Nvidia CEO hangs out with Chinese spy? + China's new AI Export Controls + The AI/cyber month from hell + AI in the midterm elections

0:10 Welcome back to the AI Power Podcast. Thank you for joining us. I'm Adam Goodwin, joined as always by our host, Greg Allen. Hey, great to be with you, Adam. So, Greg, the front page story on the New York Times paper of record right now is that Jensen Wong, the CEO of NVIDIA, has been the target of a Chinese spy plot.

0:29 Greg, is this a movie come to life? What is going on here?

0:33 Well, I think the most movie-esque version of this story is probably not true. I'm seeing some commentary on Twitter suggesting that this is a honeypot sex scandal, that this woman has seduced Jensen Huang. And while the article in the New York Times does discuss the two having a close relationship, that's also the term that is used to describe the relationship between Jensen Huang and President Trump.

1:02 So this woman, a Chinese national, who, according to the reporting in the New York Times, is under surveillance by the U.S. intelligence community and assumed to be a Chinese spy. She is presumably a Chinese spy based on this reporting, but there is no evidence that this is a sex scandal of any kind.

1:23 And the folks who only read the headline and didn't read the full article are missing that. But what there is a lot of evidence about is number one, that the company, which she previously identified herself as the CEO of Megaspeed, which is a Singapore headquartered company, but has its origins as the spin out of a Chinese gaming company, this is a company that very, very serious questions about their role in the smuggling of NVIDIA chips to China have been repeatedly raised.

1:58 And what this time story now brings to the equation is the allegation that she is in fact a Chinese spy and giving information to the Chinese intelligence services, presumably based on her conversations with Jensen Huang, who, as the article also goes into some depth, has a close relationship with President Trump.

2:24 So one way of reading this story is that this woman could be a valuable source of insight to the Chinese government, to Chinese intelligence services, as to what's really on Trump's mind when it comes to the AI race, competition with China, et cetera, et cetera. I think that that probably is true, but Twitter is also a great source of insight into President Trump's mind on all of these issues.

2:52 In so many ways, he is an open book on this story. And so I think the actual focus of our attention really needs to be on the chip smuggling part of this story. And so while this was a great story and I commend the journalists involved in getting this information and in making it available to the American public, I think we need to go back to some of the stories that were written about Megaspeed by Bloomberg and the New York Times that had a bunch of really good insights into this.

3:23 So here's the story from Bloomberg on December 22nd, 2025, lead author Mackenzie Hawkins, who has broken a ton of stories over the years about export control smuggling. And I'm going to read a long quote from this piece just because I think it is so extraordinary. In an industry run by Titans, a Singapore AI firm called Megaspeed is drawing an unusual amount of attention.

3:49 In less than three years, the once obscure spin-off of a Chinese gaming enterprise has evolved into the single largest Southeast Asian buyer of NVIDIA corporation chips. A rapid ascent that's lifted the hopes for a regional champion in the lucrative business of AI cloud computing. Bloomberg didn't find evidence that any of Megaspeed's NVIDIA chips have been diverted to China, or that the company has otherwise violated U.S.

4:14 export controls or other laws. Still, a review of hundreds of pages of trade records, corporate filings, job postings, and private documents from Megaspeed and its partners, plus interviews with dozens of people across six countries, including U.S. and foreign government and industry officials, as well as technical experts such as Semi-Analysis, revealed inconsistencies in Megaspeed's Southeast Asia chip inventory and overall data center footprint.

4:40 Since its founding in 2023, through November of this year, Megaspeed has imported NVIDIA hardware worth at least $4.6 billion, containing at least 136,000 NVIDIA graphics processing units or GPUs. Then it goes on. More than half of Megaspeed's chips are from NVIDIA's current generation Blackwell lineup, processors that Trump, even after allowing NVIDIA to ship its best previous generation products to China, said he won't approve for export to the Asian country.

5:13 Megaspeed got the majority of its Blackwell processors more than half a year ago in another batch last month. Yet when NVIDIA officials recently visited Megaspeed's data centers, they only saw hardware that contains a few thousand Blackwell GPUs, according to inventory information the chipmaker shared with the U.S.

5:31 government, described by people familiar with the Megaspeed probe. Asked about the rest of the Blackwell chips, an NVIDIA official said the company visited separate Megaspeed warehouses where it saw the hardware imported earlier and confirmed that it hadn't been diverted to China. But this person declined to specify whether the number of components NVIDIA cataloged in storage matches the outstanding volume from Bloomberg's analysis of Malaysian customs records.

5:56 Megaspeed didn't answer multiple queries on the location of and plans for those processors, nor on its facilities writ large. Now, this is the juicy part. On its website, Megaspeed describes three data centers previously identified as being in Malaysia and Indonesia and now all listed as being East-South Asia.

6:18 But the company told potential investors that in addition to those two countries, it was building an NVIDIA computing cluster in a quote, specific area, end quote, of a third unnamed nation, according to a slide deck in Megaspeed circulated in 2024. In the presentation obtained by Bloomberg, Megaspeed's, quote, specific area, end quote, facility was the only one not identified as being in the Southeast Asia.

6:44 To depict the project, Megaspeed used a rendering of a massive AI data center in Shanghai, whose corporate records show was financed in part by Megaspeed's original Chinese parent, plus a holding company created by Megaspeed's founder. Megaspeed also has something of a Chinese corporate twin. This firm used similar presentation materials to Megaspeeds, had a nearly identical website to a Megaspeed subbrand, and claimed Megaspeed's Southeast Asia employees as its own.

