B (4:44)

Yeah, absolutely. So on that note, we already know that one of the biggest spenders in AI infrastructure is Elon Musk's X AI. They recently made some big disclosures about its Colossus 2 supercluster. So what did we learn about these new computing facilities?

A (5:02)

Yeah, so I think it's really helpful that XAI is still kind of committed to these single facility, single campus computing facilities. Some of the other companies like Google, Microsoft, OpenAI have moved towards more distributed training. So splitting the workloads up amongst multiple computing facilities. And that has some performance downsides. Not horrific performance downsides like it used to have, but it also made it a little bit more difficult to sort of say, apples to apples, how big is the big biggest computing facility of today compared with like where we were not that long ago? So there's two common ways for thinking about, like, how big a computing facility is, how big a data center or a supercomputer facility is. One common way of doing it is how many GPU chips are inside it, which is a rough proxy for total computing power. Obviously it kind of falls apart because the chips of today are way better than the chips of five years ago. And then another metric that is commonly used is how much energy the facility consumes, like in megawatts or in gigawatts. And that kind of tells you how big the facility is in a way that will remain true even after older chips are swapped out for newer ones. So it's kind of hard to get a true apples to apples comparison of like the past and the future. But those are two commonly used metrics, though imperfect, and they give us something that we can use to talk about just how big. Colossus 2, because Colossus 1, the predecessor also used by Xai, was at one point the largest supercomputing facility on earth. Now it's going to be dwarfed by its successor. So what is all this coming together to say? Let's start with points of comparison historically. So GPT3, which was developed by OpenAI and was the predecessor to GPT 3.5. GPT 3.5. Hopefully folks will remember when Chat GPT first launched In November of 2022, that was GPT 3.5. So the first version of chat GPT, if you used it, that was 3.5. The one before that was GPT 3, and that was reportedly trained at a supercomputing facility with 10,000 Nvidia V100 chips. So the Volta generation of Nvidia chips, now GPT 3.5 was rumored, and here there's less reliable publicly available data, but was rumored to be trained in a data center that had something like 20,000 Nvidia A100 chips. So the next generation of Nvidia GPUs, and that was in 2022, that was state of the art at the time. Colossus 1, the predecessor to the Colossus 2 facility, started construction in September 2024 with 100,000 Nvidia H100 chips. So the Hopper generation, and it was constructed in only 122 days, which is shockingly by industry norms. And now we're headed, Sorry, there's one more thing. And then Colossus 1 was upgraded by February 2025 to have more than 200,000 chips, most of which were Nvidia H1 hundreds, but it also included the Blackwell generation of chips, B1 hundreds and B2 hundreds. So we're going from 10,000 chips is a state of the art facility in 2021, 2022 timeframe to now. Totally understandable in the industry why you would have a facility with 200,000 chips. And that's still Colossus 1. So we're going even bigger than what we're talking about here. But I want to dwell on the difference between the old chips and the new chips. So according to Nvidia, the sort of per chip performance, processing power, and here we're using the metric of integer 8 tera operations per second, which is sort of like using a certain kind of mathematical numerical format. How many calculations can you do per per second? That increased 32 fold between the V100 generation and the H100 generation of chips. So we went from 10,000 Nvidia V100 chips in like 2021, 2022 time frame to now more than 200,000 H100 chips. And so that's 20 times as many chips. And each chip is 32 times more powerful than the ones in it. So we're talking like hundreds and hundreds and hundreds of times more aggregate computing power coming up on like a thousand times more aggregate computing power used to train and inference. You know, the most advanced models out there, of which grok, which is the XAI model, is absolutely, you know, on the overall frontier. But everything I just said, that was Colossus 1. And now Xai has just announced what they're doing with Colossus 2. And some really good reporting has been done on this topic by Semianalysis, which has a great write up on Colossus 2 which I encourage folks to go read. But here's just one quote that I'm going to get from them. The Colossus 2 project was kicked off on March 7th, 2025 when XAI acquired a 1 million square foot warehouse in Memphis and two adjacent sites totaling 100 acres. Acres are normally used to like describe farmland. So it gives you a sense of like how big this stuff is. By August 22nd, 2025 we count 119 air cooled chillers on site. That is roughly 200 megawatts of cooling capacity. That's enough to power roughly 110,000 GB200 Nvidia chips. So the chips again are getting much, much better. The, the, the Blackwell generation of chips is sort of depending on how you count, somewhere between twice as good and five times as good. And that's on top of the previous 32 times as good that we were talking about. And in terms of power, we're learning that XAI is absolutely going to be building gigawatt scale energy and they're doing it, you know, right across the border in Mississippi to supply energy to Colossus 2. So Colossus 2 is not even as big as it's ult going to be and it's already on the gigawatt scale, which means we're talking like a Hoover dams worth of power. This is all getting really, really big. So when you think about like, how could it possibly be the case that this project is anywhere remotely the same size as the Manhattan Project? This is how building facilities like this and XAI is just one company, Amazon, Google, Microsoft, OpenAI others are all, you know, doing same ballpark sized capital expenditures. Maybe they're not doing it all in like one big facility the way grok is. But OpenAI for example, said that they're going to have something like a million GPUs in use by the end of this calendar year. So going from 10,000 GPUs back in 2021, 2022 to a million GPUs, like we just keep going up by orders of magnitude, multiplying by 10. And that's how you get from AI is like an exciting niche in the technology industry to AI is driving the overall returns of the stock market. I mean, I heard one economist who basically said that the amount that we are spending in the private sector on AI is comparable to like the economic stimulus packages that the government often puts in place to pull us out of a recession, like we did after the 2008 financial crisis, for example. So these are big, big numbers. And it really comes down to the relentless demand of modern AI for computing power.

