Latest / Faster, Please! โ The Podcast / ๐ The Great AI Race as a clash over compute: my interview with analyst Ryan Fedasiuk
Transcript
- 0:00SPEAKER_01: Welcome to Faster Please,
- 0:01SPEAKER_01: the podcast, the audio-visual companion of my Faster Please newsletter.
- 0:05SPEAKER_01: I'm your host, Jim Pethokoukis, and my guest this episode is Ryan Fadaschuk.
- 0:10SPEAKER_01: Ryan is a fellow at AEI and the author of the substack,
- 0:13SPEAKER_01: Choosing Victory,
- 0:14SPEAKER_01: where he focuses on U.S.-China relations, technology, and national power.
- 0:19SPEAKER_01: He's also an adjunct assistant professor at Georgetown University.
- 0:22SPEAKER_01: Recently,
- 0:23SPEAKER_01: Ryan has written about the AI race between China and America,
- 0:26SPEAKER_01: and while much of the debate has focused on the models themselves,
- 0:30SPEAKER_01: Ryan has focused on the compute and chips being produced and how they can affect
- 0:34SPEAKER_01: the global balance of power,
- 0:35SPEAKER_01: which is also the subject of a new report,
- 0:38SPEAKER_01: VoltCraft, Industrial Competition in the Age of AI.
- 0:42SPEAKER_01: Our conversation will focus on that competition and what it can mean for
- 0:45SPEAKER_01: the AI industry in the coming years.
- 0:48SPEAKER_01: Hope you enjoy it.
- 0:49SPEAKER_01: Ryan, welcome to the podcast.
- 0:52SPEAKER_00: Thanks for having me.
- 0:54SPEAKER_01: I want to start a little with first because we're going to get to this great paper
- 0:57SPEAKER_01: that you wrote.
- 0:59SPEAKER_01: But I want to just get a feel for what sort of like your working assumptions about the world.
- 1:04SPEAKER_01: And let me start with an important one.
- 1:07SPEAKER_01: To use that phrase, what is your working assumption about how fast frontier AI is improving?
- 1:18SPEAKER_00: Sure.
- 1:18SPEAKER_00: I mean, I think it's improving.
- 1:20SPEAKER_00: Incredibly quickly,
- 1:21SPEAKER_00: it's hard to measure that reliably even in a single snapshot,
- 1:25SPEAKER_00: but just the pace of Frontier model releases has compressed from, say,
- 1:30SPEAKER_00: on the order of six months a year ago to a matter of a couple of months
- 1:35SPEAKER_00: or even weeks from some leading labs today.
- 1:38SPEAKER_00: People have conflicting views on the impact of recursive self-improvement.
- 1:43SPEAKER_00: Most people I talk to in San Francisco think that this is happening and augmenting
- 1:48SPEAKER_00: the way we do AI R&D,
- 1:50SPEAKER_00: it's probably happening in a pretty jagged way and humans are still the bottleneck
- 1:55SPEAKER_00: in a lot of parts of how AI is developed,
- 1:58SPEAKER_00: but clearly progress is accelerating faster and faster and faster.
- 2:03SPEAKER_01: Well, I mean,
- 2:04SPEAKER_01: I think, again,
- 2:06SPEAKER_01: I mean, nobody knows, though I think the people in San Francisco think they know.
- 2:14SPEAKER_01: Because I would assume that what you believe about how fast frontier AI is improving
- 2:20SPEAKER_01: and that then leaves a certain,
- 2:23SPEAKER_01: I guess,
- 2:24SPEAKER_01: assumptions about the economic potential of this technology and about
- 2:28SPEAKER_01: the national security implications.
- 2:31SPEAKER_01: If you think,
- 2:32SPEAKER_01: and I think economists tend to think more like this,
- 2:34SPEAKER_01: that this is sort of like the next step in the digital revolution and we can look
- 2:41SPEAKER_01: at previous technologies,
- 2:43SPEAKER_01: whether it's You know,
- 2:44SPEAKER_01: the PC or the internet as sort of benchmarks to understand this technology,
- 2:51SPEAKER_01: you might think one thing about those implications.
- 2:53SPEAKER_01: And if you're like,
- 2:55SPEAKER_01: think like,
- 2:56SPEAKER_01: I don't know, like,
- 2:57SPEAKER_01: Demisa Saba said he described us as being in the foothills of the singularity,
- 3:02SPEAKER_01: and in a couple years, we're really good, that won't, that will be clearly obvious.