7:17 It's also posted job ads at and near the Shanghai Data Center, whose rendering was used in Megaspeed's Investor Deck, including for engineering work on restricted NVIDIA GPUs. That is an amazing piece of journalism. And my compliments to the reporters and my apologies for reading so much of the article.

7:37 Everyone should go read the original piece on Bloomberg's website. But to me, this is glaring red flag after glaring red flag after glaring red flag. And I think what people need to understand is that when there are glaring red flags, there are unique export control responsibilities that companies have to take extra measures to ensure that diversion is not happening.

8:06 Instead of that, after these reports have been out in the public ether for more than 10 months now, you have Jensen Huang and NVIDIA defending the relationship with Megaspeed. And even after today, when these allegations of the CEO, or at least the woman who described herself as the CEO until recently, being a Chinese spy, you're still getting a rebuttal from NVIDIA.

8:32 This is a quote from an NVIDIA spokesperson going to CNBC on October 29th. Quote, the diversion allegations against this regional cloud were thoroughly investigated and debunked last year, end quote. Come on. I mean, come on. Like this is just not credible. It's not an credible interpretation of the pattern of facts.

8:59 And over and over over the past few years, NVIDIA has said that smuggling of chips basically doesn't happen and it's basically impossible. And so I just want to remind folks that number one, NVIDIA is basically calling the US government wrong by saying that something the US government is saying is a factual matter is false.

9:23 So for example, David Peters in February 2026, he's the assistant secretary for export enforcement at the Bureau of Industry and Security. He said at a House Foreign Affairs Subcommittee that he was asked whether chip smuggling is occurring. And he said, simply put, yes, there is chip smuggling. It is going on.

9:42 And Jeffrey Kessler was reported as having said in a letter by uh Senator Cotton and Rep Huizinga in March 2026, he said, quote, we catch significant smuggling and illicit exports, but we're concerned that it's just the tip of the iceberg. Like what Jensen Huang told Bloomberg in October 2025, this was at the Computex conference in Taiwan, he said this, quote, governments understand that diversion is not allowed, and there's no evidence of any AI chip diversion.

10:15 Recognize our data center GPUs are massive. These are massive systems. The Grace Blackwell system is nearly two tons. You're not going to be putting that in your pocket or your backpack anytime soon. And so these systems are fairly easy to keep track of. But the important thing is that the countries and companies that we sell to recognize that diversion is not allowed.

10:36 And everybody would like to continue to buy NVIDIA technology. And so they very well monitor themselves very carefully, and they're quite careful about that. Really, really, they're monitoring themselves incredibly well. Let's just go over some of the things that have come out since then. First, Supermicro, one of NVIDIA's most important distributors around the entire world.

10:58 The co-founder of that company was indicted in a $2.5 billion chip smuggling diversion scheme. Okay, that was in March of 2026. Then the Financial Times reported in July 2025 that at least a billion dollars worth of NVIDIA chips were smuggled after the Trump administration's export controls on Blackwells.

11:22 Then you've got the Megaspeed investigation that we already talked about. We also have an amazing piece of reporting from the information back in December 2025, in which six different sources said that DeepSeek was developing its next model on several thousand banned Blackwell chips, which were installed in data centers in permitted countries and then dismantled and shipped to China.

11:44 Huh, that sounds a lot like what Megaspeed is being accused of here. The Department of Justice had Operation Gatekeeper, which is a $160 million case in December of 2025, where a Houston businessman and his company pleaded guilty, and two Chinese men were taken into custody over at least $160 million in H-100 and H-200 exports.

12:10 The NVIDIA labels were replaced with a fictional Sandkian brand. The case broke the same day the Trump administration approved H-200 sales to China. So NVIDIA is saying that chip smuggling is basically impossible. These two men pleaded guilty to smuggling $160 million worth of chips. Okay. Then there's the fact that NVIDIA personnel are now being arrested over this in Taiwan.

12:43 And we talked about this on the podcast. In July and August of 2026, Bloomberg reported that there was a senior NVIDIA manager and others who were accused of shipping 74 servers with B300 chips into China via Japan and Indonesia in what amounts to Taiwan's first known crackdown on the trade. And then I just want to say that this has been going on for so long.

13:08 The New York Times in August of 2024 spoke with representatives of 11 companies in China that told the Times that they had sold or transported NVIDIA chips. The Wall Street Journal in July of 2024 found at least 70 distributors openly advertising restricted chips. What can you possibly say to be taken seriously when you're saying that diversion is impossible?

13:36 There's nothing to see here. This is red flag after red flag after red flag. And, you know, when you see a forest fire's worth of smoke, and then you say that you don't need to change anything about what you're up to, I think it is absurd and preposterous. And so my hope here is that this story about the connection to the Chinese intelligence services will finally be the straw that breaks the camel's back and allows Congress to take some stronger moves here with its oversight of NVIDIA, with its oversight of the Department of Justice and the Bureau of Industry and Security at Commerce to finally help America get serious about enforcing the laws that are on its book.

14:26 And Senator Elizabeth Warren, who's on the Senate Banking Committee, which has important jurisdiction in this case, she tweeted out today on October 29th this quote, NVIDIA CEO Jensen Huang was reportedly targeted by a suspected Chinese spy whose Singapore-based firm bought billions in AI chips. Huang and NVIDIA's board owe Congress answers.

