← The AI Policy Podcast
The AI Policy Podcast
Trump's National AI Framework and Super Micro's Chip Smuggling Indictment
01:02:41
In this episode, we unpack President Trump's new national framework for AI legislation, including reactions from experts and policymakers.
Loading summary
Transcript47 lines
- [00:00]
A
Foreign.
- [00:11]
B
Hi, welcome Back to the AI Policy Podcast. I'm Sadie McCullough and today I'm speaking with Greg Allen about the White House's recently released national policy framework for AI and a revealing indictment describing the smuggling of Nvidia's best chips to China. Greg, thanks so much for joining again. But first, I think we should get started by asking you how your legs are doing.
- [00:31]
A
Oh, man. Yeah, it turns out that running a half marathon is something that you should do after you train for a half marathon, not before. How are your legs, Sadie?
- [00:45]
B
They're better now, but yesterday was a bit rough, even though I did train.
- [00:48]
A
Yeah.
- [00:49]
B
So I'm glad to see you're in one piece.
- [00:51]
A
Yeah, sort of. We ran the rock and Roll half marathon here in D.C. on Saturday. It was a delightful race and a beautiful day, but man, I was like falling AP at the end. We ran the army 10 miler in October for which I did train, and that went considerably better. And then in a fit of euphoria, we signed up for this race. And yeah, I was considerably under trained, but Sadie and I and our former colleagues Brielle Hill and Jordan Adamson had a great time out on the course. Just not as great of a time as we would have had if we had done more proper training. But it was quite cold this weekend.
- [01:27]
B
Yourself?
- [01:27]
A
I trained well, you're just the most disciplined of us all, I suppose.
- [01:32]
B
Well, it was a great time. But I feel like you also have something exciting you want to share about the Claude project you were working on.
- [01:41]
A
Yeah. So I think this is something that may not interest all of our listeners, but if you want to get to the hard hitting policy, go ahead and skip a few minutes. But I just wanted to share some personal experiences with progress in AI agents and using it for the types of work that I do here at csis and then also use cases that I'm kind of imagining. So I will say the two use cases for which AI agents and LLMs and chatbot assistants writ large have been good at for a while and have been useful to me. And my work for a while is sort of like a good, good path towards edge cases where Google searches aren't especially useful. So sometimes for example, I will have heard something from conducting an interview, but I'm looking for more details on it or I'm looking for authoritative search sourcing on it and that's hard to find. So for example, one time from interviewing folks, I had heard that the path that Nvidia took to degrade its chips from the H100 to the H800 back when that's the strategy the company was taking was to blow fuses on the chip that were originally part of the story of defect recovery, which is to say how do you make the chips still usable even if there are manufacturing defects? But I didn't know a lot of the terms of art, such as the term defect recovery in the semiconductor manufacturing ecosystem. That would have been very useful if I was in, if I was using Google searches, but because I didn't know the specific keywords I as I tried to put in my sort of vague sense of these terms into Google and this was, this was years ago at this stage, but I'm just recounting sort of where we were a couple years ago. It didn't. I couldn't find authoritative sourcing for this phenomenon that I knew was true in a vague sense, but I wanted to write about in a specific sense and I wanted to do it with authoritative sourcing. And that's when ChatGPT's deep research came out and that was able to scrub through Nvidia PDFs and technical documentation that are hundreds of pages long that were available openly on the Internet but if you didn't know the exact keywords to search for weren't easily findable via Google the way I was looking for them. But because LLMs enable a more semantic, semantic based search based on the ideas as opposed to the specific keywords and, and Google does some of this, but LLM took it to an entirely additional level. I was able to find, you know, the relevant passages in the relevant technical documentation and understand them so that I could cite them in my work and provide, you know, authoritative, precise descriptions of phenomenon that I knew were true in vague sense based on people I talked to who, who were experts, but took me to a level I've been able to write about. So that's like a use case that LLMs have been good at for me and my work for years now. Now as AI has built on progress just improving the quality of the models themselves, but also giving the consumer facing services and the enterprise facing services more agentic type capabilities where they can actually go out and do stuff, I've been curious to see how that works. And I've also been interested in, you know, to what extent giving these models a file structure to work makes them better. So I will say, you know, giving the model examples of my own writing, even giving them podcast transcripts and saying, hey, convert this podcast into an outline that I could use to write an article just as sort of a test experiment. Still, the results were quite low. And I used what I understand to be my, what I understand to be the best practices in prompting for such type tasks. Although maybe there's something that I don't know that I don't know. And so there's another level of performance with the existing state of the art out there to be unlocked. But basically I'm sort of at the stage where AI can do some things that meaningfully help me in my job, but it cannot fundamentally do my job, which is to say that the article draft that it spat out and the outline draft that it spit out was unusable garbage. And keep in mind, that's when I had already given it all the ideas that I wanted to express, at least in this experimental trial run. Because I was saying, convert this podcast transcript into something resembling an article, and here's what my articles tend to look like. So at least for right now, AI can't do my job. But then just over the weekend, as I was laying on the couch whining about how much my legs hurt, I decided to run another round of experiments. And this was with the coding based assistants. So in my own mind, you know, I'm not a great programmer, but I remember this game that was my sort of final project for reaching an early stage programming milestone for me, which was to create a memory game where it would, you know, display boxes, put numbers in them, and then it would hide the numbers and you would have to click the boxes in order of lowest to highest based on what you remember. And I had worked with, I've done this myself, I have created a version of this game independent of AI agents many years ago. And then a couple years ago I tried to persuade, I tried to get a combination of ChatGPT and anthropic claw to create a version of this game that was usable. After a lot of troubleshooting, I was able to get something that was ultimately usable, but it did involve my knowledge as a programmer, which, as I said, is not especially strong. And it took an awful lot of troubleshooting. And the LLMs, you know, this copy pasting going back and forth was not an especially pleasant experience. And so that was around December of 2024, I guess.