B (13:19)

Well, we're going to keep an eye on all of that AI infrastructure and specifically like the energy required to run all of those data centers on the podcast coming up soon. But on that note about chips, let's change gears to China's ban on Nvidia. Ch. So China was reported to have banned the use of Nvidia chips among its private sector companies. So Greg, can you tell us like, what's going on here?

A (13:48)

Yeah, so you just heard me talking about like, how awesome Nvidia chips have gotten. So why would China be banning them? Well, I think it's first worthwhile to ask ourselves, like, what do we actually know about the situation? So the Financial Times, which has done some really good journalism on this topic, on September 17th they reported, quote, the Cyberspace Administration of China told companies including ByteDance and Alibaba this week to end their testing and orders of the RTX Pro 6000D, Nvidia's tailor made product for the country, according to three people with knowledge of the matter. And then it goes on to say it comes after Chinese regulators concluded that domestic chips had attained performance comparable to those of Nvidia's models used in China. So, so I think it's really important to like, dwell on like, what is actually being said here. Oftentimes this is reported as China bans Nvidia chips. But actually more precisely what has happened is that they have banned this RTX Pro 6000D, you know, which is a model of chip that Nvidia was hoping to sell to China. It is not comparable to the latest generation of Blackwell chips. Right. It is not the case that China went to all of its leading hyperscalers and AI companies and said, hey, all those smuggled Blackwell chips that you've been able to import. And keep in mind, the Financial Times reported earlier this year that there was more than a billion dollars worth of Blackwell chips being smuggled into the country in just like a two or three month period. China's not saying like, like give up all those Blackwells and throw them in the garbage. It's simply saying that this other Nvidia chip, you're not allowed to buy it. So this is a noteworthy development in that the Chinese government is making demands of private sector entities when it comes to the Nvidia chips. Because earlier stuff that we've talked about on this podcast, for example, the municipal government of Beijing requiring that 90% of the chips in their computing facilities be locally produced. Those are government computing facilities. So like the, the local government of Beijing or maybe even state owned enterprises. So what's new about this mandate is that it's coming down on private sector entities, but it's coming down on a chip that is not close to the state of the art in terms of like, what Nvidia is selling American companies. So that's the thing is like, what we've seen empirically is that Chinese companies, according to reporting by Reuters and others, still want to buy the H20 chips, but they've been presumably instructed by the Chinese government to not order those chips from Nvidia. And there's, you have to ask yourself, like, what's the potential explanation for this observed behavior? And I think there are like two plausible hypotheses. The first hypothesis is that the Chinese government is trying to support local domestic chip manufacturing and chip design of GPUs. So these are the kinds of chips that are designed by Huawei, Cambercon, Alibaba, Tencent and others, and manufactured either by SMIC or by Huawei, because those are the two most advanced logic chip manufacturers in China. So one hypothesis is that the Chinese government is basically being sincere and saying that like there's going to be good enough chips and there's going to be enough of those chips that we don't really need this sort of second rate Nvidia chip overall. So here's a quote from Bill Bishop who writes the Cynicism newsletter, which is excellent and here's what he had to say about it. Quote the headline of the story is China bans tech companies from buying Nvidia's AI chips. But it's not clear if policymakers have decided that they no longer want want any Nvidia AI chips or just not the hobbled ones that they believe can be replaced by PRC chips. There is still a chance this is part of a play to get the Trump administration to approve the B30, the modified Blackwell chip that is many times more powerful than the H20. Perhaps even then the full Blackwell as well as the relaxation on export controls around high bandwidth memory. Any US China trade deal would likely include large purchase commitments from the PRC side and a massive chip order promise could help sway President Trump. And I think that second explanation is the more likely one. Essentially the Chinese government is trying to arrange a large they're trying to negotiate their way towards getting the B30 chips. So if you'll recall, President Trump said that Jensen Huang was coming to see him about legalizing the sales of the B30 chip. At least the first one of those conversations is reported to have taken place. So Nvidia, they're anti any restrictions on chip exports to China. But if you can't get all restrictions lifted, they would like at least the B30 to be allowed to be sold. This is a chip that is many, many times better than the H20 chips that the Trump administration originally banned and thereafter decided that could be sold to China. And so I think that this is a negotiating strategy on China's part. Basically saying either give us the best stuff stuff or something close to the best stuff or we'll take nothing at all and we'll pump all of that money into our local companies. So I think we have to ask ourselves like just because China doesn't want to buy h20s does that mean it is in America's interest to sell them even better chips, the B30s? Like I would argue no, this grok scale facilities like Colossus 2, those million chip centers where the chips are hundreds, hundreds of times better than what was available only a few years ago. That is specifically the AI future that export controls are designed to block Chinese companies from reaching. If you think about like the Deep seq moment and all of that. It's still the case that if they could have a million chip facility, Deepseek would absolutely love to have a million chip facility. And if that million chip facility could be full of Blackwell chips instead of, you know, H20s or even instead of like Huang Huawei chips, that is a way, way better option for them. So there's like multiple stakeholders in this ecosystem that Chinese policy is coming out of. You have chip makers, chip designers, and you have chip users. So a chip maker is a company like smic. They manufacture chips on behalf of Huawei, they