← The AI Policy Podcast
The AI Policy Podcast
The Future of Nvidia’s H200 in China and the Pentagon's New AI Strategy
00:59:36
We talk export policy for Nvidia's H200 chips and the Pentagon's new AI Strategy.
Loading summary
Transcript29 lines
- [00:00]
Greg Allen
Foreign.
- [00:10]
Sadie McCullough
Welcome Back to the AI Policy Podcast. I'm Sadie McCullough and I'm back as always with Greg Allen to talk about the future of Nvidia's H200 chip in China and the Pentagon's new AI strategy. Greg, thanks so much for being back.
- [00:23]
Greg Allen
Great to be back with you, Sadie.
- [00:25]
Sadie McCullough
All right, well, let's jump right back in to recent developments regarding H200 chips exports to China. So on January, January 13, BIS issued a new rule specifying conditions for H200 exports. The next day, China blocked H200 chips from entering its borders. So before we discuss China's decision, let's unpack the new rule for our listeners. So, Greg, can you tell us what's in it and what process does it create for exporting these chips?
- [00:52]
Greg Allen
Sure. So to help everyone understand where we are in this story, President Trump had already announced that he was going to allow exports of H200s. But that is the sort of high level decision. There's also the nitty gritty part of the decision which is like, how is this going to happen? How is this going to be implemented? And the way this has come out is a new rule issued by the Department of Commerce, Bureau of Industry and Security. And this gets to the specifics of what is going to be allowed, under what conditions, in what time frame. And we got a lot of additional detail. Essentially, this is trying to take things that Trump has already decided were going to do and figure out by what mechanism they are going to be legal and to also figure out by what mechanism they are going to be enforced legally and put into practice. So to give you one example of something that this did, President Trump had previously negotiated a deal with Nvidia CEO Jensen Huang that AI chip exports to China, 25% of that revenue would be given to the United States government. Now, if this was was an export tariff that is explicitly banned in the United States Constitution, which says that there will be no tariffs or taxes on exports. So how do you have a deal that does something that arguably is illegal under the Constitution? Well, it appears that we got the answer in the form of this BIS rule, which is saying that all of the chips exported to China are going to undergo a security review by the United States federal government. So they're not just going to say Nvidia, you say that these are the chips that you're allowed to sell. They're going to check. Right. That these are the chips that they're allowed to sell. And that is a service being provided by the United States federal government. The check and you can charge for services that you provide, even when they're two private sector entities. And so if I had to guess how this is all going to be implemented, it's going to be Nvidia is going to be charged 25% of the cost of the product as a fee for service in doing that service of the check. So this is what I mean when it says, like, how are these things going to be legal? And it really is like that. I mean, you are sometimes in the government and you are told from on high, this is what we're going to do. Find the legal authorities that would enable me to be able to do this. So that's what's going on here. Now the next big thing it does is put in place a lot of restrictions on under what circumstances the chips can be sold, how many of the chips can be sold. It's pretty clear to me that what they're trying to do is to some kind of extent thread the needle between the national security concerns that so many folks in Washington D.C. and around the world have raised about these chip exports and then also the benefits as the Trump administration sees them, both from the revenue that Nvidia will receive from these sales and also, as we've heard from the aizar David Sachs, the purported benefits of getting China a addicted to the American technology ecosystem. So how exactly is the Trump administration going to thread that needle? Well, they put it in this way. And I also want to send my compliments out there to Janet Egan and James Sanders over at the center for a New American security, who on January 16 published a really nice report that is a soup to nuts evaluation of this rule. So if you want to go into more depth, we're going to go on to this podcast. I certainly endorse your reading that report and I'm going to quote from it liberally here in this podcast. So, to begin, the rule caps total shipments to China at 50% of how many of each chip has already been bought for use in the United States. So what that means is you cannot sell all the H200 chips that you can only sell 50% of how many of those chips have already been sold in the United States for use by customers in the United States, which is is interesting. And so CNAS actually took the the step of saying, what does that suggest about how many chips China would be able to buy in the near time frame? They said, quote, China could purchase up to 850,000 H200 chips under this threshold, combined with AMD's Mi325X chips, China could acquire the equivalent aggregate compute of nearly 900,000 H2 hundreds. And what that means is that's an awful lot of chips, right? I think the largest AI comput facility on earth right now is probably Grox Colossus, which is now the Colossus 2 in process of becoming the Colossus 3. And I think they're pretty close to if they have not already exceeded 1 million chips in that computing supercluster. So what it means is that China as a country is certainly authorized to buy enough chips to build a state of the art world class computing facility now, and that's right away. And as more chips get sold in the United States, they can then sell more to China. So that number