- 3:07SPEAKER_01: If you think that, then I imagine...
- 3:10SPEAKER_01: What you believe about the economic potential and the national security implications
- 3:14SPEAKER_01: and what our public policy should be would be a different set of policies, maybe.
- 3:21SPEAKER_01: So that's why I asked you that question.
- 3:23SPEAKER_00: Sure.
- 3:24SPEAKER_00: Well, you know, let's put a finer point on it, right?
- 3:27SPEAKER_00: I've been focusing at least in some part on AI policy since about 2019,
- 3:34SPEAKER_00: a few years before ChatGPT came to be.
- 3:38SPEAKER_00: And at that time,
- 3:39SPEAKER_00: people were You know,
- 3:41SPEAKER_00: wildly unsure about AI development timelines and how long things would take.
- 3:47SPEAKER_00: Many people were skeptical that we would reach the capabilities we have already seen AI make.
- 3:51SPEAKER_00: And, you know,
- 3:52SPEAKER_00: I think at least according to definitions circa the late 2010s,
- 3:56SPEAKER_00: we're living through what I would describe as a fast takeoff.
- 4:00SPEAKER_00: We have,
- 4:01SPEAKER_00: as far as I am concerned,
- 4:03SPEAKER_00: you know,
- 4:03SPEAKER_00: artificial general intelligence capable of meeting the performance of human beings
- 4:08SPEAKER_00: or exceeding it at most tasks.
- 4:11SPEAKER_00: And I think that things are only likely to improve from here.
- 4:17SPEAKER_01: I think that's important to know.
- 4:20SPEAKER_01: When we talk about the US and China being in a race, what is the nature of that race?
- 4:29SPEAKER_01: Some people love the word race, some people don't.
- 4:32SPEAKER_01: If you're using the word race, you're assuming some sort of some endpoint where a side has won.
- 4:40SPEAKER_01: So is that what we're talking about,
- 4:41SPEAKER_01: a race where one side will win or is it really more likely to be kind of
- 4:46SPEAKER_01: a continuous technological and,
- 4:49SPEAKER_01: as you would say, industrial competition without any kind of finish line?
- 4:53SPEAKER_01: Is there a real race aspect to this?
- 4:57SPEAKER_00: I think there is a real race dynamic here.
- 5:01SPEAKER_00: I find it useful to deconstruct the US-China competition into four dimensions,
- 5:07SPEAKER_00: some of which are zero sum and some of which are not.
- 5:11SPEAKER_00: So I think about number one,
- 5:13SPEAKER_00: The race to develop frontier capability,
- 5:15SPEAKER_00: you know,
- 5:16SPEAKER_00: whose models are scoring better on frontier benchmarks,
- 5:19SPEAKER_00: able to do new stuff,
- 5:22SPEAKER_00: discover proofs to crazy math problems that have been unsolved for decades,
- 5:28SPEAKER_00: finding cures to ailments and things like that.
- 5:32SPEAKER_00: The second dimension of competition is about adoption and uptake by enterprises
- 5:37SPEAKER_00: and actually using the technology that we have made available in a lab environment.
- 5:43SPEAKER_00: More of a difficult question to answer or measure.
- 5:46SPEAKER_00: It's about how much computing power is available to enterprises.
- 5:50SPEAKER_00: It's about energy availability.
- 5:52SPEAKER_00: It's about the cost of tokens and running a model at scale.
- 5:57SPEAKER_00: The third dimension of competition is I pay a lot of attention to this.
- 6:27SPEAKER_00: Dimension of competition as kind of a US-China foreign policy guy
- 6:32SPEAKER_00: and national security aficionado.
- 6:35SPEAKER_00: I really worry about whose models are forming the lifeblood of the global intelligence economy.
- 6:43SPEAKER_00: And then the last dimension of this contest is not zero sum.
- 6:47SPEAKER_00: It is a contest to make our societies resilient to AI's disruptive effects.
- 6:54SPEAKER_00: This is something that I think fewer people have paid attention to,
- 6:58SPEAKER_00: but it's really starting to become almost the whole ballgame in late 2026
- 7:03SPEAKER_00: as frontier models become just so capable.
- 7:08SPEAKER_01: I think some people would view the race almost purely in that first dimension.
