, like no. No, no. Um, Adam, this was a ton of fun. Uh, thanks so much. Um, and uh for those of you out there, thanks for listening to the AI Power Podcast. We'll be back again uh next week digesting all the big stories of AI. Thanks, Greg. Thank you.