is not a static number that's going to continue going up as chip sales go up. One caveat there of course is that there's not just one company in China that would like to buy these chips. Alibaba wants to buy the chips, Baidu wants to buy the chips, Tencent wants to buy the chips, bytedance wants to buy the chips and on and on. And so that those 900,000 H2 hundreds that could theoretically be sold would have to be divvied up amongst those companies unless Nvidia basically picks a winner, which I think they're extremely unlikely to do and says all the chips get to go to one place. So what you really have is that the top companies in China are probably going to each be able to buy something like 100 to 150,000 to 200,000 chips apiece. Maybe some of them will be more excited to buy more, some of them will be more excited to buy less. Now the next thing is that there are certain provisions that Nvidia or AMD or whoever the exporter is has to certify in order to be able to sell these. And one of them is clearly inspired by the Gain AI act, which has since died in Congress. But you can see that some of the ideas from that debate live on because the rule states, according to the CNAS summary quote, the rule requires exporters to certify China bound shipments will not delay U.S. orders or divert foundry capacity from the domestic market. That is very much inspired by Gain AI, which as you all will remember, gave US companies sort of a right of first refusal. Now what's interesting is it doesn't give the US customers the right of first refusal. It just says that Nvidia has to certify that it's not selling these chips to any Chinese customers. When they would have been bought by U.S. customers. And that kind of makes me wonder, like how loud are U.S. customers willing to be in terms of complaining that Chinese companies are getting chips that they would like to buy? I think that is really a political decision because Trump clearly wants to allow these chip sales. So is a company like Microsoft or Google or OpenAI, are they going to say, hey, those are my chips that I would love to buy that you're shipping to China in? Nvidia's certification was deceptive or in error. I don't know if they're willing to get that political. But this at least does give them that opening. Nvidia has said that there is no trade off. So I don't expect them to have any reluctance in certifying this. The question is then, is it going to be contested? Are US Companies going to get loud? Is the US Department of Commerce going to put any question mark on those kinds of certifications? The second thing is this third party review which we already talked about. So here again I'm quoting from the CNAS summary. The new policy requires AI chips to undergo a third party review by a US based testing facility before export. This serves a dual purpose. It creates an independent verification layer that confirms technical capabilities match the license application and it might plausibly be function as a workaround to constitutional restrictions on export tariffs. There's one other thing that I think they might be doing here, which I. Sorry, that not I think they might be doing here that Chinese commentators have speculated they might be doing here. There has been some commentary on Chinese social media that this is actually a ploy to put spyware or malware in the chips by for example, the National Security Agency. So effectively we're going to put these sort of poisoned chips into the Chinese ecosystem and that this is all an elaborate trap to set back China's AI race. That is something that the United States government did do at times during the Cold War. Is selling known defective or worse than defective parts intentionally to the Soviet Union? I don't think there's been any credible reporting that that's actually what's going on here. But if it was what's going on here, would there necessarily be any credible reporting? I don't know, but I think it's interesting that China is concerned about this step in the equation and might factor into their decision about whether or not to buy chips at all. Now I think the next thing that we need to say is a quote from our friend of the podcast, Chris McGuire, who over at the Council on Foreign Relations has written a, I think it's fair to say pretty strident criticism of the new policy. The title says a lot. Quote the new AI chip export policy to China. Strategically incoherent and unenforceable. So here's, here's one quote that I think elucidates his basic criticism. Quote While the restrictions are included in the rule are preferable to unconditional exports, they are still severely problematic, creating a false sense of security while authorizing very large numbers of AI chip exports to China without enforceable guardrails that will prevent their misuse. This will allow China's AI industry to close the gap with the United States and build the largest data centers in the world and will not stop China's military or intelligence services from using these chips. And when, when he says that this will not stop China's military or intelligence services from using these chips and that they're not enforceable, I think he's referring to the end user requirements that exist in these chips. Because you're allowed to sell chips to China but you're not allowed to sell them to the Chinese military or to the Chinese intelligence services. And how does that actually get implemented? Quote the rule requires exporters to certify that chips will not be accessed by prohibited end users, including military linked and blacklisted entities. But here's the thing. The U.S. department of Commerce during the Biden administration specifically said that they were taking a China wide prohibition on the exports of these chips because they had assessed prior end user based restrictions on chip sales and found that they were completely ineffective in stopping them from getting to the PLA or the intelligence services.