- 7:13SPEAKER_01: The race at the frontier,
- 7:18SPEAKER_01: under the assumption that there is an endpoint,
- 7:22SPEAKER_01: that once the models become capable enough,
- 7:25SPEAKER_01: almost those other three, well,
- 7:27SPEAKER_01: maybe two of the three don't much matter because it will be so amazing,
- 7:30SPEAKER_01: you know, super intelligence.
- 7:32SPEAKER_01: And I've read some, you know, war games, for instance, on the national security implication.
- 7:35SPEAKER_01: I've looked at some war games and...
- 7:38SPEAKER_01: The way those war games get run about whether you want to call it super intelligence
- 7:42SPEAKER_01: or artificial general intelligence is that it is a race
- 7:46SPEAKER_01: and someone gets there first.
- 7:48SPEAKER_01: And there is a massive advantage to getting there first and there is
- 7:51SPEAKER_01: a massive disadvantage for being the runner up.
- 7:58SPEAKER_01: So is that first dimension really the most important of the four?
- 8:05SPEAKER_01: I know they're all important.
- 8:06SPEAKER_01: I'm not saying that.
- 8:09SPEAKER_01: Can you imagine a scenario where that's actually, that's a real race?
- 8:13SPEAKER_00: Yes, I can.
- 8:15SPEAKER_00: I think a lot of people are hesitant to embrace that conclusion,
- 8:20SPEAKER_00: but I think that's kind of unavoidable.
- 8:23SPEAKER_00: We are obviously living through this uncertainty.
- 8:26SPEAKER_00: When Mythos was unveiled on April 6th,
- 8:30SPEAKER_00: I think it sent a real tremor through the Chinese national security apparatus
- 8:35SPEAKER_00: and kind of a moment of, oh no, what is this?
- 8:41SPEAKER_00: What do the Americans have?
- 8:42SPEAKER_00: What is this system capable of achieving?
- 8:45SPEAKER_00: When can we produce our own?
- 8:48SPEAKER_00: The release of GLM 5.3,
- 8:50SPEAKER_00: we're recording this in mid-August,
- 8:53SPEAKER_00: has kind of created a glass wing type capability sharing infrastructure.
- 8:59SPEAKER_00: Maybe you would say with Chinese characteristics.
- 9:01SPEAKER_00: And so it's interesting to see China's internet and cyber defenders create
- 9:08SPEAKER_00: a shared coalition of capabilities of their own.
- 9:11SPEAKER_00: But clearly,
- 9:12SPEAKER_00: I think we are living through a situation where every frontier enterprise
- 9:18SPEAKER_00: that touches any kind of critical infrastructure,
- 9:22SPEAKER_00: whether it's financial,
- 9:23SPEAKER_00: communications,
- 9:24SPEAKER_00: transportation,
- 9:25SPEAKER_00: health and water,
- 9:28SPEAKER_00: In every country wants access to the best frontier AI system they can have
- 9:34SPEAKER_00: to make sure that their cybersecurity suite is capable of defending
- 9:38SPEAKER_00: against frontier attacks.
- 9:41SPEAKER_01: Given the national security implications,
- 9:45SPEAKER_01: we just talked about the cyber implications,
- 9:49SPEAKER_01: is there a risk that we end up viewing AI too much through a security lens?
- 9:57SPEAKER_01: To the detriment of making sure that people can use these models and they defuse
- 10:04SPEAKER_01: and they get adopted and we end up locking them down
- 10:08SPEAKER_01: and we end up limiting their capabilities because we're concerned about
- 10:13SPEAKER_01: the security part of it.
- 10:15SPEAKER_01: I'm not sure I would have asked this question four months ago, but now to me it seems relevant.
- 10:20SPEAKER_00: Most definitely.
- 10:21SPEAKER_00: And I'm right there with you.
- 10:24SPEAKER_00: I have been surprised to see the direction of travel of AI policymaking in
- 10:31SPEAKER_00: the Trump administration,
- 10:32SPEAKER_00: which has really oscillated In ways
- 10:35SPEAKER_00: that I think very few people would have predicted a year ago,
- 10:38SPEAKER_00: the debate around the voluntary framework to vet frontier AI models,
- 10:44SPEAKER_00: questions about how voluntary that framework really is or what even is in it,
- 10:49SPEAKER_00: because it has not been made publicly available,
- 10:52SPEAKER_00: is a great example, I think, of what you're talking about.
- 10:56SPEAKER_00: Clearly, the administration is seized with AI's security risks and vulnerabilities.
- 11:02SPEAKER_00: Like I said, I think the Mythos moment freaked out a lot of people in the administration.