- [07:47]
B
Yeah, this first version, yes.
- [07:50]
A
So now Fast forward, it's March 2026 and I'm using Claude Cowork, which is the more non expert friendly adaptation of Claude code and still has, you know, a really strong suite of coding capabilities. And it created an extremely high quality, at least based on my, you know, limited version of this game. Very quickly I was able to create additional games that had like fun animations that were pretty easy to persuade the model to add. And then I finally decided to create something that I thought my kids, my, my twin sons are five now, and I'm trying to get them to practice math with me. And they're also taking karate classes now. And so I created this game that I call Karate Math, which has little animated sprites that hit each other every time you do a math problem. And the amount of damage it does is based on the math problem you do. So if the, if the math problem is, you know, 25 plus 10 and you get it right, you do 35 damage to the opponent.
- [09:00]
B
And it was fun.
- [09:01]
A
Yeah, it was pretty mind blowing to me to watch Claude cowork chew through this problem. You know, I was still very much involved as the sort of overarching designer and as the detector of bugs, including some bugs that were like, you know, browser compatibility issues. It works in Chrome, it doesn't work in Safari, but this is one of those cases where it created a capability that I thought was genuinely interesting, that my kids actually found genuinely fun. And it was vastly beyond what I could have created myself. As I said, a pretty lousy elementary programmer. And so it just really was kind of an aha moment for me about just how it was like a way for me to viscerally understand just how far the coding capabilities have come in the past two years, even though some of the writing and analytical work, while it has also come far, is at least not right now at the stage where it's going to be putting me out of a job. But I also kind of understand why the programmers are out there saying, hey, what's happening to us right now is going to happen to all of you in a matter of years. It's only a matter of time.
- [10:16]
B
So do your kids actually like doing math now?
- [10:19]
A
At least they did this morning. But that also could have been this. They didn't want to go to school yet and they were trying to convince me to let them play the game more so that they didn't have to go to school right away.
- [10:30]
B
Well, whatever works. But that's definitely. It was a very cool, cool game that you created that I saw this morning.
- [10:37]
A
Yeah. So that's my advice out there to everyone is if you've been waiting for these agentic capabilities to get to the point where they are idiot friendly, they're there. You almost certainly, if you're listening to this podcast, can start coding up stuff and this, you know, in my situation was recreational, but just given the amount of capability on display, it's very easy for me to imagine coding capabilities that would be useful to me in my professional life.
- [11:08]
B
Great. Well, I keep, I'm looking forward to seeing what you keep creating. But let's get into the policy discussion for today.
- [11:14]
A
Yeah.
- [11:15]
B
So on March 20, the White House published its National Policy Framework for Artificial Intelligence Intelligence. So we'll get into the recommendations this framework makes and its future on the Hill. But first, could you give us some context on where this framework comes from and what the goals are?
- [11:31]
A
Yeah. So the, the basic requirement for this framework comes from the Trump administration's executive order. Back in December 2025, it said it was going that, that the Special advisor to the President for AI and crypto, which is David Sachs, and then also the Assistant to the President for Science Technology, which is Michael Kratz of the White House Office of Science Technology Policy, shall jointly prepare a legislative recommendation. And as you look through, as you look through this national policy framework for artificial intelligence that just came out, a lot of it is Congress should, in fact, almost every bullet point starts with Congress should. And it's still a very high level document. It's only when you print it out, three pages, four if you include the title page. And it is basically just a list of objectives. Objectives. But that is sort of the starting negotiating position of this Trump administration. Now that the national congressional moratorium on state level AI regulation that the White House was in favor of, now that proposal failed, now they're coming back and saying, okay, we do, we still want a moratorium on state level AI regulation or at least certain categories of it. But given that we recognize that Congress is not willing to do that until there's some kind national level framework for governing AI, this is the, the White House's starting position for what that should look like. And it's a really interesting data point. And you can kind of interpret it in two different ways. One way that you can look at this document is what the White House thinks is most important from an AI policy perspective. So the first item, for example, is protecting children and empowering parents. Maybe that's what the White House thinks is the number one most important AI policy issue to work on right now. But another way that you could read this document is from a politics perspective, what the White House thinks is most important from a politics perspective when it comes to AI regulation. And there you can basically say that they acknowledge and recognize that Congress and the American population is quite frustrated with the state of AI as it concerns child endangerment and you know, the parents being able to control what these technologies are doing vis a vis their children. So that is kind of interesting and how. I would say you can think about it. The second thing that I thought was worth noting is that the first section, which is protecting children and empowering parents, ends with a bullet that says Congress should ensure that it does not preempt states from enforcing their own generally applicable laws protecting children, such as prohibitions on child sexual abuse material, even where such material is generated by AI. So you might ask yourself, like, wait, what is this contradiction? But actually, I don't think there is a contradiction in this documentary, or at least not at this stage. The original moratorium as was proposed in Congress and supported by the White House was really looking, looking unfavorably on regulations coming out of states that regulated AI development, but looking pretty comfortably at AI regulations that concerned harmful use of AI and that that paradigm is retained through this policy framework. So even as they continue to support a moratorium that does not include on certain categories of laws protecting children related to the harmful use associated with AI, and even as the White House looks to go comparatively light touch when it comes to regulating the developers themselves.
