AI’s expanding attack surface

Reid and Aria unpack the geopolitical battle over chips, from U. S. export controls to China’s push for self-sufficiency, and how the race for compute is reshaping global power. They then turn to how AI is rapidly expanding the attack surface, driving more frequent breaches and exposing new vulnerabilities deep in the software stack as speed and scale outpace traditional defenses. Finally, they explore why enterprise AI adoption has been slower and more uneven than expected, and how network effects, organizational inertia, and trust constraints are shaping the path forward. Together, these forces show how AI is not just advancing technologically, but quietly transforming the foundations of security, competition, and economic power.

• 0:00Speaker 00

  Reid, it is a delight to be back here with you again. Thanks for being on. We are once again going to be talking about AI and specifically about chips. So according to Reuters, data shows that Chinese firms now control about 40% of their domestic AI chip market, and that's up from roughly 35% just a year ago. So not huge changes, but Nvidia's share has fallen to the mid 50s, and analysts are expecting for Chinese homegrown chip production to keep increasing. Obviously, US export controls have accelerated Chinese push for self-sufficiency. Companies like Huawei, Alibaba, Baidu, they're building their own chips and creating entire AI ecosystems around them. Chinese firms are already optimizing

• 0:44Speaker 00

  models to run efficiently on domestic hardware, even if the raw performance sometimes still lags by several years at the very cutting edge. And so when people are thinking about AI hardware chips, how should they be thinking about this? Is this just a supply chain issue? Or is this more about this is the foundational layer of geopolitical power? Like, how important are chips right now?

• 1:09Speaker 01

  Well, the short answer is chips are very important, although not decisive to the end. People have this tendency to be extremists. Either chips don't matter at all, because in fact, you know, China is going to do its own, you know, kind of chips, or chips matter everything, nothing else matters. And as all of these important things, it is a very important factor, but not the only factor that matters. And there's a lot of reasons why the leading chips, which are, you know, currently Nvidia as the key leading chip, really matters is because its compute density, it's because it's a question of as you get to larger and

• 1:50Speaker 01

  larger and less effective ecosystems, less efficient like chip ecosystems in terms of how it operates. You get more probabilistic failure in the system. You get longer time horizons in terms of training. You might even have certain capabilities. There are a lot less, and that's why the leading chips do in fact matter. And that's the reason why, for example, all of the US players who are leading this do have at minimum intense training clusters with Nvidia. And so that really matters. Now, there's a set of things that then come after, come off of this, though, which is,

• 2:33Speaker 01

  for example, China has had some leading innovations in efficiency of compute. And part of, you know, we are learning from that within the Silicon Valley ecosystem. It's one of the reasons why, you know, the Chinese government has, generally speaking, started imposing certain kinds of checks and controls because they don't want to have, you know, kind of Chinese leading edge IP leaking to the West. And so, you know, you've got some challenges there. You've also got, you know, kind of the questions around like, you know, how do open source models work? You've got the questions around distillation. You know, as far as we can tell with Qwen, Kimi and other leading Chinese models, there's

• 3:17Speaker 01

  actually a lot of distillation that happens from OpenAI, Anthropic, Gemini, Copilot, et cetera, in terms of these. So you've got that, you know, kind of playing out as well. So the short answer is chips matter. The short answer is the TSMC generated chips, which include Nvidia, include TPUs, include, you know, other leading edge chips, you know, really does matter. But it's only one major factor around a set in how AI future plays out.

• 3:56Speaker 01

  Now, the geopolitics of it are, we are, you know, beginning to move from the talk of AI and compute fabrics being geopolitical power to the reality of it. Like, it's this year, next year, where that will begin to show some really substantive questions, whether it's anything from You know, obviously, we saw this kind of dust up around, you know, the Pentagon. And what is the compute matter there with Anthropic and others? But also, like, what does this mean for industry adoption? And what's actually happening? Because as you see the coding revolution kickoff,

• 4:41Speaker 01

  the enterprise and work transformation will be in various, what we call in the industry jagged edge, will be also picking up speed this year. Now, I don't think you're going to see layoffs this year because of it. I think you'll see people claiming layoffs because of it. But I think what we've already seen in software engineering is much more Jevons paradox, which is as these tools have gotten efficient, it just increases demand pace for what there is for software engineering.

• 5:14Speaker 00

  Absolutely. And so going back to a little more of the geopolitics, does this, as sort of China and the US become these poles, are people going to have to take sides? Like, as we go, is Europe, are companies throughout Europe, throughout South America, throughout Asia, are they going to have to decide where they're getting their chips? And does that lead to sort of more fractured society as opposed to closer collaboration? Because you're going to have to choose.