manufacture chips on behalf of Alibaba, et cetera. They love this ban, right? This ban is wonderful news for the chip makers because it essentially guarantees a certain amount of demand is coming their way. Then you have the Chinese chip designers, they also love this, right? Because it again, it means a certain amount of demand is guaranteed to be coming their way. They really hope that the Chinese government doesn't get rid of this Nvidia ban and they hope that the Nvidia ban extends even more widely. But then you have the Chinese chip users and some this is a little bit confusing because Alibaba designs its own chips and also uses its own chips. But just think about the Chinese chip users, the big tech companies in China that want to build their own equivalent of Colossus 2. They hate this policy. They hate this policy because they want to be able to buy the best chips that are going to lead to the maximum return on investment of the tens of billions, billions of dollars in capital that they want to invest. And that is Nvidia chips. Right? The reporting all suggested that all these companies want to buy H20 chips. Even though that's not the best chip in the world, that's still the best chip they're illegally allowed to buy, including the alternatives available locally in China. And I think that that's likely to be the case for a while. If you think about the types of things that can be made locally by Huawei, by smic, by others. You know, Huawei and SMIC are still stuck at the 7 nanometer manufacturing node. That's something that TSMC was making chips with all the way back in 2019. So like six years ago. And if you think about like where Huawei might be in 2028, where SMIC might be in 2028, the best expectation we have now is that they're still going to be stuck at the 7 nanometer node. Maybe they'll get down to the 6 nanometer or the 5 nanometer node using a process that will be hugely inefficient. And so the point is, if we are effective in enforcing the semiconductor manufacturing equipment controls, then and that big AI buildout that takes place in China, it's going to be with really, really lousy chips. And that's going to affect their overall competitiveness in the AI ecosystem. And moreover, the amount of money that's going to be plugged back into smic, that's going to be plugged back into Huawei, I mean, that is money that the Chinese government is willing to spend anyway. The most recent semiconductor and AI government guidance funds all in like the tens or hundreds of billions of dollars thing. You heard it best from Deepseek when he said we are not capital constrained, we are chip constrained. So from my perspective, if you look at the Chinese government's stated behavior in 2020, when we were willing to sell them any Nvidia chips, they had a government policy that said we're going to ban the use of Nvidia chips in three years. They were already willing to mandate this transition to again. Now we see them being interrupted in that plan by being blocked from accessing tsmc, by being blocked from accessing certain categories of advanced semiconductor manufacturing equipment. But that commitment to local production on the part of the Chinese government is pretty gosh dang unwavering. And so I don't see a plan here where we can affect what China wants to be and where we should expect them to be in terms of local production of chips in say, 2028. But in terms of where they are in AI in 2028, I think we have an opportunity to make a big impact on where they are. Because keep in mind, as I've always said, and as I said in the Deep Seq moment, biggest impact of export controls produce a lagging impact. You know, Deep Seek's achievement was in December of 2024 with chips that they bought in 2023. So giving up on export controls today tells you where China is going to be in AI in 2027. And a lot of folks, Elon Musk, for example, says that we are one year away from AI superintelligence. Anthropic recently made a similar claim. Now, you can believe that or not believe that, but I still think, you know, this is a really important time for AI leadership. And nothing about what China is doing suggests to me that we are going to be able to persuade them away from indigenization by being willing to sell them more. But I do think we have an opportunity to impact the efficacy of that indigenization. And that starts with not selling them H20s, not selling them other kinds of chips that are equivalently good or better, and also doubling down on the semiconductor manufacturing equipment export controls. The reality is that the only reason why Huawei is even able to produce these 7 nanometer chips and why CXMT is within striking distance of producing high bandwidth memory is because the export controls that we have for semiconductor manufacturing equipment took too long and had an awful lot of loopholes. You know, if you could go back in time and do the export controls on equipment, right, and do that back in late 2022 and early 2023, boy, would China be in really, really, really deep trouble right now. They're still in deep trouble, just not as bad as it could have otherwise been. And that's kind of where I think we are in this story. Now. China has put out a lot of information designed to persuade America that they're not afraid of not having access to H20 chips, that the local production ecosystem is good enough. And I just don't buy it. I just don't buy it in terms of the quantities that they will need to produce their ability to produce the surrounding ecosystem of chips, most notably high bandwidth memory, which eventually China is going to produce in large quantities. But I think there are ways away from that. And similarly, you know, when you hear like the Huawei production roadmap, they, you know, it sounds scary. Oh my gosh. You know, Huawei's talking about making more and better chips, but those chips are still going to be way, way, way, way worse than what Nvidia is producing. So that's kind of where I come down on this story. I think the claim, but like, what Alibaba is designing, what Huawei are producing, are better than the H20. I just don't think that's borne out by reality in terms of the performance of the overall systems. And I definitely don't think it's borne out when you take into effect producing enough of the chips at a high enough level of reliability and quality and having the surrounding software ecosystem that's good enough to do impressive AI work work. It's definitely the case that China is making progress on indigenization. It's definitely the case that they're going to become more and more capable. But the question is like, does that mean that we should give them all the sources of our current advantage? I would say no.