- [15:33]
B
Sure. So can you break down some more specific policy areas that the framework focuses on and what are some specific recommendations that you found interesting?
- [15:43]
A
Yeah, so I guess we should, we should probably just go line by line. Right. So there are seven sections. The first is protecting children and empowering parents. The second is safeguarding and strengthening American communities. The third is respecting intellectual property rights and supporting creators. The fourth is preventing censorship and protecting free speech. The fifth is enabling innovation and ensuring American AI dominance. The fifth is educating Americans and developing an AI ready workforce. And the seventh is establishing a federal policy framework preempting cumbersome state AI laws. Right. So they didn't want to put the moratorium front and center even though they are still committed to it. And that's why I think, you know, interpreting this as maybe both a policy document and a political document is the, the most astute one. So we already talked a little bit about the, the, the child one. I think what's really interesting is that there's one recommendation in here that says Congress should empower parents and guardians with robust tools to manage their children's privacy settings, screen time, content exposure and account controls. Now when you say Congress should empower parents with robust tools. Well, Congress is not going to build anything technologically. Right. So effectively this is going to come in the form of regulations saying that the developers have to provide these tools in order to provide these capabilities. So that would be something more. But this is also the kind of thing that many of the social media companies are already emphasizing they're creating. Right? As you walk down the streets of Washington, D.C. it's not uncommon to see an ad on a signpost like a billboard where Meta is talking about how they're creating new teen accounts for Instagram and new types of parental controls. And I think, you know, those types of controls are seen as the political alternative to what Australia has done, which is ban kids under a certain age from even having social media accounts. And so, so I think this is putting in these types of protections that is still a pro industry position, but it's just a less extreme pro industry position as industry and government have acknowledged the pretty unfavorable politics on this story and how it's changed. So that's what called out to me on protecting children and empowering parents. The second one on safeguarding and strengthening American communities. This is kind of an interesting section because it mixes energy and AI safety topics in one area. So here's one line that I thought was quite interesting. Congress should ensure that the appropriate agencies within the national security enterprise possess sufficient technical capacity to understand Frontier AI model capabilities and any associated national security considerations and establish plans to mitigate potential concerns, including through consultation with Frontier AI model developers. So it doesn't say, you know, we love Casey, the organization at NIST that has an important role and was previously known as the AI Safety Institute. It doesn't call out, at least right here, the Department of Energy, although they also do important work in this area. But I think one thing that's, that's noteworthy here is that the national security concerns associated with Frontier AI model capabilities, which could be be lowering the barriers to weapons of mass destruction development. It could be the AIs themselves presenting some kind of a risk through cybersecurity or something else. It, the White House still thinks, think that this is important enough to put it on the first page of their policy proposal. And I was glad to see that because I think it does deserve their time and attention. But the second, which is in the same section, which again I thought was a little bit odd, but it's talking about the data infrastructure build out and how it should strengthen communities. And so here's the first bullet. In accordance with the ratepayer protection pledge, Congress should ensure that residential ratepayers do not experience increased electricity costs as a result of new AI data center construction and operation. And look, energy prices are already a real political salient point with the war in Iran, for example. But AI and the energy is definitely perceived as a big source of pain. And whether or not it's the case that every place across the country where rates have gone up, it's because of AI, it is certainly the case that many politicians have seen a benefit of blaming AI. I mean, you're seeing Senator Bernie Sanders, for example, call for a moratorium on new data center construction. I don't think that's an especially widely held view in Washington or maybe even around the country. But the point is, you know, concern about increasing electricity prices is real. And what the White House has said here is that Congress should ensure that that doesn't happen. The only thing that it really says to do that would potentially reduce the chance of that happening is streamlining permitting so that more of the new electrical generation can take place in terms of behind the meter power generation. So which is to say, building something that's off the grid, or mostly off the grid, but still, I mean, as you think about all the demand for all the natural gas turbines and every other new way of generating electricity, at least. It's tough for me to see how at least the ideas that are raised in this proposal would in and of themselves be enough to make electricity prices go down. But you could just call that the White House punting to Congress and saying, you figure it out. You figure out how to.
- [21:37]
B
Yeah, just some helpful starting points.
- [21:39]
A
Yeah, exactly, exactly.
- [21:40]
B
More or less.
- [21:41]
A
Okay, so the third thing is on intellectual property rights and supporting creators. And here is something that I thought was. Was quite interesting in their first bullet quote. Although the administration believes that training of AI models on copyrighted material does not violate copyright laws, it acknowledges arguments to the contrary exist, and therefore supports allowing the courts to resolve this issue issue. Similarly, Congress should not take any actions that would impact the judiciary's resolution of whether training on copyrighted material constitutes fair use. Now, this is pretty interesting. And at least as. As of right now, the courts have not resolved this, right? The New York Times very famously has a lawsuit open with OpenAI, which you. You encounter every time you read a New York Times article that mentions open AI.
- [22:34]
B
We've talked about it a lot on the podcast.