- 11:09SPEAKER_00: And I don't think that's without good reason.
- 11:14SPEAKER_00: I am also worried about the cybersecurity implications,
- 11:18SPEAKER_00: as well as uplift for malicious actors when it comes to Synthetic biology.
- 11:24SPEAKER_00: I think a lot of your listeners probably saw this New York Times piece from
- 11:28SPEAKER_00: the past couple of weeks outlining how AI has been used
- 11:33SPEAKER_00: to synthesize novel pathogens in the real world for the first time.
- 11:38SPEAKER_00: And so I think we're going to keep seeing quite disruptive effects from AI.
- 11:45SPEAKER_00: At the same time, the concern is how do we make sure that we stay ahead
- 11:53SPEAKER_00: How do we make sure that our frontier capabilities are the best in the business?
- 11:58SPEAKER_00: The pacing letter that 1,300 Frontier Lab employees signed was an interesting kind
- 12:06SPEAKER_00: of counter signal to what we often talk about with the US-China AI race.
- 12:11SPEAKER_00: The fact that slowing down development risks seeding frontier advantage.
- 12:19SPEAKER_00: But I don't know where this goes.
- 12:21SPEAKER_01: What do you make of the pacing letter?
- 12:23SPEAKER_00: Oh, man,
- 12:24SPEAKER_00: you know,
- 12:25SPEAKER_00: our colleague Shane Tews just had on some good friends,
- 12:28SPEAKER_00: including Adam Thierer from the R Street Institute.
- 12:31SPEAKER_00: And I think a lot of people are rightly concerned about slowing down America's advantage.
- 12:36SPEAKER_00: You know, the point of a frontier AI company is to build frontier AI.
- 12:42SPEAKER_01: I will say there are, I mean,
- 12:42SPEAKER_01: certainly there are some analysts when they look at these issues,
- 12:46SPEAKER_01: their sort of working assumption is They're not concerned
- 12:51SPEAKER_01: about like frontier capabilities.
- 12:53SPEAKER_01: They very much look at this as like,
- 12:56SPEAKER_01: you know,
- 12:56SPEAKER_01: another technological general purpose technology and why are we slowing down?
- 13:02SPEAKER_01: And I tend to think that for a lot of those people,
- 13:06SPEAKER_01: like these kind of frontier concerns
- 13:09SPEAKER_01: where you have some other analysts saying we need to audit the models and like
- 13:12SPEAKER_01: that doesn't really play like into their worldview.
- 13:18SPEAKER_00: I think that's right.
- 13:20SPEAKER_00: I wrote on Twitter some weeks back some low confidence predictions about the future.
- 13:25SPEAKER_00: Those
- 13:26SPEAKER_01: are great ones.
- 13:26SPEAKER_01: Those are my favorite guys.
- 13:29SPEAKER_00: I love, as every good AI policy analyst, I love opining about what could come to be.
- 13:38SPEAKER_00: But I wrote a couple of things.
- 13:40SPEAKER_00: One, it's very clear that A handful of hyperscalers
- 13:45SPEAKER_00: and large labs are pretty pot committed to building large compute clusters at least
- 13:51SPEAKER_00: over the next six months to 12 months.
- 13:54SPEAKER_00: And so really the question is what kinds of technical capabilities will emerge
- 13:58SPEAKER_00: from those investments?
- 14:01SPEAKER_00: Two,
- 14:02SPEAKER_00: I think that large clusters of Blackwell class GPUs Probably will be good enough
- 14:08SPEAKER_00: to squeeze out another order of magnitude improvement in Frontier AI capabilities
- 14:14SPEAKER_00: through the automation of AI research like we've been talking about,
- 14:18SPEAKER_00: maybe even before the introduction of Vera Rubin chips circa 2028.
- 14:23SPEAKER_00: Three, if that's true,
- 14:26SPEAKER_00: then that could mean we will really soon start seeing really disruptive effects
- 14:31SPEAKER_00: from Frontier AI, like the autonomous hacking of Hugging Face last week.
- 14:37SPEAKER_00: Or a few weeks ago by an unreleased OpenAI model.
- 14:42SPEAKER_00: And then lastly, four,
- 14:44SPEAKER_00: it is true that Frontier Labs will continue emphasizing AI's catastrophic risks
- 14:51SPEAKER_00: to justify AI's careful and centralized development,
- 14:56SPEAKER_00: while their competitors will continue downplaying AI's rapidly improving
- 15:00SPEAKER_00: capabilities and the legitimate risks it might bring.