• 5:37Speaker 01

  I don't think you're necessarily going to have to choose. I think one of the things when it gets to the rest of the world, the fact that they can, you know, bid them against each other is good. And I think this is part of the reason why, you know, a kind of a international trade policy of tariffs and threats and retaliation is terrible, because that will then offset some natural advantages we had, which is a higher trust than the Chinese ecosystem relative to, you know, the kind of the US setting, the global platform, the US being the, you know, having the companies who are the providers of these things. And I think that the, you know, last year of alienating, you know, kind of friends and partners and allies as a general strategy, whether it's, you know,

• 6:22Speaker 01

  tariff threats or tariff actualities, threats on Greenland, you know, speeches to say, you know, piss off, you know, etc. All of which will mean that this, that we move much more rapidly to a bipolar or multipolar provider of these. And I think that that's going to be true for chips. True for data center architecture and true for software, all of which we would want to be as close to a US technological ecosystem as we can for the economic prosperity of the US and for the economic prosperity of US companies.

• 7:07Speaker 00

  It feels like someone needs to send the White House a copy of How to Win Friends and Influence People, and maybe they can take some hard-won lessons. All right.

• 7:16Speaker 01

  That's if they read.

• 7:18Speaker 00

  Yeah, exactly. Well, yeah, yeah, yeah, we're starting slow. All right, so moving from geopolitics to actually another concern that people talk about with AI, which is cybersecurity. And so, anyone who's been reading the news, or at least my Twitter feed, was full of hot takes about what's happening in cybersecurity recently, because we've seen a wave of breaches and leaks, and this is some big names in the space. So, we saw that Merkle got hacked, and the attackers claimed they stole four terabytes of data. They said the breach was tied to the compromise of Lite L L M, which was an open-source tool in the AI stack, meaning the company was hit through a supply chain attack rather than a direct hack on its own perimeter.

• 8:01Speaker 00

  At the same time, we also saw recently that Anthropic accidentally leaked more than 500,000 lines of Claude Code source code through a bad MPM release. And so, they were exposing internal implementation details, unreleased features. A lot of people are talking about what does being sort of secure mean in this new world? So, how should founders and operators right now be thinking about security in a world where it seems the attack vectors are endless? And we always talk about, yes, AI helps secure against cybersecurity, but it also is increasing the ability to have these attacks.

• 8:39Speaker 01

  So, I think this is just the beginning. And I think that part of the question is, is, you know, it's almost, there's a couple of reasons why cybersecurity opens up much more intensely in the age of AI. One simple one is, now that things are moving so much faster, that there isn't as much time for, like, the previous kind of iterations about how one did security. So, the fact that there is all these tools that are being deployed in production and everything else, where you're like, well, actually, in fact, The security like intensity of of kind of like, you know, red teaming, multiple attacks, et cetera, et cetera, hasn't happened yet. So I think there's just the speed. Second

• 9:22Speaker 01

  thing is, obviously, on a set of AI things, including, you know, how one deploys L L Ms, even if one is deploying open source models of L L Ms in one's own production environment, is that these L L Ms are inherently insecure because they're probabilistic systems that we don't understand that well. And so even though we do a whole bunch of alignment training, you know, part of what happens with even the frontier models, which do an intense amount of alignment training through their chat interfaces, is we still get them doing odd things, which can include, of course, cyber attacks or other kinds of things as ways of doing that. And so there's, there's actually, in fact, some

• 10:07Speaker 01

  new surfaces in the software stack platform that are unclear how you secure them. And this is, of course, within the standard of cybersecurity, which is inherently never fully secure. I mean, the way that they, that you, that you make a fully secure system is you air gap it, right? Which is you disconnect it from the network. And, you know, that's, that there's very limited set of systems where would you can do that. And so, so all of this opens up an intensely new, intense kind of new waves of insecurity. And given speed and iteration, new increase, like, like, even though you said, okay, well, we secured that last thing. Well, here's a set of new things that have changed and are now

• 10:52Speaker 01

  insecure, which of course means we're gonna have to evolve how we play the cybersecurity game. Like, you know, is there gonna be agents and tools that are specifically not only on the, you know, the pure code hacking form, but on phishing, you know, because AI generated tools are the, some of the best phishing amplifiers. Well, what about phishing defense? And how does that play out? And so there's gonna be a whole stack of like new approaches to security and new

• 11:20Speaker 01

  Needs for defense that haven't existed yet. Now, in the venture business is very happy because you know, like Greylock and other places that that are kind of like the the cutting edge of enterprise software, you know, like, you know, Palo Alto Networks and a whole stack of other things. Greylock was at the beginning of, and there's a couple other venture firms that are equally enterprise. You know, cybersecurity tends to be an evergreen category with new companies being started every couple of years. That's really important. And obviously, that's one of the things that, you know, Sheem and Sam and the other folks at Greylock are really focused on.