- [22:35]
A
We've talked about it a lot on this podcast. So what's going to come out of the yet uncertain. I think one plausible outcome of the courts is that AI for training data is not a copyright violation, unless perhaps you are violating explicit instructions not to train on that data. So If a website, for example, includes a robots. Txt file that's supposed to be read by web crawlers and scrapers that says do not scrape this website, and you scrape it anyway, you know, maybe there, there might be a different kind of violation. But assuming that it is made openly public, then training on it does not in and of itself copyright the copyright violation. But perhaps at the stage of generating AI content, if you generate a copy of something or if you generate something that is associated with intellectual property rights, then you might be violating copyrights. So that would be, for example, on the Internet, there's a lot of, of Disney imagery. And so you would say including that imagery in the training data is not in and of itself a copyright violation. But if you create an image that includes Mufasa and Simba from the Lion King, especially if you create one that's like an actual still image from the movie, then that could be a copyright violation. That could be an intellectual property violation. I think that's a very plausible outcome for where the courts might land. And it seems like the White House is sort of endorsing the courts acting because they think that that's where something like this might be headed. But one other thing that's kind of interesting here is that there's no mention of the patents part, right? So the, the intellectual property. In the United States, there's three big pillars of intellectual property rights, and those are copyright, trademark, and patent. And in the, in the patent area, the courts have ruled that AIs cannot be the creators of anything. So even if, you know, an AI agent goes off and invents some kind of new way to slice bread, the authorship of that patent has to go to the person who entered the prompt or the person who owns the AI or something like that. You know, AI's in and of themselves cannot have rights. And I think that that is also sort of understood here in the copyright thing when you're, when you're embracing the approach taken by the courts. That's one of the important court rulings that we have as a precedent in the intellectual property sort of a thing. Okay, there's one other thing which gets to rights that I thought was interesting in this section, and I'll have to read the long quote quote. Congress should consider establishing a federal framework protecting individuals from the unauthorized distribution or commercial use of a generated digital replicas of their voice, likeness, or other identifiable attributes. While providing clear exceptions for parody, satire, news reporting, and other expressive works protected by the First Amendment, Congress should prevent persons from abusing such a framework. To stifle free speech online. So that's kind of interesting. So what would that say? That would say that when OpenAI generated a voice that sounded an awful lot like Scarlett Johansson from the movie her, where she played an AI assistant, if they indeed, you know, trained that with a direct intention of resembling Scarlett Johansson, that would very likely be violating her rights as a creator and her likeness and voice. Likeness without her permission, without compensating her. However, if, you know, somebody was creating a satire of the movie her, or parody or something like that, then they could do it because that has certain kind of explicit First Amendment protections. So that is pretty reasonable. That's the same way that we apply to satire for political cartoons, for example. Political cartoons can create parodies of Disney properties when they want to do that in the. In a way that I could not if I wanted to just create a children's entertainment movie, not a satire. But that is going to be a thorny thing to think about implementing at AI scale. So, again, the White House throwing it over to Congress to figure out how to do that.
- [27:07]
B
A lot of fun stuff for them to deal with.
- [27:09]
A
Yeah, yeah. So here's another thing. The fourth pillar, preventing censorship and protecting free speech. Well, it relates to what we were just talking about, but it also includes this bullet. Congress should prevent the United States government from coercing technology providers, including AI providers, to ban, compel, or alter content based on partisan or ideological agendas. And now, what's interesting here is that there's a lot of ways in which this government is really trying to stamp out what it sees as WOKE ideology finding its way into AI platforms. The example that was very, very prominent during the Trump administration, sorry, during the Trump campaign was the Gemini example, where it had a. A specific thing that it would add to image generation prompts to add racial diversity among the composition of the image. And this was true even when you were depicting historical examples, where from the real historical context, there almost certainly would not have been racial diversity in that image. So they're basically saying we don't want Congress to act and sort of say, AIs have to talk, talk a certain way. And I just sort of wonder, you know, how is that going to encounter this administration's approach to how they've been, you know, handling their owner, their control of the fcc, to how they've been handling federal procurement rules, where they've sort of said that we don't want WOKE ideology in federally procured AI stuff. A lot that sort of has to be ironed out with some precision. Here to see what the White House really means and what Congress really might be willing to bring about.
- [28:56]
B
About.
- [28:57]
A
Okay, so that takes us to the, the fifth item, enabling innovation and ensuring American AI dominance. There's stuff that you would expect like Congress should provide resources to make federal data sets available to industry and academia and AI ready formats for use in training AI models and systems. That was something that the first Trump administration was very gung ho about. That's something that's very consistent with some of their major initiatives like the Genesis Project at the Department of Energy. There's also this thing about establishing regulatory sandboxes for AI applications that help unleash American ingenuity and further American leadership in the AI development and deployment. I'll just say that I've been hearing a lot about regulatory sandboxes in the context of the EU AI Act. That was something that Dragos Tudorak, one of the co rapporteurs for that legislation at the EU Parliament was very gung ho about. And I don't, I don't, don't, don't believe I've heard anybody bragging about how well it's going. Since I think that the best regulatory sandbox, to be perfectly honest, is the states. Like the fact that the states have a different approach to regulating. For example, autonomous driving, extremely permissive in Arizona, much less permissive. And I don't know, let's say Maine or Minnesota or somebody with a lot of snowy weather where they don't want to be so risk taking in the embracing of this kind of new technology. So what regulatory sandboxes really means here, I don't know, it just seems like one of those things that often shows up in regulatory debates and I'm still waiting to see a great version of it. Okay, here's one that comes back to our discussion earlier about the growing capabilities of AI which is educating Americans and developing an AI ready workforce. Now, a lot of the public anxiety about AI does relate to, you know, is it going to take my job? Are my kids going to have a viable path to employment? And what's interesting here is that there's basically three things in this section. The first is that Congress should come up with non regulatory methods to ensure that existing education programs and workforce training and support programs, including