B (28:57)

So now that we've heard a lot about your perspective, what are other analysts and experts saying about these new developments?

A (29:04)

Well, I think There's a lot of interesting claims out there. So David Sachs, who is the AI czar in the Trump administration's White House, he wrote On X on September 18, quote, the message is clear. China is not desperate for our chips. It is producing its own and intends to compete globally in the semiconductor market. I think he's right, that that is absolutely the message that China's government is trying to send. I think there's a real question though as to whether or not it's true. True. Right. All the reporting of people who interview executives at top Chinese AI companies say they are desperate for Nvidia's chips. But if they were to say that on the record instead of off the record, well, they are at risk of being, you know, Jack Ma in the Chinese political system. Jack Ma, the famous Alibaba entrepreneur who was forced to go into exile for multiple years for pissing off the Chinese political establishment. I mean, that's why they can't say anything like that. I think China is posturing here. And, and that's it. Now I, I want to also quote here from a semi analysis analysis that came out on September 8th, which I thought was astute, quote, while there is significant messaging that China has decreased interest in the H20, we do not believe this to be true. Specifically, we do not believe this is reflective of China's demand for compute or foreign chips. It is an incorrect policy being pushed from the top down, which will be reversed as soon as Huawei runs out of HBM to produce more advanced. Sorry, to produce more Ascend chips. Or it could even be orchestrated brinksmanship to get approval to a more powerful chip. I think that's right. Here's Chris McGuire who is a former member of the National Security Council who has worked on these policies in the past. He tweeted on 9-8-18th and he's talking about Huawei's plan, Huawei's future semiconductor manufacturing roadmap. Quote, it's not very impressive. Huawei does not plan to produce a chip as powerful as the B30A, which Nvidia wants to ship to China now until the fourth quarter of 2028. In 2027, Nvidia's best chips will have 27 times the processing power of Huawei's best AI chips. So that's kind of where I think we are. The export control policy of the first Trump administration, of the Biden administration and even, you know, the, the second Trump administration, because they've continued important parts of the export control policy even as they changed some of it, is not having the full range of impacts that, that we could have had if we had had this policy designed correctly from the outset of October 2023 to. But it's still having a massive impact. If we didn't have this Export Control Policy Colossus 2 and like a million chip facility that would be under construction in China today and the chips powering it would have that Same sort of like 32x or 64x improvement that we've seen over that same period of time. That is the advantage that we've given our American companies. And I think that's part of the reason why, you see, OpenAI has 700 million weekly average users and is still, you know, has it facing a shortage of computing resources. So that's, that's kind of where we are.