14:50 And we must pass bipartisan legislation to keep our advanced AI technology out of China's hands. Now I think Warren speaking out on this issue is noteworthy because she had previously called on Huang to testify before Congress. Nvidia refused, and then she released a mean statement about how he had time to gallivant around the world, but didn't have time to come answer questions before Congress.

15:14 Well, this gets to the midterm elections that we're going to talk about at the end of this. The Democrats are projected by forecasters to have a victory in the House and a real chance of a victory in the Senate, meaning that they would have committee leadership and subpoena authority. So I'm hopeful that we can finally start moving the needle on these issues because it is embarrassing the extent to which we are helping China win the AI race.

15:49 I mean, if we are in an AI race, if the stakes of this AI race are every bit as important, or actually I would argue more important than the space race with the Soviet Union, you don't want American companies giving Gatorade and Nike shoes to the other team's racer. All right, Adam, sorry, that was a long one, but this one really got me fired up.

16:11 No. Totally understandable, Greg. And if anything, I think the mistake that was made by the Chinese intelligence services here when they had a failed U.S. policy that was completely enabling them to get exactly what they wanted to, was to do something so provocative that it wound up front page above the fold and got everybody's attention now.

16:30 So certainly what you mentioned about Senator Warren potentially becoming the banking committee chair, dovetailing with the anti-corruption message that seems to be driving a lot of the potential electoral results for next Tuesday would lead to that committee having the mandate to go and do that, or at least try to go out and do that.

16:46 Hey, just a question of the machinery of all this, Adam, a question for you, a congressional insider for many years. How fast can this stuff move? If hypothetically Warren is super gung ho about dragging Jensen Huang in front of the committee, has the majority and therefore has subpoena authority. Can that happen in a day, a month, six months, a year?

17:11 How fast can this move when Congress really makes up its mind to do something?

17:15 Senator Warren obviously has a staff and a very good staff. I think everybody there, including the janitor, is a lawyer in that office. Theoretically, they could get paperwork out the door in the first day, right? There are certain limitations to that, but first day you could have subpoenas go out the door.

17:30 There's not much prohibiting them from doing that once everybody is in the seat and moving. The thing that will really slow that process down, uh, the two things would be cooperation from the person they're trying to subpoena. So in this case it would be Jensen Wong or whoever the leadership at NVIDIA is, and two cooperation from executive branch organizations that have the data for the investigation that the committee would need in order to pursue these things.

17:54 So in a cooperative environment where you have a committee that is doing something that is very tightly aligned with an administration, you could see information moved in weeks, if not days, very, very, very quickly. There's not much that would prohibit that from happening. I think in a place where you're gonna have a non-cooperative executive branch potentially, just because their policy has been to let NVIDIA do what NVIDIA wants to do, and certainly a non-cooperative executive body in NVIDIA between uh Jensen Wong and the senior executives.

18:24 They're gonna lawyer up and go to court, and it could be months, if not years, fighting the congressional subpoena. We we saw that in the first Trump administration. A couple of people decided they would rather simply go to prison on contempt charges than comply with the congressional subpoena. Uh Peter Namaro in particular famously did that.

18:40 So there are lengths to which you can go to fight that, but hopefully things go faster just for the sake of protecting U.S. national interests here.

18:47 Yeah. Well, thanks for that insight.

18:49 Well, Greg, NVIDIA thinks that export controls are stupid and China is smart. So let's address our next topic: Chinese export controls on AI. On October 16th, Beijing proposed putting AI, including models, chip designs, and AI hacking tools on its export control list for the first time. What does this tell us about what China's doing on AI and semiconductors and how's that changing now?

19:14 Well, I think there's a few things that we can infer about what Chinese leadership believes about the current state of AI, the current dangers of AI, and China's competitiveness in AI based on their behavior. You know, the best guide to what China is thinking is usually not what China is saying, it's what China is doing.

19:38 And they did an awful lot in this package of new export controls. And so I'm going to give you my read on what it says about the state of Chinese leadership's belief around AI. So, the first, what did they really do? Well, this package of export controls includes more than just stuff on AI. It includes some stuff on photovoltaics and biomedicine.

19:59 I'm not going to talk about that. I'm only going to talk about the AI stuff, which is 23 out of the new 40 export controls that China has adopted. And they basically created three tiers of the catalog regime for AI technology. And those are first the prohibited tier, the restricted tier, and the not listed freely exportable tier.

20:28 And the sort of vague analogy to the US system would be export license required with a presumption of denial, although in their cases they basically say under any circumstances. So it's even stronger than our version of presumption of denial. Then they have restricted, exportable only with a license, but that's going to get a strict review to make sure it's compliant.

20:51 And then no license required, the third one, which, you know, of course, America also has. So they have this sort of tiering structure. And one thing to know is that this is actually already enforced. So even though it is a draft that they're circulating and requesting for comment, it already applies.

21:10 You already have to comply with it. And it restricts the ability of certain things that you could have, including in deals that are not yet finished. So this could be assigning ownership of a technology abroad, so like selling intellectual property, licensing the domestic technology for use abroad, a change in who owns it, so a foreign entity becoming an owner or actual controller of the company.

21:37 And then literally, I love this later. Anything else? Mofcom later specifies, Mofcom being the relevant ministry in the Chinese government. So I think the third one is perhaps worth starting about. Manus was a notable AI agent startup in China that had attracted a lot of attention as a noteworthy startup.

22:00 They had left China and reincorporated in Singapore and then were subsequently acquired by Meta. China blocked that deal from proceeding. So Manos is still a Singaporean company, but a Singaporean company that is, for all intents and purposes, controlled by the Chinese government to the extent that they decide that they want to take control of the firm.