• 11:57Speaker 00

  And on the other side, how would you think about it as a consumer? So I feel like every week you're getting an email that's like, oh, your, you know, data was breached and there was a leak and your passwords, etc. And then at the same time, you know, we're heading up to April 15th, you see a million vibe-coded apps telling you to, here's how you can do your tax returns with this new vibe-coded app. And, you know, you don't necessarily want to put your W2 into some vibe-coded app that hasn't sort of looked at the security side. How do you think that'll either affect consumer behavior? Or if you were a consumer, what would you do? Would you, you know, sort of trust these new things that were coming along?

• 12:32Speaker 01

  Well, so fundamentally, consumers are not terrifically informed in these areas. It's like, for example, it's one of the reasons why, generally speaking, if you're in a mall and you you have a have a Ferrari sitting in the mall, and you say, here is a five-page fill-out of all of your personal form, like your children's names, you know, every single place that you've lived in your last, you know, in your whole life and all the rest, and you go, okay, I'll fill it all out because I want the chance of the Ferrari. And you're like, sure. What are they doing with all of this data? Because the way that, because it's an economic model by which the Ferrari is the expense by which they sell all this data. And consumers broadly, like, you know, maybe 10 or 20% of them understand this,

• 13:15Speaker 01

  but the majority does not.

• 13:18Speaker 00

  Totally.

• 13:18Speaker 01

  And so what they generally look for is companies to keep them safe. And if they encounter something, to have that be trusted. Some of this may end up becoming government regulation. That's part of the reason why you've had government regulation on, you know, finance and other kinds of things for for for how this plays out. And it may need to be there in some ways. But like, you know, previously, like, you know, the number of just call it.

• 13:48Speaker 01

  You know, train wrecks in cybersecurity and data that you see even from, you know, high-quality companies like Merck or Anthropic. Then you've got all the vibe-coded apps that people are putting up. And like, you know, like dating apps where all the information has suddenly been leaked and posted and all the rest of this stuff. And so I actually think that part of what it's going to end up being is that people are going to go, all right, I, I might trust the startup, but the startup is going to have to go harder to establish its trustworthiness credentials.

• 14:26Speaker 00

  Yeah.

• 14:26Speaker 01

  Because I think it'll grow over time that people will be much less, you know, kind of like, sure, whatever I encounter on the internet or whatever I encounter on my mobile app, that'll be fine and the data will be fine. And now I think there'll be a much, you know, I think there'll be a growing concern. And I think that we may need to actually even, you know, put some extra juice into growing that concern because of how insecure the environment's becoming.

• 14:56Speaker 00

  Yeah, it'll be interesting to see if, if this, we obviously think that, you know, startups, AI-native companies are going to be the ones who sort of win everything here. And yet I do think sort of brand, trust, security, like that will become more of an issue as we sort of hear more of these horror stories. And to your point, maybe people should be more worried than they are, but we will see. And so stepping back from security, it's actually really interesting if you look at the divide between consumer and enterprise. So on the consumer side, I mean, ChatGPT in particular, but all of the AI chatbots have, have been adopted at enormous pace. I mean, faster than social media, faster than sort of all other technologies that we've seen.

• 15:37Speaker 00

  But I think that people expected this spread to happen faster in the enterprise space. We expected sort of companies to take it on more quickly. And in practice, it's been slower and more uneven. You see pockets of people going crazy and using it intensely, but I think other companies are unchanged from five years ago. And, and people probably underestimate how long it takes for these new technologies to sort of go through these enterprise systems. So. Why do you think, like, what do you think that people underestimate about how new technologies diffuse through large organizations? And can we see this pace? Will it speed up, slow down, continue as it has been?