B (32:29)

So what do all of these new developments mean for export control policy?

A (32:33)

Well, I think, you know, we're still back at that conversation, which at my, the best of my understanding is happening between President Donald Trump and Nvidia CEO Jensen Huang. One thing that I think is kind of different. Different is that when the Trump administration released the H20 deal, basically saying that H20s are no longer banned for export to China, they got some of the first important pushback from Republicans in Congress that they had received during this entire administration. And so I think if there was an assumption that the Republican Party was unified on this issue or that the Trump administration was not going to face significant pushback from within his own party, I think that that is now demonstrably false. And so President Trump, as he is, you know, running the analysis in his own mind on the pros and cons of allowing B30 exports, he has to, you know, take into account the fact that this is an unpopular move even within his own party.

B (33:39)

Okay, Greg, thanks for that great breakdown. We always follow chips on this podcast. So more to come in following episodes, but let's move on to Anthropic's copyright case case. So Anthropic recently reached a $1.5 billion settlement for pirating over 7 million books to train its chatbot. Claude, we covered this case in our July 2 episode on this podcast shortly after the judge's initial ruling was announced. So, Greg, can you just refresh our memory on why Anthropic was being sued and what the judge's decision was?

A (34:10)

Yes. So let's start with the case itself. There are three authors here who are suing Anthropic for pirating their books and then using those books to train their large language model. So now this large language Model has capabilities that it would not otherwise have if those books were not part of the training data set. Now it's only three authors suing, but they're suing as a class action. So what that means is that this legal team is now claiming to represent, if the judge grants them that status, the 7 million authors. Well, probably not 7 million authors, but the authors of the 7 million in books that Anthropic admits that it downloaded from pirated sources illegally originally. And so they're trying to determine, number one, is piracy legal? Answer, no. Anthropic does not even contest that. They did not contest that in the original lawsuit. Then the question then is the transformation of the works. So Anthropic is saying, basically, look, if you could log into, you know, Claude, which is Anthropic's chatbot, and say, print out the text of any one of the books in our training database, and then you would get that, they admit that that would obviously be the massive copyright infringement. But they also say we don't do that. Right. What we might do is we might use the information that the LLM learned by including that book in its training data into an output. So, for example, let's say that one of the books that they have is some kind of economics textbook, and it describes a certain kind of methodology for estimating inflation. Well, if you then ask Claude to estimate inflation given some inputs, and then it gives you an output, and that output is drawing on the methodology that was in that book. Is that transformation of the book and that transformed output, is that fair use in copyright terms, or is that essentially piracy? Again, by not having fair use of the input that you've sort of infringed on the intellectual property property. So here's what's very interesting. This, as we talked about on that July 2nd podcast, this is the the version of the case where the judge basically said had three big conclusions in that original ruling. Number one, they are a class action. So these three authors do represent the authors of everyone else. But then storing and using pirated copies is not fair use. Storing purchased books is fair use. So remember that after they had the pirated version of all the books, they actually bought hard copies of the books, cut off the binding, and then ran them through like an ultra high efficiency scanner and scanned literally millions of books by taking pictures of each paper page and then digitizing that. So the ruling said that that is fair use. And then it also said that training on the books is fair use. So that's a very, very big deal. So then what's left over is kind of like, okay, given that you also pirated these books, you know, what do you owe the authors? And that kind of comes back to this settlement. One final thing it's worth pointing out is that like, this case happened right around the same time as a lawsuit suit against Meta. So Anthropic and Meta both faced major copyright cases. They were both class action and both were, you know, both of the plaintiffs were groups of authors arguing that their copyright was infringed by training. So the Meta judge did not see fair use from that transformed training version. But the Anthropic judge did see transformed fair use. So like, all of this was super unclear. And into this now runs this settlement. So let's, let's talk about the settlement. Anthropic is supposed to pay at least $1.5 billion. So that is roughly $3,000 per work for the 500,000 works that were included in their training data set where there is some kind of copyright infringement that is adjudged to have taken place. So they have to find like, who was the owner of the intellectual Property, find these 500,000 works, which are the subset of those 7 million works, and they have to pay $3,000 per. So on the one hand, this is like the largest, according to the court filing, the largest publicly reported copyright recovery in history Street. On the other hand, it is actually kind of small compared to just how bad it could have been for Anthropic. The original petition in this court case was talking about damages up to $150,000 per work, meaning like hundreds of billions of dollars in legal liability, which would almost certainly result in Anthropic's bankruptcy. So where we're going to end kind of depends upon what the judge decides. And this is Judge Alsupp, who is as the original Anthropic ruling in a Sept. 25 hearing. According to the Associated Press, he was critical of the settlement and said, quote, we'll see if I can hold my nose and approve it, end quote. Which is a kind of juicy statement for, for a judge to be making about this. And that really leaves us, you know, wondering what's going to happen now. Again, this is really all about the pirated content content. And the original ruling on the, the, the, the, the non pirated content is still kind of up in the air. It kind of is going to require an appeals court to reconcile the differing views that came out from the judges in the Meta case and the Anthropic case.