apprenticeships, affirmatively incorporate AI training. So that's basically saying we need our workers to be ready to use AI in the course of their jobs and that like the jobs of the future are going to involve using AI as one of the tools of doing that job. So to the extent that we are already supporting workforce retraining that needs to be a part of it. The second thing is about collecting more data about understanding the issues of what the AI workforce, what the impacts of AI are going to be. And so that's expanding federal efforts to study trends in task level workforce realignment alignment driven by AI in order to inform policies supporting the American workforce. I think there's a lot of support for this. I think the, the Bureau of Labor and Statistics would love to do this. I think the Federal Reserve would love to do this. This is a pretty obvious point. And then the final one is about land grant institutions, which is, you know, how the federal government weighs in on a lot of collegiate education at the state level. And there again, it's more about technical assistance, developing AI youth programs, so an education and retracement training. None of those are bad ideas as far as I can see. But I think it's also worth pointing out that, you know, there's no big bang ideas coming out of this administration on what to do about the AI workforce problem. That could be because they principally don't foresee a workforce problem, that they think that the labor disruptions associated with AI are going to be analogous to those of the Internet or computers when they first came out, which is to say some disruption, but fundamentally a good news story and will create more jobs than it gets rid of. And then we come to the seventh thing, which is about establishing a federal policy framework preempting cumbersome state AI laws. So they are still committed to the preemption thing. And I kind of already said my biggest point here, which is, I'll read the version of it that they have here. Quote, states should not be permitted to regulate AI development because it is an inherently interstate phenomenon with key foreign policy and national security implications. So they're about development as the issue. They're not against states that want to regulate the harmful use. But then it says that states should not be permitted to penalize AI developers for a third party's unlawful conduct involving their models. But now think about that in the child endangerment context. When you think about, for example, what was happening with Grok on X not that long ago where Grox AI was at the request of users, undressing many women and girls on the platform, including underage girls, that would imply that, you know, it's not grok's fault that users are doing bad things with its technology. And you have to contrast that with some of the proposals, like for example, those of Senator Marsha Blackburn, that are coming out in Congress right now. Now, which would give the AI developers a duty of care on some of these issues related to child endangerment.
- [34:26]
B
So great that you mentioned Marsha Blackburn, because I was going to ask you next that it's very interesting that two days before the White House released this framework, Senator Blackburn released a full discussion draft implementing the same December 11th executive order. So how does her draft compare to the framework that was released East?
- [34:44]
A
Well, for one thing, she called it the Trump America AI Act. So she is trying to claim the mantle of being more Trumpy than the Trump White House. But I think it's pretty clear that there's a lot in this proposal that the White House does not support. And the first one is what I just Talked about. Title 1 places a duty of care on AI developers in the design, development and operation of AI platforms to prevent and mitigate foreseeable harm to users. So we get back to that issue of do you want the regulations to target the developers and what exactly they put into that system versus do you want to target the users? And if people use these tools for bad, then we should punish them for their bad use, but we should not necessarily punish the people who created those tools. Well, the legal term of art, a duty of care, would imply a great deal that companies would need to do to ensure that this does not happen, even in the context of malicious users. And they could be found legally liable in civil lawsuits, for example, when they are found to have not exercised an appropriate duty of care. So when you see that that's there, you should foresee here come the lawsuits for that kind of a thing. Then Title seven of the legislation enables the US Attorney General, state attorneys generals, and private actors to file suit to hold hold AI system developers liable for harms caused by the AI system for defective design, failure to warn express warranty, and unreasonably dangerous or defective product claims. So that is a lot of legal hooks to go after these companies. And of course, what you would expect is that they're going to respond to that extreme, I say extreme to that significantly increased legal exposure by presumably putting in more safeguards and being more conservative about what they deploy and when they deploy it. And I think notably Senator Blackburn has basically said that she, quote, welcomes the White House to, quote, this important discussion and looks forward to working with my colleagues to codify the president's agenda. But she also says that her Trump AI act is, quote, the solution America needs. So this very strange, you know, kind of dance that politicians do to talk about who actually is carrying the appropriate mantle of Trump in this stage. So I don't know where this is going to go politically. Michael Kratzios of the White House has said that he met with the Senate Majority Leader and the House Majority Leader. Those are obviously both Republicans. He said that he thinks that there's a path forward to bipartisan legislation here. But at least from the public communications that he's put out there, I haven't seen him talking about talking to Democrats. So what the White House, if anything is going to do to try and make this a bipartisan outcome unclear at this stage age. And I think one thing that's, that's worth touching on here is that the, the politics of AI in the United States right now are not especially kind to either party. At least there was one survey put out by NBC News that was asked, you know, who do you think which party do you think would do a better job of handling artificial intelligence intelligence? And neither the Republican Party nor the Democratic party pulled above 20%, which is to say by far the most common answer was neither would do a good job of handling artificial intelligence. Exactly. So that's, that's a remarkable thing here. And I think the, the reality is that both parties do foresee that AI is going to be on the ballot in a certain sense in the upcoming election. You know, we mentioned how two state legislatures in California and New York who are very prominent in their state's AI regulatory debates are now running for Congress. And there's also large super PACs that are putting a lot of money in this and running ads on these issues. And but the question is like, it's not obvious to me that either party is happy, happy that AI is on the ballot, which is to say do they, do they foresee a significant advantage? Because elections are not just about, you know, your positions on the issues. It's also about which positions are prominent in voters minds. And usually parties try to skew the, the election to be more about the issues that they perceive that the, the public trusts them more on. And with AI pretty low trust in both parties at this stage.
- [39:43]
B
So you've outlined that there are many action items for Congress potential for a lot of new legislation. So what are policymakers and other experts who have been, you know, reading this framework and thinking a lot about it had to say about the future and potential of this framework?