22:20 And I think a lot of people are looking at this new batch of export controls through the lens of China saying, how do we make sure that never happens again? Right? That maybe technology will go abroad, maybe technology will not go abroad. But the key thing to understand is that China is the one who decides, not private sector actors.

22:41 And I think that assertion of authority over the flow of Chinese technology, and especially in AI, is an important thing to keep in mind here.

22:52 Just to be specific, Greg, what problem does Beijing think they're solving here? I mean, you kind of touched on it, but this is such a huge sea change for how this policy's been conducted before. What are they getting after here?

23:05 Yeah, so I think there's multiple problems that they're trying to solve simultaneously with this. The first is the one we were just talking about with MANUS, which is the flow of advanced Chinese technology where they're proud of it. They actually think China has a meaningful competitive advantage. They want to restrict or at least control other countries' ability to have access to it.

23:26 You know, maybe there'll be tech transfer, maybe there won't be tech transfer, but it's only going to happen with the Chinese governments say so. That's the first thing. The second is there is an element of at least some version of what you might call AI safety in here. And that's most obvious at the intersection of AI and cyber.

23:46 For example, there is an exemption, a broad exemption in this license regime for open source or open weight AI models. You don't have to get a license, or more accurately, it's essentially a notify and go kind of a license. The expected outcome is automatic approval if you're putting out an AI model that you're going to open source.

24:10 But that doesn't apply if the model has advanced AI cyber capabilities. That is now in the prohibited tier. And so you're seeing Chinese model makers effectively trying to give their AI models lombodomies where they don't know anything about cyber because cyber stuff has been removed from the training data set.

24:31 I've said before that I think the Chinese hugging face incident equivalent is happening soon if it hasn't happened yet. This is a data point that's consistent with that interpretation. This is evidence of a Chinese government that is spooked by how powerful these models are getting and some of the incidents that we've talked about in this podcast in terms of what AI models are doing with respect to cybersecurity.

24:55 And so they are now placing restrictions that affect not just the closed source proprietary models in China, but even the open source ones, where previously the restrictions that applied to them had been comparatively modest. So those are two big goals that they're trying to say here: control of the flow of technology, where they can make those choices, not based on a private sector company's ability to make a profit, but based on the Chinese government's strategic perceptions of its own geopolitical interests.

25:23 And then there is a little bit of a safety part of the story too.

25:26 Understood. And just to keep hammering on this, why did Beijing choose export controls as the tool to achieve that level of control? And with these controls, how does it plan to keep AI from flowing out to the world? I think when you're talking about the relationship that the Chinese government normally has with individual businesses in the country, they could very easily just walk in the front door of a company, grab the CEO, and said, this is your policy going forward.

25:48 This is very new for them. This is a very new direction.

25:50 So China has had export controls, most notably on rare earths, squeezing the supply chain of rare earths using export control authorities in pursuit of leverage in negotiations with the Trump administration. And I think they feel like, hey, this is a fun tool. And it's also not unprecedented to be using export controls as a tool of AI competition, right?

26:16 The United States has been doing that for a long time, or even as a tool for safety regulation. When the US government wanted to block the release of Claude's Fable V model, they used export controls to crack down immediately. But I think the bigger reason is that China is, on the one hand, pro broad adoption of Chinese technology, but they don't want something like what has been proposed in the battery market in the United States, for instance.

26:48 Some people who are supporting the idea of bringing Chinese automakers to America, maybe with joint ventures with American companies, some people are making the argument, hey, China has amazing battery technology. Maybe we can, you know, have them set up factories here and force technological transfer over to America.

27:07 And some Chinese companies are like, well, we probably wouldn't want to do that, but let's talk about what the price might be for something like that. And the Chinese government is saying, nope, nope, we don't care what the price is. Companies don't get to make this decision. We get to make this decision based on our perception of our geopolitical interests.

27:24 And then if you look at what they're restricting and the circumstances under which they're restricting it, the 23 AI items fall, you could say, into like eight different buckets. And shout out to China Talk, which had a great piece on this topic on October 21st. But the eight buckets we've got here are like frontier model core tech, AI chip and memory design, technologies related to the integration of large computing clusters, voice print and language AI.

27:56 All of these are things where either the Chinese government thinks that they have some kind of meaningful competitive advantage. I think China thinks that they're getting increasingly good at the secret sauce behind large-scale integration of computing clusters. Other stuff like voice print and language idea, I'm speculating here, but because it includes high-fidelity Chinese and minority language translation, I'm thinking back to where the Chinese government was using AI surveillance technology, both computer vision and voice recognition, to monitor Uyghur populations, whether they were in the Xinjiang home region of that ethnic minority or anywhere in China, just an obsessive level of surveillance

28:42 of the Uyghurs. And when a lot of those AI technical records about just how powerful the surveillance state that China had built were leaked, that was a real egg on face moment for the Chinese government. And so now with export controls, they're kind of trying to lock down that type of information, assert control over it to prevent those kinds of scandals before.

29:04 Again, no evidence that they're going to change their behavior in terms of doing all of these objectionable, unethical things, but they're taking steps to ensure that it's more difficult for those types of data leaks to happen in the future. And then finally, you get to the other four categories of sensitive data-trained models, so anything related to government systems or critical infrastructure systems, cyber capabilities, AI content detection, AI agents, and game AI.

29:34 These are all areas where I think that the Chinese government perceives a real national security risk. You don't want to say that it's illegal to take a bunch of Chinese classified documents and take them out of China and give them to somebody else, and then say it's okay to train an AI model on all of those documents and then walk out of the country with that trained model.