B (40:27)

Okay, thanks for breaking down the agreement with Us, Greg, so how are the authors and other stakeholders reacting to all of this?

A (40:33)

This. Yeah, so one of the plaintiffs, Kirk Johnson, was reported to have said in an interview with Reuters that the settlement marked, quote, the beginning of a fight on behalf of humans that don't believe we have to sacrifice everything on the altar of AI. And there I would sort of emphasize, like, beginning, because this is about this, this, this settlement was about piracy. And people knew that piracy was illegal before this court case even started. Anthropic switched over to that scanned book methodology because they knew that piracy was illegal and that they were putting themselves at legal liability. And so in that regard, I'm not sure that, you know, this, this, this court case tells us very much. And I think if you look at the Authors Guild CEO Mary Rassenberger, what she told npr, I think is, is more in line with the reality of this situation. Quote, the impact of this decision for book authors is actually quite good. The judge understood the outrageous piracy, and that comes with statutory damages for intentional copyright infringement, which are quite high per book. And that's kind of where I think we are. The bigger AI implications really depend upon the resolution of the disparity in the Meta case and the Anthropic case. For what is the overall relationship between AI companies that take in copyrighted works as part of their training data and then transform them. Where are the boundaries of fair use in that story? That's the one that I think will have pretty major implications for the overall industry.

B (42:11)

Okay, so what precedent, if any, does this set for future copyright cases against AI companies?

A (42:19)

Yeah, and I think I just sort of said that this one is not the case that's going to have the binding legal precedent that we're still waiting on. But I think Luke Madonna, who is an associate professor of law at the London School of Economics, said in Fortune, quote, while the settlement doesn't set a legal precedent, it could serve to legitimize the author's claims and may set a business precedent for similar cases. And what I mean by that is, like, Anthropic is not generally believed to have been the only AI company that downloaded a whole heck of a lot of pirated content that was copyrighted in order to use it in its training data. So as a business precedent, you know, for what needs to happen for these other companies that are presumably at legal risk for having pirated stuff and using it in their training data sets, that I think is important for that legal precedent. We're still waiting for, you know, as I said before, the higher resolution there's also a couple of other major outstanding court cases, not least of which was OpenAI being sued by the New York Times for copyright violations. So that case, the meta case, the anthropic case, and the sort of higher level versions of those cases, if they go to appeal, that's where we'll start getting some clarity on the overall legal precedent. One thing to note here is that, you know, the higher level courts of appeals, they can set a binding precedent, which means if you are a district court, it is not your job to determine, you know, what should constitute fair use. Your job is just to take the instructions of that higher level court and then apply them in your case. So you will get a binding legal precedent from the appeals court. But just given the stakes of this, just given how much, you know, the Motion Picture association of America or the Authors Guild of America, and just given how much all of these different industries have, I would be be not surprised at all to see a case or two here end up in the Supreme Court. And so the final sort of binding legal precedent will probably come from them until and unless the US Congress passes a bill that provides additional clarity.