• 16:19Speaker 01

  Well, one of the things that I think is kind of funny is obviously in the valley, you know, everyone says network effects and relatively few people understand it in depth, even in the canonical places of network effects, which is like, you know, everything from fax machines, the messaging clients, the social networks, et cetera, marketplaces. And they don't track, like, for example, other kinds of network effects. Like, for example, perhaps the first and most important economic amplifier network effect is cities. Like, you build the technologies, you can make villages and cities. And now, not only does that allow all the different things like specialization, but it also allows a massive amount of economic

• 17:00Speaker 01

  productivity, not just because of the Adam Smith specialization, but because all this kind of trade and knowledge and information can all spread. And that's, of course, the reason why, you know, we've just gone through our recent set of idiots declaring peak Silicon Valley, Silicon Valley over, you know, whether it's Florida or Texas or whatever else. It's like, no, you don't understand network effects. Silicon Valley has network effects. Well, the same thing is true of companies. Companies have network effects. Now, generally speaking, some of these network effects are very positive, but network effects are kind of a portion of lock-in. And so, part of the reason why work transformation within companies tends to be more slowly is like you say, well, the work that locks in a company and makes,

• 17:43Speaker 01

  you know, even back in the day, forward with the Model T more effective, is it locked in a set of network effects in terms of how the company operates, which means the transformation's difficult. So unsurprising, when you look at, you know, kind of transformation of work.

• 17:57Speaker 01

  Right now, that transformation of work happens most intensely in startups, mostly intensely in small groups in big companies that are kind of doing their own thing. They've hopped on their AI ATVs and they're kind of going through the dune on their, you know, on their cognitive industrial revolution buggies, you know, kind of doing things. And, and the temps for companies to say, well, I, I, I do approve of concept and I have three people doing it's like, that's not, it only works if in a coherent work group. Now that being said, part of what I think is interesting about happens in transformations, network effects and in transformations like this, is it's slow and then fast.

• 18:39Speaker 01

  And the fast is we're now moving entirely to this new network paradigm and we're abandoning the old one, we're changing the new one. And I think it will happen, but it happens more slowly than people predict because it's first slow, then fast. And so that's the, the pattern that we should be thinking. And now, you know, like, when does it be, you know, whatever it was a couple of years ago saying, hey, when I was saying, hey, everyone's going to have their own coding assistant for doing the work, that will happen. Obviously, the coding assistant is an amplifier. So like, when you have coding assistants, a long version thinking and compute as applied to your particular problem, which might be law

• 19:21Speaker 01

  or accounting or finance or anything else. And I think that's part of, you know, how that operates. But anyway, so that's, I think, the, the, the, the slow than fast is a feature of network effects.

• 19:35Speaker 00

  I also think it's really interesting because some, one thing that some people don't realize is one of the reasons why Silicon Valley was successful was because non-competes weren't allowed. And so you had this diffusion of folks to other companies or starting their own or sort of bringing their knowledge with them. And still in so many parts of the country, it's actually those non-competes which slow down innovation. And so not that they're going to topple Silicon Valley, but we actually could see greater innovation in other cities and states if we got rid of those non-competes so that people could take their knowledge with them, whether it's on AI or something else.

• 20:07Speaker 01

  Well, 100%. And some of this stuff was some of the good stuff that Lina Khan was doing. And so I think that having, because it's redefining the network as opposed to having the not, the anti-compete or non-compete as keeping you as a structural lock-in per company, it's opened it up to the local ecosystem. And you know, Annalise Exinean's book Regional Advantage covers this really well from a viewpoint of, you know, venture capital was invented in Boston, a bunch of technical universities, venture capital, et cetera, et cetera. Why did Silicon Valley outstrip Boston? It was because it actually loosened the ability to lock in network effects as companies in terms of the anti-compete and spread it

• 20:50Speaker 01

  to the region, the regional advantage. And, you know, just to kind of, you know, add to my earlier, like, oh, yeah, these idiots describing peak Silicon Valley. My own point of view has been, we want as many Silicon Valleys as we can have, both within the US and within the Western world. And so I'm supportive whenever, you know, when Mike Bloomberg came out to say, how do I, you know, do Silicon Valley in New York, or, you know, various governors. And so that notion of trying to help as many Silicon Valley areas start as possible is, I think, actually a really good one. It's just that you've got to presume the density of the network effect of the existing Silicon Valley in your strategy.

• 21:34Speaker 00

  Absolutely. Reid, thank you so much. Appreciate it.

• 21:37Speaker 01

  Always a pleasure. Possible is produced by Palette Media. It's hosted by Ari Finger and me, Reid Hoffman. Our showrunner is Sean Young. Possible is produced by Tanasi Delos, Katie Sanders, Spencer Strasmore, Imo Zoo, Trent Barboza, and Tefadzwa Nirmurundwe.

• 21:53Speaker 00

  Special thanks to Surya, Yala Monchili, Sayida Sapieva, Ian Alice, Greg Beato, Parth Patil, and Ben Ralas.