- 15:03SPEAKER_00: I think both of these things are true at the same time.
- 15:07SPEAKER_00: As far as I'm concerned,
- 15:08SPEAKER_00: there's really no denying that we are living in a very different world
- 15:11SPEAKER_00: than we were six months ago.
- 15:13SPEAKER_00: And the prospect of a completely autonomous cyber compromise is one we will have
- 15:20SPEAKER_00: to contend with where we haven't before.
- 15:24SPEAKER_01: In your paper,
- 15:25SPEAKER_01: Voltcraft,
- 15:25SPEAKER_01: which we will link to, what is the core of the competition that that paper is about?
- 15:35SPEAKER_00: Yeah, so we've just released this,
- 15:37SPEAKER_00: thanks for mentioning it,
- 15:38SPEAKER_00: this very big project that's been in the works for about six months that looks at AI
- 15:45SPEAKER_00: as an industrial systems competition.
- 15:48SPEAKER_00: So we've discussed four dimensions, right?
- 15:50SPEAKER_00: Frontier capability, national adoption, international diffusion, and societal resilience.
- 15:55SPEAKER_00: This paper is really about international diffusion of AI,
- 15:58SPEAKER_00: and we try to suss out,
- 16:00SPEAKER_00: you know,
- 16:01SPEAKER_00: How much computing power will be available to enterprises in the United States,
- 16:06SPEAKER_00: to enterprises in China, and where will it be installed around the world?
- 16:11SPEAKER_00: We basically make this argument that
- 16:14SPEAKER_00: Although frontier capabilities are democratizing,
- 16:17SPEAKER_00: they're getting more efficient with mixture of experts architectures,
- 16:20SPEAKER_00: they're becoming less compute intensive to run.
- 16:24SPEAKER_00: The fact is that to make the most out of frontier AI and do really disruptive stuff
- 16:30SPEAKER_00: with it and to run it at scale,
- 16:32SPEAKER_00: you still need tons and tons of devices that turn electricity into tokens.
- 16:36SPEAKER_00: You still need tons of compute.
- 16:38SPEAKER_00: And really,
- 16:39SPEAKER_00: I think the name of the game is going to be who can build
- 16:41SPEAKER_00: and install Computational power around the world.
- 16:45SPEAKER_00: And so the paper,
- 16:46SPEAKER_00: which we've just written for AEI,
- 16:49SPEAKER_00: tries to estimate how much compute will be produced by companies in
- 16:52SPEAKER_00: the United States and by companies in China.
- 16:55SPEAKER_00: And we basically find that the United States currently has a massive lead,
- 17:00SPEAKER_00: about a 20x lead over China's computing ecosystem.
- 17:04SPEAKER_00: A permanent lead?
- 17:05SPEAKER_00: Is it a permanent lead?
- 17:07SPEAKER_00: I think,
- 17:07SPEAKER_00: you know,
- 17:08SPEAKER_00: the lead is Likely to last for a while,
- 17:13SPEAKER_00: but it's narrowing is the best way I would put it.
- 17:16SPEAKER_00: We estimate that by 2028, China could grow capable of meeting a third of its own compute demand.
- 17:24SPEAKER_00: And the US system currently makes about 20 times more compute than China.
- 17:29SPEAKER_00: By 2028, it could make So, there is a rate
- 18:02SPEAKER_01: Which China will be catching up to,
- 18:05SPEAKER_01: in which each country, which they will be able to export, compute.
- 18:12SPEAKER_01: And you would rather have us be the country exporting the compute.
- 18:16SPEAKER_00: A hundred percent.
- 18:17SPEAKER_00: This is really, I think, the area that I'm most concerned with.
- 18:21SPEAKER_00: If compute is the substrate of the global intelligence economy,
- 18:25SPEAKER_00: we want to make sure that it's a US designed compute
- 18:28SPEAKER_00: that other countries are choosing to buy and install.
- 18:32SPEAKER_00: Rather than compute manufactured in China.
- 18:34SPEAKER_00: This is because of inordinate security risks that we know emanate from hardware made
- 18:38SPEAKER_00: in the PRC, which are likely to follow compute infrastructure.
- 18:43SPEAKER_00: Plus, we also know that there are follow on effects where if you buy
- 18:47SPEAKER_00: an American made chip,
- 18:48SPEAKER_00: you're more likely to train your developers to program with it.