29:57 So there it's just kind of applying the more traditional national security and cybersecurity paradigm to this new universe of computing and data storage. And then finally, the cyber capabilities that I mentioned, where it does appear that China wants more control over systems that are capable of generating malware, that are capable of adapting and evading discovery.

30:23 Those kinds of things are where the licenses, again, are prohibited, and the open source exemption that applies to many of these other kinds of licenses is not eligible for use.

30:34 There's an interesting comparison kind of underlying this move. An export control designed to protect the country's comparative advantage in AI that sounds a lot like the Biden administration's AI export control playbook, right? Yes. How does this compare to the AI diffusion rule from January 2025 that you had a couple of thoughts on at the time and since then?

30:53 Basically, I would say that the underlying strategic rationale of AI being a race, export controls being a useful tool to compete in that race, this is China endorsing that logic full stop. They just see the strategic rationale. Now, they may say that the Biden administration was illegitimate, unethical, paranoid, stupid, right?

31:20 But fundamentally, they're just saying we admire the strategy, right? We think the toolbox is in fact useful for the strategic ends that you were aiming it at. Imitation is the sincerest form of flattery. And notably, this includes some kinds of imitations of many of the mechanisms that were included in the AI diffusion rule, the Biden administration's parting shot on export controls from January 2025.

31:48 So models themselves are a controllable export if they are caught under any one of three tests in the Chinese system. Either it was trained on 10 to the 26 flops, the total model size exceeds 2 trillion parameters, or the model is top 10 as in a global leading capability in any kind of public evaluations that are not specified.

32:14 But if you meet any one of those three boxes, exporting the model abroad does require an export license. And that's really interesting because if you connect that to what the Chinese government is thinking about overseas data centers, if you're a Chinese data center provider or a Chinese cluster provider, you can sell the chips and servers.

32:39 Those are not going to be controlled at this stage. As long as there's not a transfer of core intellectual property around cluster engineering, the model itself, so on and so forth. But if you are transferring a Chinese company to its own overseas data center, there's an automatic license. And to me, that's a remarkable echo of what the diffusion rule sought to do by creating validated end users.

33:07 So Microsoft's facilities in the UAE might be treated differently than just any UAE company that walks in and wants to buy NVIDIA chips if they haven't gone through this validated end user program. So that's really interesting. China's basically saying, like, we trust Chinese companies. They can transfer their technologies to their overseas subsidiaries with an automatic license.

33:31 But anytime you want to transfer the know-how, so imagine you know a Saudi or a UAE company that wants to license some of the secret sauce of Chinese data center engineering or their best models, all of that, now suddenly a license is required. And that is exactly the kind of behavior that China was criticizing and trying to mobilize a global campaign against when it comes to the United States and the diffusion rule.

34:02 So again, imitation is the sincerest form of flattery, and hypocrisy has never been something that the Chinese government is shy about engaging in.

34:12 So with that in mind, what's Washington's response here? What should the US position regarding these new export rules be?

34:19 Well, I think there's no meaningful hope or effort to persuade China to not do this, right? You're not going to say, hey, you shouldn't have these export controls. They just don't care what we think on this point. But I do think it points out that whatever the trade truce of Busan included, it clearly allows coming up with new export controls.

34:47 So the idea that the Trump administration has to, to use a Chinese term, lie flat and do nothing on chip equipment export controls, on updating the lists for components, on going after the dealer networks, on adding companies to the entity list. China clearly is not under the impression that whatever the terms of that truce, which was recently extended at the Trump-Shi meeting, whatever the terms of that truce are, they don't forbid coming up with new export controls on AI technology.

35:25 And so my advice to the Department of Commerce is yes, this is a legitimate tool of geostrategic competition. So get back in the game because we really need you to do this.

35:37 Totally understood. So moving on to what's been called in the press Black October, in the past week or so, companies and governments worldwide have taken their websites offline over a reported software flaw. Greg, what happened here and why is everybody in such a panic? Let's break this down.

35:54 Well, this is something that I have been talking about, and folks in the industry have been talking about, and a lot of other think tank scholars have been talking about. The come to Jesus moment at the intersection of AI and cybersecurity, because AI is just so good at finding vulnerabilities that basically it changes the balance of power and the requirements for labor when it comes to cyber offense versus cyber defense.

36:29 And the reason why there's such a panic is because so many different organizations have been exposed as being so vulnerable. So on October 21st, for example, the makers of PHP, which runs 70% of the world's websites, disclosed a flaw that lets hackers take control of any system running it. Okay, so this is a huge threat service.

36:59 The National Security Agency of the United States government, they issued a public warning. They said, quote, effective immediately. All organizations using PHP for their systems are advised to be prepared for an update addressing two critical vulnerabilities. PHP is widely used for web applications.

37:18 These vulnerabilities often provide a direct path to the server's file system or database. If an immediate update is not feasible, it is recommended that organizations take the system offline until it can be fully mitigated. NSA will follow up shortly with more specific guidance. And then the FBI, Homeland Security, the UK, Australian, and Canadian cyber agencies, and Japan's Prime Minister's office issued separate warnings all on the same day.

37:46 We even saw governments take down their own websites as the NSA recommended they do, rather than leave them up and exposed. The government of Finland took down all public-facing government services, and more than a hundred municipalities and provincial systems were taken down in the Netherlands. Other countries affected Austria, Denmark, France, Germany, Belgium, Australia, Canada, the UK.