- [12:38]
Sadie McCullough
So how do they say that they're going to enforce the end user restrictions?
- [12:43]
Greg Allen
That's the thing. They're going to require that Nvidia certify, right. The exporter certify that they are selling them to a company that is not part of the defense intelligence apparatus. And also that the customer understands and agrees not to resell the chips or use of the chips to a military intelligence customer.
- [13:05]
Sadie McCullough
And is there any incentive for Nvidia to actually do that? Or if they.
- [13:12]
Greg Allen
So I think if they were found to knowingly be selling to customers that had military intelligence ties, they would be subject to criminal penalties and executives involved would also be subject to criminal penalties. But I think the, the criticism here, at least the criticism from the Biden administration about the prior, you know, regime of end use checks and these are criticisms that have been echoed by folks like Marco Rubio who's now in the second Trump Administration is essentially, it's not that Nvidia is selling to, you know, company A, which they specifically know is a front company for the People's Liberation Army. It's just that they know that Chinese companies are perfectly willing to lie to Nvidia, and that Nvidia or other companies do not have a credible mechanism for assessing which ones of their customers are or are not willing to transfer chips to the Chinese military. So that's what I think folks like Chris McGuire mean when they accuse this of being unenforceable.
- [14:15]
Sadie McCullough
Sure, that makes sense. So what have other experts had to say since this rule was published on January 13th?
- [14:23]
Greg Allen
Well, I think there's a few things that are worth saying. First is a quote that comes from Anthropic CEO Dario Amudai, who I think there was folks who know that Dario has been very. Oh, sorry, Dario Amadei. I pronounced his name incorrectly, as I often do on this podcast. My apologies. But Dario has been very loudly opposed to these export controls for a very long time. But something interesting happened in recent months, which is that Anthropic received a financial investment from Nvidia and Microsoft and appeared in a joint video with Nvidia CEO Jensen Huang. And so I think there was folks who were asking themselves, is Amadei going to walk back his criticisms of Nvidia selling chips to China in the wake of that financial investment? And on January 20, at the Davos World Economic Forum conference, we got an answer to that question. And here's the quote from Dario, quote, it would be a big mistake to ship these chips. I think this is crazy. It's a bit like selling nuclear weapons to North Korea. And so my I tip my cap to Dario, who is not moderating his sentiments one iota in the face of what is argu, you know, financial or political pressure. He still thinks this is a mistake. And he's basically saying, look, maybe we're going to do it, but we haven't done it yet, so don't ship the chips. Kudos to him. I think there's another quote that I want to highlight, which comes from a conference in China on January 10, 2026. This was their AGI Next Frontier Summit. And interestingly, they got together. They're four of the leading lights of China's AI ecosystem, representing multiple companies who are leaders in AI. And the full transcript of this panel was made available by the Geopolitics substack, which published a full English translation. So kudos to them and thanks to them for doing so. And I want to highlight A quote from Lin Junyang, who was the tech lead behind Alibaba's Quinn model. So this is one of the best open source models in the entire world. World. Arguably the best open source model in the world. This is a smart person. And when Lyn was asked, how likely do you think a Chinese company will lead globally in the next three to five years? His response was looking at the US China gap. And he said the following. This is a long quote, but I think it's worth reading. QUOTE US compute is one to two orders of magnitude ahead. But more critically, OpenAI and others are putting massive compute into next gen research here. We're barely keeping up with daily delivery, which already exhausts most of our compute. This creates the contrast between rich people innovation and poor people innovation. Rich labs can waste GPUs. Poor labs, out of necessity, are forced to tightly optimize algorithms and