- 18:52SPEAKER_00: And those developers are more likely to build applications that make use
- 18:56SPEAKER_00: of AI software built in design in the United States.
- 18:59SPEAKER_00: This is the idea behind selling an American tech stack.
- 19:03SPEAKER_01: And what is the role in government there?
- 19:07SPEAKER_00: This is a great question.
- 19:08SPEAKER_00: The Trump administration has taken a very forward-leaning role in trying to export
- 19:14SPEAKER_00: an American tech stack.
- 19:16SPEAKER_00: It's announced an American AI exports program led by the Commerce Department.
- 19:23SPEAKER_00: People have different ideas about what the role of government should be and how much
- 19:27SPEAKER_00: of this should be left to the free market.
- 19:30SPEAKER_00: My concern is basically that China is poised to run the same playbook it's run
- 19:35SPEAKER_00: in every other industry,
- 19:36SPEAKER_00: electric vehicles,
- 19:38SPEAKER_00: solar cells,
- 19:39SPEAKER_00: steel and aluminum,
- 19:40SPEAKER_00: etc., which is propping up the production of massive quantities of a certain good.
- 19:48SPEAKER_00: With state-led subsidies,
- 19:50SPEAKER_00: undercutting everybody else in the market,
- 19:52SPEAKER_00: achieving at scale volume,
- 19:56SPEAKER_00: And then crowding out every other competitor before jacking up prices.
- 20:01SPEAKER_00: This is,
- 20:02SPEAKER_00: I think,
- 20:02SPEAKER_00: something that we could see happen in high-end computational power probably
- 20:07SPEAKER_00: the next 10 to 20 years.
- 20:10SPEAKER_00: People have different timelines as to when China will achieve at-volume production
- 20:15SPEAKER_00: of Frontier Compute.
- 20:17SPEAKER_00: We think it's going to achieve at-volume production pretty soon,
- 20:21SPEAKER_00: probably within the next five or six years.
- 20:23SPEAKER_01: They don't need, right now, they don't need to sell the best.
- 20:27SPEAKER_01: They need to sell good enough at a reasonable price.
- 20:33SPEAKER_01: So we have this time period, which they can't really do that.
- 20:40SPEAKER_01: And during this time period, we should be doing what?
- 20:43SPEAKER_00: That's right.
- 20:44SPEAKER_00: Right for the next...
- 20:46SPEAKER_00: Three years at least.
- 20:47SPEAKER_00: Really, it's the US-led compute ecosystem that has a monopoly on high-end compute production.
- 20:54SPEAKER_00: And so we basically have to make difficult choices as to who
- 20:58SPEAKER_00: to sell high-end compute to as it continues to be a seller's market and
- 21:04SPEAKER_00: in ridiculously high demand.
- 21:07SPEAKER_00: Probably we want to sell a good chunk of that compute to enterprises in the United States.
- 21:12SPEAKER_00: We want American companies to reap the benefits of AI.
- 21:16SPEAKER_00: At the same time,
- 21:17SPEAKER_00: we need to serve certain markets that are going to become really important partners
- 21:21SPEAKER_00: to the United States that have lots and lots of energy to spare,
- 21:25SPEAKER_00: that can actually install some of the chips that are rolling out
- 21:28SPEAKER_00: of our fabrication facilities before China's alternative comes online.
- 21:34SPEAKER_00: We think that this basically boils down to a handful of partners in the Gulf,
- 21:40SPEAKER_00: in Southeast Asia,
- 21:41SPEAKER_00: in Japan, and in a couple of other developed economies with reliable electric grids.
- 21:46SPEAKER_00: But it really will depend on a case by case basis
- 21:50SPEAKER_00: of what security measures those economies are willing to introduce to make sure
- 21:54SPEAKER_00: that compute isn't surreptitiously sold to Chinese enterprises or made available
- 22:01SPEAKER_00: to China's security institutions.
- 22:06SPEAKER_01: I'm assuming that you don't think this will just happen and that there needs to be
- 22:09SPEAKER_01: a government policy steering these chips to some of these markets.
- 22:14SPEAKER_01: Is that what you're saying?
- 22:15SPEAKER_01: That we need to be nudging,
- 22:17SPEAKER_01: nudging the sellers of AI infrastructure to these markets in some fashion.
- 22:24SPEAKER_00: It's a bit of both.
- 22:25SPEAKER_00: We see competing pressures in one direction or another.
- 22:29SPEAKER_00: I'll give you an example.