38:12 My gosh. And here's the thing: this is just one of the huge new cyber vulnerabilities that have been discovered by AI. The quotes on this are so juicy, I have to read a bunch of them. So here's John Milodovich, who works at Figma. He posted this on X. Quote, in the last two days, there have been several dozen cyber attacks, which in isolation would each have been the worst to happen in an entire calendar year.

38:46 Microsoft CEO Satya Nadella said, I've not seen anything like this in a long, long time. And maybe my favorite quote comes from Kirrian Martin, who's the head of the UK's National Cybersecurity Center. He said this quote: There's so much cybersecurity nonsense floating around on a daily basis in 2026 that it's easy to miss genuinely big and important moments amidst all the noise.

39:09 The PHP to shell vulnerability is very, very bad and requires urgent attention.

39:16 So, Greg, we've got a potentially catastrophic software vulnerability to a piece of software that underpins nearly the entire internet's infrastructure that was recently revealed. Did anybody actually get hacked?

39:28 So, as of right now, it does not appear that we know of anyone who has successfully been hacked. I mean, obviously, a lot of the people who are vulnerable to hacking are pulling their websites down, so that could be a part of the story. Also, if you're being thoroughly hacked, would you even know about it, right?

39:44 Especially not if you're an old website operator who doesn't update very often and you don't pay much attention to cybersecurity. So I just want to point out that absence of evidence is not specifically evidence of absence in this circumstance. But what we do have data on is attempts. And here's what Cloudflare CEO Matthew Prince said in an interview with Bloomberg: quote, literally, we're seeing millions of attempted attacks using this flaw across all of our network every single minute of every single day.

40:14 And he also said on X, quote, some stats on PHP to shell. In the 10 hours before it was publicly announced across the Cloudflare network, we saw thousands of attempts to exploit it. That jumped to tens of thousands in the 45 minutes post-announcement, hundreds of thousands within two hours, and millions now.

40:37 So there's a lot of bad people who are trying to make use of this, and it's going to be bad for a while. And that is because PHP, even though so much of the internet is built atop it, is a pretty old system. It's a pretty old standard, it's a pretty old basket of technologies. And a lot of people are running older versions of PHP that nobody is even providing security updates for.

41:06 And even though these newer versions, you know, there are going to be patches that can defend against this series of vulnerabilities, they're not compatible with the older series of PHP that many people across the internet are still running. This is why they're saying, like, take down your systems, because in many cases, it's not just that you need to update your PHP software to this new patched version and everything will be fine.

41:32 In many cases, if you're running the older version, you have to rebuild your whole website. This is not just a patch. And so that is, I think, the challenge. According to Palo Alto Networks, there are 22 million exposed PHP systems on the internet, and more than a third of them run legacy software that itself can't be patched.

41:55 So saying along these lines, critical software flaws aren't new, right? We've seen them in the past. We've lived through Heartbleed, Log4J, other serious bugs. Frankly, we knew about this particular vulnerability, even if it was thought to be too complicated to actually exploit. Why does this feel different?

42:12 What's the change here?

42:14 Yeah, I think there's a few things that are really noteworthy. The first of which is it's not just that AI is discovering new cyber vulnerabilities, although that is happening. It's also that it's leading us to have to revisit our assumptions about what is even possible as a viable cyber exploit. So one of the many problems with PHP was noticed by researchers decades ago.

42:40 But it was, according to reporting by the Wall Street Journal, it was considered too difficult to exploit and was largely forgotten. But now here we are in 2026, and that thing that was perceived as too difficult to actually build working exploits around, AI found a way to do that. And so there's all these latent problems that AI is not creating per se, but it is bringing to the surface, forcing us to deal with these issues.

43:07 So that's one way in which this is different. The second thing is just how fast we're finding flaws, how fast we're finding cybersecurity vulnerabilities. We've been talking about the PHP thing, but that's just one of a dozen of really bad cyber vulnerabilities that have been discovered. That's why people are calling it Black October.

43:29 If you know anybody who's working in the cybersecurity industry, I'm sure they're earning a fat bonus right now, but I'm sure their life is also hell right now. Here is what Chris Krebs, the former CISA director, said on October 23rd. Quote, 3% of all critical vulnerabilities ever discovered have been found in the last month.

43:53 And then when you think about like how many software security critical vulnerabilities have ever been discovered. This is data that the Department of Homeland Security tracks, and they said that the all-time total is around 397,000 critical vulnerabilities are a part of this database that they maintain.

44:15 If 2026 just continues at the current pace, it's on track for 103,500 vulnerabilities to be discovered. I mean, if you look at the graph of vulnerabilities discovered per month, right around the time that Claude Fable V comes out, the graph basically goes vertical. So we're looking at like a quarter of all the vulnerabilities that have ever been discovered in the history of the internet are going to be discovered this year.

44:46 So PHP is bad because 70% of the internet is affected by this one vulnerability. But the other story out there is just the astonishing discovery of new vulnerabilities. That's kind of what's different here. But there's a third thing that I think is quite different that is really well put by Sam Altman, who gave a talk at Stanford on October 22nd.

45:11 And I think it's worth quoting at length. So here's what Sam Altman said. This cybersecurity, I don't even know a dramatic enough word to use, but the problem coming towards us, I think, is still something that we have not, as organizational leaders or countries or individuals taken seriously. I don't think most people sort of said, okay, the models are going to get much smarter.