B (44:34)

Great. Thanks for that recap and update, Greg. And last but not least, let's move on to our final topic. ASML's investment in Mistral. So we've seen several big investments by chip making companies this past month. Most recently on September 22nd, Nvidia announced a $100 billion investment in OpenAI. This came after Nvidia announced a $5 billion investment in intel on September 18th. And Dutch chip making equipment company ASML announced a $1.3 billion investment in the French AI startup Mistral on September 9th. So let's start with the Nvidia investments. What is Nvidia hoping to get out of these deals?

A (45:10)

Well, I think just looking at all three of these deals briefly together for a second, they're a really interesting example of vertical integration. So if you think about in like economics terms, when one company buys another company, that can either be horizontal integration, which would be like Coca Cola buys Pepsi. They make the same products and they compete against each other. Or they could be vertical integration which would be like, like Coca Cola buys a bottling company or Coca Cola buys a sugar farm or Coca Cola buys a grocery store chain. Those are vertical integration because they're buying either their suppliers or their customers. So none of these deals are a merger, but they are pretty sizable investments. And they're along the vertical axis of corporate integration because these are customers investing in suppliers or vice Versa. So ASML is the provider of EUV lithography systems. They make the machines that are at the heart of modern semiconductor manufacturing facilities. And so, because AI chips are such a critically important driver of overall spending in the semiconductor industry, you know, ASML buying an AI model provider, Mistral makes AI models. They compete with OpenAI, they compete with. They compete with Google, et cetera. That's a pretty interesting degree of vertical integration, right? Secondarily, you know, Nvidia investing in OpenAI, that is a supplier investing in a customer. And the relationship between OpenAI and Nvidia is so much of the story of the AI boom of the past five years. Jensen Huang very famously hand delivered a DGX1AI supercomputer to OpenAI something like five years ago because he was so bullish on what that company was going to mean for the future of AI and for the future of GPU AI chip demand. Wow, was he right on that kind of a bet? So there's a couple ways of interpreting, you know, what's going on here. Number one, could be like, okay, this is all just part of the AI bubble, right? That like, like, these companies all have insane valuations. These insane valuations are in part driven by ridiculous requirements for AI capital expenditure, right? That Manhattan Project level of investment in AI computing power. Nvidia is the most important supplier of what's going into those data centers. So they would like that Manhattan Project or greater than Manhattan Project level of space spending to continue. If that level of spending is going to continue, then OpenAI needs enough money to buy all of those chips. So where can they get enough money? Well, how about from Nvidia, which is now like the most valuable company in the history of humanity because their stock price is through the roof. So you could say that, you know, this is like a snake eating its own tail in order to grow larger. Like, this is all just part of some kind of bubble. But, you know, you can kind of see that there is a little bit of strategic logic here, which is to say that the relationship that Nvidia has had with OpenAI is part of the reason why Nvidia has been so good at seeing the future. Like, a lot of Companies in the 2021, 2022 timeframe were designing chips that were specific for running AI applications. But what they were optimizing for in many cases was lots of, like, facial recognition or, you know, audio generation and other kinds of, like, speech recognition. Like, that's the AI chips that they were making. Nvidia rather brilliantly was like, Nope. The future of computing demand is large language models. And part of the reason why they were in such a great position to see that future is they had this super intimate relationship with OpenAI, which sort of let them see what they were building and just how transformational it might be. So you can imagine it's not crazy that Nvidia would want that incredibly close relationship to continue. Especially since OpenAI was rumored to be opening up its own chip design team or rumored to be thinking about doing what Google has done right, which is create TPUs, an alternative to Nvidia chips that only works when you're inside the Nvidia. Sorry, inside the Google cloud ecosystem system. So this kind of locks down OpenAI as the biggest overall customer of Nvidia for the foreseeable Future, which if OpenAI retains its market share, you know, it's by far the most widely used AI chatbot in the world. You can see why Nvidia might like something like that. But if you're Google, which uses TPU chips and uses Nvidia chips, if you're Amazon, which owns a good chunk of anthropic stock and uses Nvidia chips, but also uses its in house Trainium and Inferentia chips, you're looking at this relationship and wondering, okay, wait a second here. Is this actually in our best interest to keep supporting the Nvidia ecosystem when they're supporting one of our big AI model competitors? Because Nvidia's great strength in a lot of this market is the that they work with everybody's model. They're sort of the basic platform for AI computation and that universal compatibility is a real strength. You can imagine that some of OpenAI's competitors might start having a stronger and stronger incentive to try and help TPUs or Trainium or what have you, be stronger competitors to Nvidia chips and the Nvidia chip ecosystem.