- 22:30SPEAKER_00: Last year,
- 22:32SPEAKER_00: there was a big debate about what was called the GAIN-AI Act,
- 22:35SPEAKER_00: which originally was a piece of legislation meant to keep high-end compute sales
- 22:41SPEAKER_00: within the United States to make sure
- 22:42SPEAKER_00: that American hyperscalers would be served first over foreign competitors.
- 22:47SPEAKER_00: That bill was ultimately changed to facilitate the sale of compute Abroad
- 22:53SPEAKER_00: to US hyperscalers building large projects in other countries.
- 22:59SPEAKER_00: And this, I think, reflects the reality of what I'm trying to describe.
- 23:02SPEAKER_00: There are certain constituencies around the world where we can expect
- 23:07SPEAKER_00: the developer ecosystem in Singapore is going to take off like wildfire
- 23:11SPEAKER_00: and start building and using a lot of AI applications.
- 23:14SPEAKER_00: Maybe we want to make sure that Singaporeans are building on American compute Just
- 23:20SPEAKER_00: so I'm clear, so would this mean building?
- 23:50SPEAKER_01: Data centers in other countries that would be run by companies in those countries?
- 23:57SPEAKER_01: Or would these be American hyperscaler data centers that were located in other countries?
- 24:04SPEAKER_00: Yeah, this is another big part of the fight.
- 24:06SPEAKER_00: If I had it my way,
- 24:09SPEAKER_00: I think a preferable policy outcome is to make sure
- 24:11SPEAKER_00: that it's American hyperscalers building data centers that they have some kind
- 24:16SPEAKER_00: of operational influence over.
- 24:19SPEAKER_00: In central nodes around the world and making sure
- 24:23SPEAKER_00: that those data centers can serve American AI workloads,
- 24:26SPEAKER_00: or at least have a fraction of their capacity dedicated
- 24:30SPEAKER_00: to serving American AI workloads.
- 24:32SPEAKER_00: One of the big problems that I see,
- 24:35SPEAKER_00: and you've had a number of recent guests on to discuss,
- 24:38SPEAKER_00: is the fact that people don't like data centers in the United States
- 24:40SPEAKER_00: and they don't want to build them here.
- 24:42SPEAKER_00: And so I want to make sure that America's compute needs are met,
- 24:45SPEAKER_00: possibly even through data centers located overseas.
- 24:49SPEAKER_01: These data centers are going to get built somewhere.
- 24:51SPEAKER_00: That's right.
- 24:52SPEAKER_00: These data centers are going to get built somewhere.
- 24:55SPEAKER_00: And I'd rather basically that it is American compute getting chips in sockets where
- 25:02SPEAKER_00: the sockets are available between now and about 2028 and 2029
- 25:06SPEAKER_00: when China can conceivably serve them.
- 25:10SPEAKER_01: What does the world look like where we don't do anything?
- 25:13SPEAKER_00: This is a pretty scary world.
- 25:16SPEAKER_00: I'm old enough to remember the inexorable march of Huawei's 5G equipment
- 25:22SPEAKER_00: through almost every market on earth when a lot of our friends in Europe and around
- 25:27SPEAKER_00: the world were tempted to install Huawei's 5G base stations is the basis
- 25:32SPEAKER_00: of their telecommunications infrastructure,
- 25:34SPEAKER_00: which presented enormous collection opportunities for China's defense
- 25:37SPEAKER_00: and intelligence services.
- 25:39SPEAKER_00: I think we risk repeating that experience,
- 25:43SPEAKER_00: but in a much more severe and visceral way,
- 25:46SPEAKER_00: as the physical AI hardware,
- 25:50SPEAKER_00: as well as the AI software that Global publics elect to use originates in Beijing
- 25:57SPEAKER_00: or Shenzhen or Hangzhou instead of in San Francisco.
- 26:01SPEAKER_00: That's potentially very scary.
- 26:03SPEAKER_00: It's talking about the kinds of services that people choose to interact with,
- 26:08SPEAKER_00: not just for work,
- 26:09SPEAKER_00: but as personal confidants and with
- 26:12SPEAKER_00: which they're sharing extremely sensitive personal information.
- 26:17SPEAKER_01: Does anybody else worry about this?
- 26:19SPEAKER_01: Does Europe worry about this kind of thing?
- 26:21SPEAKER_01: Do they worry about whether it's a Chinese or American tech stack that dominates?
- 26:29SPEAKER_00: Yes and no.