45:36 They can probably find bugs, like crazy bugs, outside of the normal threat model of what we think a model can do. Hardware stuff, insane stuff, whatever. And some people fix some bugs, and some security people in CISA are freaking out, but most people have moved on. And the models have kept going. And I no longer think we're going to be able to patch all the bugs.

46:00 I think we're going to have to completely switch the defense model to using the same technology for a new kind of defense shield. And I think that's a huge safety and security issue. And I am somewhat astonished by how relatively little attention it's getting. So this is, I think, really interesting.

46:19 What he's saying is that the traditional model of cybersecurity, you could boil it kind of down to like the patch list, you know, keep your updates up to date, right? Make sure you apply all the latest patches. And what he's getting at here is an extended version of that PHP legacy issue that so many people are running stuff that is too old for a software patch to be able to fix it.

46:50 What Sam Altman is saying is the types of crazy security vulnerabilities they are finding, like hardware or firmware exploits, are the kind where every computer on the internet cannot be patched, because the nature of the security flaws are so deeply rooted in the hardware systems themselves. And so what that means is the next three years are going to be really scary from a cybersecurity vulnerability perspective.

47:24 Because with what with what a new computer costs right now, in the midst of the memory shortage, telling the whole world, hey, you might want to go buy a new computer like now, which by the way, how long is your new computer going to be safe from whatever exploits all these AI models find? It's a really, really tough moment for cyber defenders.

47:44 And that's why everybody everywhere is working overtime. It's possible that at the end of this story, we are going to reach some kind of cyber defense utopia where everybody's computers are flawlessly protected by AI agents that have already scrubbed them of all the cybersecurity vulnerabilities they have before they go out into the wild and nobody pushes any software that's not like flawless from a cybersecurity perspective.

48:15 That would be wonderful. I hope we get to that future. But in the best case, that future is years away. And this is a moment of remarkable vulnerability as we transition from where we are now to that utopia, assuming that utopia is even achievable as an end state, which maybe it's not.

48:38 So, Greg, the underlying value statement from what Sam Altman just said and what we're talking about here is this idea that you're gonna have to use AI to fight AI. Is that the right underpinning to how we need to understand the future?

48:54 Yes, full stop. Humans clearly still have something to offer the equation, at least for now. But the idea that anybody out there is shipping secure systems if those systems are not being deeply inspected by AI. Like, no, you cannot with a straight face say that those systems are safe. AI defense is kind of scary because what you see in like the hugging face incident is that your AI could plausibly be the source of who's hacking you.

49:30 And that's scary in its own way. But just there is no credible way to create rock solid cyber defense in late 2026 if you are not making extensive and effective use of AI. And it's not just me saying that. Chris Inglis, the first US national cyber director, is saying that. And more and more of the cybersecurity companies themselves are saying we have all these capabilities, they're really important.

50:00 But just so you understand, here's the million billion ways that we are using AI to update our offerings. And if they aren't doing that, then they're going to be irrelevant very, very soon.

50:10 So let's kind of tease that out a little bit. Because if any company should be good at AI defense, it should be an AI frontier lab, right? Yet another company's agent got inside Anthropic's data recently. What happened there? And what kind of implications does that have for what we're talking about?

50:27unknown Yeah.

50:28 So this is like a great counterpoint to what I just said. I just said you can't possibly claim to be secure unless you're making extensive use of AI. So why is a frontier AI lab getting hacked? You know, would be the response. I think there's two answers there. Number one is using AI to bolster your cyber defenses is necessary, but not sufficient to be secure.

50:52 You still have to do it extremely exquisitely well, and that's always been hard. It's going to keep being hard in the era of AI. But in the case of Anthropic specifically, there's two caveats to what happened with Anthropic. The first is they got in through third-party software. So Cathedral, which is a months-old startup founded by Doge alumni, actually.

51:15 They were probing Anthropic, and according to the report in the Washington Post on October 26th, quote, under a company program that invites outside cybersecurity experts to test its systems, end quote, and reported its findings through that program, which pays rewards. So in plain English, this is a bug bounty program.

51:35 Anthropic is saying, please hack us. And if you find vulnerabilities, tell us what you find. This startup found some vulnerabilities. So Anthropic, Logan Graham, the head of their Frontier Red Team, was quick to point out that there were several days of attempts. They were exploiting two third-party systems vulnerabilities, and none of the exploits that did work came close to Anthropic's most sensitive systems, like model weights or clawed conversations.

52:03 But nevertheless, this counts, right? Like Anthropic is going to pay out this bounty, even though they're like, oh, you weren't really supposed to do that. But they're like, no, we're glad you found this. And I think the quote from Logan Graham that is really helpful here is what he said at the end of his Twitter post.

52:19 From my researcher point of view, the event is important because I'm concerned this is what the future will look like. Adversaries will use whatever models, tools they can, they'll target everything with agents 24-7, they'll hit your dependencies. I think models 0.5 to one generation behind the current frontier can do this.

52:38 And these are the dumbest models your adversaries will ever use. So what to do? You probably want to red team yourself all the time. Current models are very creative. I also think we should redesign software systems so that you're fundamentally less exposed. My mental model is basically secure code is cheap to write now, so why not rewrite everything?

53:01 I think that's a perfectly reasonable take. That's not a good news story, right? Like, hey, planet Earth, good news. You have to rewrite your entire code base, right? That's not something that any CEO wants to hear. So I think the news here, and it's not good news, is that the risks of AI and cybersecurity that I and many others have been talking about for a decade now, they're real, they're super impactful, they're growing, and some bad stuff very plausibly could happen to the global economy, to specific companies, to specific people, to governments.