- [40:01]
A
Well, the joint statement from House Republican leaders, including Speaker Mike Johnson, Majority Leader Steve Scala, Lease Commerce Chairman Brett Guthrie, Judiciary Chairman Jim Jordan and Science Chairman Brian Babin. So that's Most of the committees that cover a lot of the AI policy touch points said that the Trump administration took a critical step in releasing a framework that gives Congress a roadmap to pursue legislation and basically that will, you know, make sure all kinds of good things happen. So that's a, that's a pretty strong statement of support that they're going to try and make something like this happen. Then we have Dean Ball who of course used serve in this Trump administration. He said, quote, the White House's proposal for a nationwide AI law is a thoughtful document that will serve as an excellent foundation for the legislative work ahead. I would be happy to see these principles, if translated well into statute, become law. And we also have Sam Hammond, who is also a think tanker at FAI and on the conservative side of the spectrum said, quote, strong framework, much more logical and right sized than the Blackburn monstrosity. I feel like we get a lot of our funniest tweets from Sam Hamlin on this podcast. So that's some of the noteworthy folks who were in favor of what the White House put out. There was also some noteworthy criticism, including from Democratic Representative Gottheimer who released a statement saying, quote, today the White House released its AI framework, which broadly blocks state AI laws and lacks key consumer protections around AI models and agencies safety. While this framework takes steps in the right direction, including child safety and lower energy costs, it is still a half measure that falls short of what's necessary for smart AI regulation. AI is too important and too vital to our global competitiveness, economy, jobs and families to do anything short of a full measure. And then one other that I thought was a noteworthy action out there comes from Brad Carson who is the president of Americans for Responsible Innovation and he wrote on X quote, I think it's like this. If you think the current state of play and social media guardrails are a okay then you'll be fine with the framework. If like most you believe we made catastrophic mistakes re social media, then you should fervently oppose this vacuous framework. And I think it's worth pointing out here that Brad Carson does have ties to one of the super PACs that's out there putting funding into elections related to AI as an issue for voters. And so, so that's indicative of probably where that entire constituency is going to go when it comes to election time messaging for gosh, only eight months till November.
- [42:48]
B
Yeah, well thanks so much for the helpful breakdown, Greg. I'd like to switch gears now and talk about super micro smuggling indictment. So as I Mentioned at the top of the podcast. On March 19, the Department of Justice indicted the co founder of Super Micro, a Fortune 500 tech company that builds servers for AI data centers, and several other individuals for diverting billions of dollars of Nvidia chips to China. So before we get into the details of the indictment, can you provide our readers with some additional context to what happened here?
- [43:22]
A
Yeah. So when we talk about Nvidia selling AI chips, most customers, the vast majority of customers, do not buy the chips, which is to not say they do not buy little pieces of silicon. Nvidia integrates those using advanced packaging into what's called the AI module. And that module is what's called the H200 or the B200 or whatever. But customers don't even really buy those. What they buy are servers. And Supermicro is one of those key partners that buys from Nvidia the chips and the associated networking hardware and integrates them into these big beefy servers and then actually goes and installs those on behalf of customers. So supermicro is a super important part of Nvidia's supply chain. I think they are responsible for about 9% of the chip maker's revenue. And it's not like Supermicro is like a big cloud hyperscaler, but big cloud hyperscaler might buy from somebody like Supermicro. So in terms of the pipeline of how Nvidia chips become, say, OpenAI servers, a company such as Supermicro, maybe not always Supermicro itself is a really important part of the story. Dell also provides this type of functionality. There are other companies out there who do what Supermicro does, but they're unambiguously an important part of the AI supply chain. So when you find out that the co founder of Super Micro is now alleged by the United States government to be involved in a $2.5 billion AI chip smuggling ring to China, boy, oh boy, is that bad news. But it is entirely predictable. In fact, it's what I predicted on this podcast and in my reports. And it's not to say that I'm like psychic or anything. A lot of reporting has been saying that stuff like this has been going on, not only saying Supermicro by name, but certainly giving you enough facts to intuit that somebody like Supermicro, whether them or somebody like them, had to be involved in the story. So let's get to what Nvidia had been saying until up to this moment. Let's start with a quote from Jensen Huang in May 2025. Quote, Governments understand that diversion is not allowed out. Diversion meaning basically smuggling. Right. When you, when you say you're going to buy it for one customer and you divert it to another customer. So diversion is, you know, export control speak for smuggling. Governments understand that diversion is not allowed. And there's no evidence of any AI chip diversion. Recognize our data center GPUs are massive. These are massive systems. The Grace Blackwell system is nearly 2 tons. And so you're not going to be shipping, you're not going to be putting that in your pocket or your backpack anytime soon. Soon. And so these systems are fairly easy to keep track of. But the important thing is that the countries and the companies that we sell to recognize that diversion is not allowed. 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. No, they're not. Apparently. Apparently their top executives are personally involved in billion dollar smuggling rings, which in a sense, like, of course they were. Recall that back on February 23, Reuters reported that Deepseek was smuggled, trained on smuggled Blackwell chips, and before that had been training on smuggled H100 chips. We discussed that in depth on our February 2025 episode. But now I want to go to specifically one of the quotes that is in the indictment, which is about the timeline here. So in the indictment it says, quote, beginning in or about 2024, and then it goes on. The defendants conspired to divert billions of dollars worth of the US Manufacturers servers to China. Well, this validates stuff that I was writing and stuff that journalists were writing in 2024. You know, basically there was not significant demand for smuggled chips. Chips, as long as the H800s, which is a degraded form of the H100, were still legal to sell. But once that became illegal, once that loophole was closed in October 2023, that created the demand for smuggled AI chips. And around early 2024, smuggling got going. And by mid 2024, it was already on a massive scale. And here's the thing. Reporters have been doing a fabulous job finding out about the tactics that smugglers have. You know, we have reports of the New York Times reporters going to Open Air electronics markets, markets and being offered to purchase smuggled Nvidia chips right there on the spot. And also being shown text message receipts of transactions for hundreds of millions of dollars worth of smuggled chips. You have people on TikTok and it's a, you know, Chinese equivalence and competitors also talking about how they've got large Shipments of smuggled AI chips. Let me just read from this article from the information in August 2024 that I think just this shows, like how good