- 26:30SPEAKER_00: I mean, Europe's got its own answer to this problem,
- 26:33SPEAKER_00: and it's got its own fears about becoming too dependent on either Washington
- 26:37SPEAKER_00: or Beijing.
- 26:38SPEAKER_00: I think there are certain economies in Europe that have convinced themselves
- 26:42SPEAKER_00: that they're going to create a euro stack,
- 26:44SPEAKER_00: a third pole to compete, which would be extremely expensive.
- 26:48SPEAKER_00: And just frankly, I don't think it's going to work out super well.
- 26:53SPEAKER_00: Really, the question before many economies now is A, How to build grid infrastructure.
- 27:01SPEAKER_00: B, how to buy large amounts of AI hardware to install and hook up to that grid infrastructure.
- 27:08SPEAKER_00: C, whose models to procure and which will allow permissive licensing.
- 27:14SPEAKER_00: For now,
- 27:14SPEAKER_00: that's been Chinese open source,
- 27:16SPEAKER_00: which doesn't aggregate data in the same way as US frontier AI systems used
- 27:23SPEAKER_00: by APIs offered by Anthropic and OpenAI.
- 27:27SPEAKER_00: But increasingly, U.S.
- 27:30SPEAKER_00: open-weight competitors are coming online that can be privacy-preserving,
- 27:34SPEAKER_00: that can offer open licenses and data localization.
- 27:38SPEAKER_00: And so those three parts of the tech stack, every country is now racing to figure out.
- 27:45SPEAKER_00: And countries that can't build the energy infrastructure or
- 27:47SPEAKER_00: the compute infrastructure that we're talking about will have to come to rely
- 27:51SPEAKER_00: on either the United States or China to serve their AI needs.
- 27:57SPEAKER_01: I'm going to finish with a national security question.
- 28:00SPEAKER_01: A few years ago,
- 28:03SPEAKER_01: there was a essay that came out by a guy named Leopold Ashenbrenner,
- 28:06SPEAKER_01: whose investment fund just ran into some trouble.
- 28:09SPEAKER_01: And the essay was called Situational Awareness.
- 28:12SPEAKER_01: And one of the scenarios was that as frontier labs pushed the AI frontier
- 28:19SPEAKER_01: into something that could be called AGI or super intelligence.
- 28:22SPEAKER_01: That basically the governments would de facto take over those labs because they were
- 28:27SPEAKER_01: so important and so risky.
- 28:31SPEAKER_01: Is that a scenario you can imagine?
- 28:34SPEAKER_01: How realistic is that?
- 28:36SPEAKER_00: Man, what a question.
- 28:40SPEAKER_00: I'll be provocative, and I'll say yes.
- 28:44SPEAKER_00: I think that we are already seeing some shadow of this, even if it isn't declared.
- 28:49SPEAKER_00: I think the US government has been able to use its hefty contracting authorities
- 28:54SPEAKER_00: to push and cajole Frontier Labs into signing security partnerships
- 28:59SPEAKER_00: and offering preview access to their products in a way
- 29:02SPEAKER_00: that very few other industries have ever seen before.
- 29:07SPEAKER_00: And maybe the relationship will grow even tighter or even more formalized.
- 29:12SPEAKER_00: I think it's too early to say with certainty.
- 29:15SPEAKER_00: But back when Mythos was released, a couple of reporters asked me the same question.
- 29:21SPEAKER_00: How could it be that the US government would ever feel comfortable having
- 29:25SPEAKER_00: a private company develop what some people are calling a cyber nuke
- 29:30SPEAKER_00: or something like this?
- 29:33SPEAKER_00: I think there's a real point there.
- 29:36SPEAKER_00: Aschenbrenner described what he called the project,
- 29:40SPEAKER_00: capital T, capital P, as a single joint government private sector collaboration.
- 29:49SPEAKER_00: I think we will likely see a continued molding and merging
- 29:54SPEAKER_00: of private sector capability with the resources and infrastructure of
- 29:58SPEAKER_00: the national security enterprise.
- 30:01SPEAKER_00: But the real question to my mind is how the development of the project will coexist
- 30:08SPEAKER_00: with a really robust private sector and the proliferation of little tech,
- 30:14SPEAKER_00: which has come to rely on AI.
- 30:17SPEAKER_01: Great stuff.
- 30:17SPEAKER_01: Thanks for coming on the podcast.
- 30:19SPEAKER_00: Thanks for having me, Jim.