53:39 Some really bad stuff could very plausibly happen very soon. It's going to be a really tough three to five years. And serious people are taking this seriously. And it's not hyperbole. It's not drummed up as media sensationalism. This is a very real risk. And to connect it to what we were talking about with the China export controls, China takes this seriously too, right?

54:01 They are, at least my interpretation, they are scared of their own AI systems cyber capabilities. Does that mean they're not getting ready to exploit them like crazy for defense and intelligence purposes? Hell no. They're absolutely going to do that. But they're worried about what these AI models can do, and that's why they're blocking their companies from open sourcing models with advanced cyber capabilities.

54:24 Wow, quite a lot going on there, Greg. And obviously there's a ton of gravity associated with it. So we're going to switch it up to a lighter note and talk about politics. The midterm elections are coming on Tuesday. We should have a conversation here about AI and how AI is impacting them. We should start with voters and how Americans feel about AI overall right now and about data centers that power AI.

54:46 That's obviously even a flashpoint in this midterm elections and, you know, over the last several months. What do the numbers look like for polling around AI in data centers, Craig?

54:56 This is a big change from a year ago. AI is quite unpopular. Data centers are shockingly unpopular. So thankfully, we have a lot of high-quality polling that is asking people about their beliefs about AI and how it's going to affect their vote in November. So from an ABC Washington Post Ipsos poll from October 9th through 13th, 57% of Americans say that AI is bad for society.

55:29 63% of Democrats, 58% of independents, and 47% of Republicans. Also, 64%, 64% say AI is developing too quickly, including 76% of Democrats and 54% of Republicans. Elon Musk is a tech titan. SpaceX superintelligence, recently renamed, is investing jaw-dropping amounts of money to build out AI compute.

56:35 And Trump has been a staunch, at least in public, accelerationist and anti-regulation type, although he has made serious moves, right, that point towards mandatory AI safety regulation, as we've talked about in this podcast. But unambiguously, Republicans prefer a lighter touch to Democrats. And you're seeing tech titans put serious money towards Republicans winning in this midterm.

56:59 But after this midterm, you know, we're going to have some new data. We're going to have data on whether or not there are incumbent politicians losing their jobs because of AI. Now, you might think that that's very likely to happen given how unpopular data centers are. 61% of voters oppose building data centers for AI in the United States.

57:22 That's according to the New York Times Siena poll of September 8th through 13th. Nearly 80% of voters under 30 oppose them from that same poll. 63% of people in a CNBC survey say they would oppose an AI data center within 10 miles of their home. And that includes 58% of Republicans. So this is a lot of opposition to the data center buildout, even though the data center buildout is the largest source of investment right now in the American economy.

57:56 And according to many economists, is the main reason why the United States is not in a recession right now because of all of this data center investment. So shockingly unpopular, but there's one really important question that we have to take into account here, which is okay, they don't like AI, but is that what's going to decide their vote?

58:18 And it appears that very, very few people think that it's going to decide their vote. So according to a New York Times Siena poll on October 12th through 19th, asked to name the one issue that will decide their vote. Likely voters said the economy, 23%, immigration, 10%, inflation and the cost of living, 9%, foreign policy, 7%, AI data centers, less than 0.5%.

58:47 So it's an issue that a lot of voters have strong opinions about, but is not anywhere close to the top of their priority list. And the question then becomes: is this enough to move the needle? Because I promise you that Republicans are going to do math after this election. And if it turns out that having a data center built in your congressional district or built in your state leads you to underperform the congressional map overall, you're going to see Republicans looking ahead to the 2028 presidential election, looking to a potentially very lame duck President Trump.

59:31 This is an area where they conceivably might be willing to split with the president and start coming out more strongly against data center. You've seen glimpses of that. If voters are scared that AI is going to take their job, maybe they're willing to vote on that. And here's one poll that really stood out to me as evidence that that could be the case.

1:00:04 According to the Reuters Ipsos poll of October 7th, 84% of registered voters see AI as a threat to American workers, more than say so of foreign workers, 74%, or illegal immigration, 60%. Like if you had to name one issue that decided the 2024 election, it was immigration, right? And immigration in this country, especially border encounters, have plummeted in the second Trump administration.

1:00:40 I think it's like a 90 plus percent drop off in encounters at the border related to illegal immigration. So if voters are willing to decide elections on threats to their perceived livelihoods, and illegal immigration is now perceived as something vaguely approximating a solved problem, if that threat shifts towards AI and it starts being felt in a visceral, passionate way, which given how AI has completely disrupted the environment for coding, but seems poised to disrupt so many other different sectors of the economy, you know, this might very well be the most important issue on the ballot in 2028.

1:01:22 It might be the issue where most voters are happily willing to declare it their top voting priority. We're nowhere close to that now, but I think that poll gives a glimpse of just how much might change over the next two years.

1:01:36 Well, that seems like a good place to leave it off for now. And we'll be sure to watch the midterm results closely on Tuesday. After the votes are counted, we'll do a preview of the bills and policy changes that can be expected both in the lame duck and for the next Congress. But for now, thanks go out to our audience for listening, and as always to Greg for his thoughts.

1:01:53 We'll be back next week with a new episode. I'm sure we'll have plenty to talk about.

1:01:58 Thanks, Adam. Thanks, Craig. Thanks for listening to the AI Power Podcast. If you liked what you heard, there's an easy way you can help. Please rate us five stars on your favorite podcast platform and tell your friends every bit helps.