some of the journalism has been on this story for a long time. Quote Several months ago, an electric appliance company in eastern China put in a $120 million order for 300 servers. Servers powered by eight of Nvidia's cutting edge H100 chips. The order was for chips that US export controls bar from sale in China. To get around those rules, the company didn't go to one of Nvidia's authorized distributors, but to a chip broker in Malaysia. The broker arranged for the Chinese buyer to establish a shell company in Malaysia, concealing any link to the parent company in China. He also helped set up a corresponding corporate website and corporate email address to enhance the fake company's legitimacy. The broker even rented space in a Malaysian data cent to temporarily house the servers when they arrived as a way of enfooling Nvidia staffers who wanted to check if such a larger order of servers was installed properly. In a matter of weeks, the servers were in China, having first passed through Malaysia. According to the broker, who gave his first name as William and who didn't want to be identified by his full name, the episode demonstrates how smuggling Nvidia chips to China has become an organized multinational enterprise this year, partly made possible by the complicated distribution network Nvidia uses to sell its chips. I mean, this was in August 2024, and for the two years since, Nvidia has been saying that smuggling is impossible because the stuff is too heavy, as though shipping containers magically stop working and trucks magically stop working when they're used by bad guys. I mean, these are extensive descriptions of the tactics that they're using, including later in reporting by the same author, switching the serial numbers on servers so that when inspectors come to look at one set of servers, they see the serial numbers and say, oh, okay, yeah, I guess the chips are still in Malaysia when in reality, you know, that's a different shipment that was recoded and the original servers have already been shipped to China. Well, that was an incredibly sophisticated thing and now we know that part of the reason why some of these Malaysian pass through entities were able to succeed is because they had help from a trusted Nvidia distribution distributor, Super Micro. And at least as it concerns the co founder, the co founder, it appears that Trump's trust was horribly misplaced. Now the indictment also talks about pass through Southeast Asian companies, so entirely consistent with that Reporting, you know, talking about going through Malaysia and obscuring the uses for end users, preparing false documents, repackaging boxes, concealing the scheme with dummy servers, pressuring the compliance team to just sign off on everything. That's how you make this sort of smuggling on this scale happen. And I think it's just been painful for folks like me and folks in the US Government even to know that this is going on, to see evidence that makes it obvious that this is going on, and to hear from leadership in Nvidia and at times, David Sachs in the White House, who has tweeted stuff about how the servers weigh 2 tons and therefore can't be smuggled, when it's just obvious, obviously wrong. And so I hope that this will bring some sanity to the debate about what needs to be done. You know, the challenge here is that BIS has gotten a plus up in terms of its budget allocation, but the situation is still not healthy when it comes to staff and expert staff. And right now, you know, the trade negotiations that we're having with China have taken on such sales that it's hard to remind folks that, like, no, we really do need to have this competency. It is great that we finally caught this guy, but it is terrible that we're catching them two years after journalists were already on the trail of all of this kind of activity. Journalists should be learning about it from bis. BIS should not be learning about it from journalists. And until the intelligence support to BIS, until BIS's own investigatory authorities are strong enough that that's the case, you know, we have a big mismatch between our priorities and the power and resources we put behind those priorities.
- [53:16]
B
All right, Greg, thanks for breaking that down for us. Those are some pretty crazy export control violations. So what do you think the government should be doing to prevent this from happening again in the future?
- [53:27]
A
Sure. So I want to point towards Chris McGuire, who was previously at the White House National Security Council, where he worked on this issue specifically. And he had some stuff that he put out on that I thought was pretty astute. So let's just go line by line with what he said. Quote, first, we need to know where these chips are going. All AI chip exports to Southeast Asia, the nexus of Chinese smuggling operations, including this operation, and potentially globally, must require a U.S. export license. Now, saying that you need an export license globally is going to be very controversial because, you know, these, as I said, BIS capacity has been limited. And so sometimes times, even if theoretically, they're just supposed to rubber stamp the licenses. Right. When the chips are going to say the uk If BIS is really backed up or shorthanded, that can be a problem. But I do think that we need. One of the reasons why you might want a licensing regime is just so that all this data is coming into bis just so that they know where all the shipments are going to those kinds of customers. So I don't know that I agree with Chris that we're always going to need a license. But it's very clear that, you know, just the standard data collection is probably not going to be enough. And I hope that there's a way that we can do something that's less burdensome than licenses. But just to give BIS greater visibility into all of these export networks. The second thing Chris says is stopping purchases by U.S. based Chinese companies. So quote, second, Chinese companies inside the United States should not be allowed to purchase AI chips. It is absurd that the only country in which Chinese companies can buy AI chips is the United States itself. A loophole that DOJ has highlighted in past indictments that Chinese smugglers routinely exploit. Gotta agree. Yep, that's pretty obvious. Okay. Tighter compliance for US companies. Third, much tighter compliance measures are needed by US companies. US companies have demonstrated that they cannot be trusted to self police companies must have stricter end use reporting requirements and or face stricter liability. Export control enforcement must become more like financial sanctions enforcement if it is to become effective. And I think that, you know, if you read about financial sanctions it is not trivial the internal compliance measures that companies have to have. And I, I do kind of have to ask the question, right? Like why was it not obvious to Nvidia that this was happening? It is kind of an indictment of their entire compliance process and I think they need to be asking themselves those questions and Congress needs to be asking the questions of like how could this go on and not be detected and not be stopped again, you know, why are we learning about it from the government? Sorry, you know, why didn't the companies catch it themselves? Okay, then we go to Senator Tom Cotton who is of course a Republican. Quote, the DOJ indictment of Super Micro Computers co founder and other employees for smuggling Nvidia chips raises serious concerns about our export control enforcement. And then he goes on to say I've asked at Commerce Gov to begin implementing some provisions from my Chip Security act to prevent more large scale smuggling of advanced AI chips to Communist China. And it's worth noting that the House Foreign Affairs Committee will mark up the Chip Security act this week. So that's not law. But Senator is already asking or Instructing, to the extent that he has the authority to do so, the government to start doing some of that stuff.