B (51:15)

Sure. Okay, well, let's switch over to ASML's deal with Mistral AI. So can you tell us more about the who these tech companies are and what exactly they agreed to in this deal?

A (51:28)

Yeah, so I already mentioned that ASML is the big provider of lithography for chip manufacturing. Mistral is kind of the darling of Europe, at least in political terms, as a local developer of advanced AI models. So, you know, the European Union, for example, is trying to use government to build big data centers. They love having Mistral be sort of the anchor customer for some of those government backed data centers. And when France had the AI Action Summit back in the beginning of this Year I'm told that French President Emmanuel Macron was like, encouraging people in meetings to download the Mistral app, literally, like while they were sitting in front of him, because he really loved them. As a European company that is headquartered in France. So ASML is based in the Netherlands, and so you could envision this deal being sort of like, oh, is this part of Europe trying to come up with an alternative to the American dominance of AI? At least the ASML CEO Poe is denying this. He said, quote, we don't decide on partnership based on location or geopolitics. We picked Mistral because they are, we thought, the best partner to execute on what we want to do. And the Mistral CEO also de emphasized the political angle. It's definitely the case, however, that the media is seeing this as a deeply political move and definitely the case that many European politicians are cheering this move move on political grounds. So just how impactful we overall see it, we will overall see it to be. I don't know at this stage, but it is a really noteworthy development.

B (53:22)

So bringing it back to Nvidia, where we started this conversation, they also just made a big investment in Intel. What do you think about that?

A (53:30)

Yes, this I think is something that certainly seems like a savvy move, at least politically. Right. The Trump administration just oversaw a government investment in intel whereby something like 10% of the company is now going to be owned by the US Federal government. Well, what's interesting is that that deal, one of the principal reasons was to make sure that intel stayed in the chip manufacturing business. One of the key provisions of the contract is that the government is going to be able to vote its shares and veto stuff if intel ultimately decides to split its chip design business from its chip manufacturing business. And the number one thing that intel needs is customers for its chip manufacturing business. Intel used to be such a big deal in the overall world of computing that just themselves as their own customer was enough that they could afford the quite expensive cost of staying on the technological frontier of semiconductor manufacturing. But as the new Intel CEO has said, that is no longer the case. Either they find external customers or they simply do not have the economics required to stay at the manufacturing frontier here. So when in, when Nvidia invests in Intel, a lot of folks I think are immediately thinking, oh my gosh, is this the investment that's big enough to ensure that Nvidia will now become a huge customer of intel chip production and thereby ensure that intel has enough demand to justify frontier chip manufacturing investments? And specifically frontier chip manufacturing investments investments in the United States. That was certainly my first reaction when I saw this deal surprise. Like none of that is mentioned in any of the deal specifics that either of the companies have talked about. I mean there's stuff that makes sense for these companies. Like Nvidia is going to have additional advantages when competing in the data center market. You know, there are some data centers that run on x86 which is an intel dominated standard. There are some data centers that run on arm, arm, which is an alternative competing standard. Nvidia works best in the ARM ecosystem. Now they'll have some co design opportunities with intel to make intel more attractive and competitive when operated in x86 style data centers. Intel is probably getting out of the GPU business as a whole except and exclusively to the extent that they're partnering with Nvidia. So that's probably good with Nvidia, but nobody's saying anything, anything at this stage about chip manufacturing. And so from my perspective, what that means is this is probably a savvy business deal on part of both companies, but we don't yet have anything that would actually talk about what could matter from a geopolitical geostrategic sense. If you want this deal to matter in geostrategic terms, you have to see Nvidia start betting on intel as a chip manufacturer in America. That's what the US Government wants when it invested in intel, when it said you can't divest your manufacturing business. So in the absence of that, we're still waiting. I mean, maybe that announcement will come later, maybe not. And whether it does or it doesn't, it's a very big deal for the future of the industry in America.

B (57:06)

Well Greg, I think that that's a great place to wrap up. We've covered many topics today and thank you as always for a great breakdown of all of them. And thank you all listening to another episode of the AI Policy Podcast.

A (57:18)

Thanks Sadie. Great to talk with you. Thanks for listening to this episode of the AI Policy Podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review on your favorite podcast platform and subscribe and tell your friends. It really helps when you spread the word. This podcast was produced by Sarah Baker, Sadie McCullough and Matt Man. See you next time.