- [57:04]
B
Sure. So that was a really helpful breakdown. And I have one last question for you also on Nvidia, but they also received some good news in the last two weeks and they're finally able to sell H200 chips to China. So we've obviously covered H200 hundreds many, many times. So where did we leave off and what's new with that story?
- [57:27]
A
Yeah, so, I mean, as the smuggling networks prove, there is still strong demand amongst the user community for Nvidia's chips in China. Right. We've seen the Chinese government say, oh, maybe we'll block the imports of these chips because we're so eager to have self reliance. But you know, that can be a negotiating tactic because if you're at the same time you're like facilitating and I do think there's reason to believe that the Chinese government allowed this smuggling activity to happen. It would be reasonably easy for them to catch that if they were insistent that they wanted to do so. So, you know, when you say, oh, we're going to allow the smuggling, but we're not going to allow the official shipments. Right. That's kind of about sending a message to the Trump administration. Basically, hey, legally sell us the Blackwells, don't just sell us these H20s which stink, or these H2 hundreds which are good, but not as good as the Blackwells that we wanted, the Vera Rubins that we want. So that's one part of the story that I think this brings into clear focus now, the fact that Nvidia is getting to allow some of these shipments through, I mean, I think it sort of raises the question of, you know, what is the current state of the debate inside the Chinese regime? I have always maintained that there's multiple factors that the cloud providers, that the AI frontier model developers, they want to be able to buy Nvidia, but folks like Huawei and smic, they want more requirements about local use to sort of strengthen their own businesses, even though they're not especially competitive vis a vis Nvidia on the merits alone, at least not right now. And so what this tells you is that the Chinese government is sort of publicly acknowledging that, at least for a while, they're going to continue needing Nvidia and they're going to allow these shipments to go, go forward. Worth noting that Nvidia had previously said that, you know, its Chinese customers have to pay in full, in advance because they don't know what's going to happen in terms of the US interrupting this deal or Chinese government interrupting this deal. And so they're saying, look, you, you, the customer, need to take on the risk here. We don't want to get stuck holding a bunch of inventory in this regard. I think it's really confusing to me how this specifically got allowed. And what I mean by that is that, you know, the BIS review, the BIS rule for these licenses, you know, had to say that it was not going to get in the way of American companies buying these chips. And it's worth pointing out that, like the spot prices for renting H200s out there on the cloud market right now, H200s are still incredibly valuable. A Google executive said not that long ago that like, not just one generation ago, right. Hopper is one generation before the Blackwells, but chips that are like four and five generations old in the TPU stack are still fully subscribed because the demand for AI chips is so, so high right now. So how you can say that these chips going to China is not going to get in the way of chips being delivered to American companies? Companies. I think that's a really weird argument to make. And I don't understand how BIS greenlit the purchase on those grounds because their rules said that they had to, had to greenlight it on those grounds for the sale to go forward. And that's why I thought that Peter Wildefords from the Institute for AI Policy and Strategy, his post on X was so clever on this. So here's what he wrote. Amazon CEO Jassy every provider would tell you, including us, we'd grow faster if we had all the supply we could take. Google CEO Pichai We've been supply constrained even as we've been ramping up our capacity. Nvidia we're redirecting supply to produce for China. Now, the Jassy quote is a true quote. The Pichai quote is a true quote. The Nvidia quote is a inference based on their behavior that Peter is making there. But I think it pretty accurately reflects the state of affairs that we are hyperact right now, which is pretty disappointing. It would be lovely if China could suffer from the chips shortage that this policy was originally designed to make them suffer from. But in the meantime, the American economy will suffer from not a huge chip shortage, but at least a chip bottleneck as we continue to wrestle with export controls. Now going into. Gosh, we're on our fourth year here.
- [62:01]
B
Well, thanks for a great breakdown, Greg. I learned a lot. I'm sure our audience did too. And it was great to be back with you on another episod. So looking forward to more in the future. But thanks again.
- [62:13]
A
Yeah, thank you, Sadie. Take care. Thanks for listening to this episode of the AI Policy Podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review on your favorite podcast platform and subscribe and tell your friends. It really helps when you spread the word. This podcast was produced by Sarah Baker, Sadie McCullough and Matt Mann. See you next time.