infrastructure, things rich labs don't always have the incentive to do. And then he goes on and he says, if I had to assign a number, 20%, and I consider that optimistic because the historical head start is real. So he's saying that once again, you know, this compute gap between the United States and China is a key pain point for a company like Alibaba. They're barely able to keep up with their inference demands. And that he sees only a 20% chance of China surpassing the United States in terms of having the best overall models over the next three to five years. And he thinks that's optimistic. I think that tells you a lot. And I want to underscore one part of the argument that he's making, which is how having a compute advantage reinforces an innovation advantage. And that is simply because when you're trying to make better AI algorithms, well, you need to run experiments, you need to have a hypothesis, oh, maybe if I did this, it would make my algorithm better, and then I could, you know, wring more intelligence out of each unit of data input or computing resource input. But you have to have a hypothesis, and then you have to run experiments to test that hypothesis. And those experiments are themselves computationally intensive. And so what he's pointing out is that companies like Alibaba, Chinese companies, are limited in the amount of computing resources they have because of these export controls. And that actually dampens their ability to be innovative and to develop the next generation of awesome algorithms because they can run fewer experiments because they're so bogged down just trying to serve their inference requirements. Now, he also points out that they have this incentive, you know, to focus on algorithmic Innovation because they can't afford to waste GPUs, as he puts it. So it's not all bad, but it is almost all bad. And I think that's just one more quote that we have from this just laundry list of executives from Tencent, from Deepseek, from Alibaba, from other leading Chinese companies that, boy, these export controls hurt and they wish they could buy all the Nvidia chips that they want and need. So there you have it. There's two really interesting AI leaders around the world, in different ways, making the case for export controls on AI chips. So that's kind of where we are right now.
- [20:11]
Sadie McCullough
Great. So now that we've broken down what this rule actually is, let's talk about China's response to this rule. So we heard that, that we heard about Beijing's decision to block the H2 hundreds from entering China, at least for now. So can you break down this decision more for us?
- [20:28]
Greg Allen
Yes. So this is happening in the very recent past. So we've got two high quality reports on this topic. The first is from Reuters on January 14, which said, quote, chinese customs authorities told customs agents this week that Nvidia's H200 artificial intelligence chips are not permitted to enter China. Government officials also summoned domestic technology companies to meetings on Tuesday where they were explicitly instructed not to purchase the chips unless necessary. Which I think is an interesting little addendum there. Then on January 16, there was a report in the Financial Times, and the article was titled, Nvidia suppliers halt H200 output after China Blocks Chip Shipments. So there's a lot of interesting stuff in there. Earlier reporting had suggested that Nvidia was requiring Chinese customers to pay in full in advance before they would ship the chips, which is not a norm necessarily in the industry where it, you know, is either some, some amount of money up front, cash on delivery, that sort of transaction is, is more typical. And so requiring the Chinese customers to pay fully upfront, I think reflects Nvidia's uncertainty as to whether or not they're going to get hurt again. Because remember, Nvidia had said that they took a $5 billion write H20 work in progress inventory that they were not allowed to ship to China. So where does that leave us overall? I think it goes back to stuff that I've said earlier on this podcast, which is there are multiple constituencies in China and they have different preferences and they have different incentives. You have the Alibaba, Tencent, Baidu, Deepseek, you know, the AI leaders of China, the hyperscale cloud Companies of China, it's pretty clear, unambiguously, they want to be able to buy Nvidia chips right now. They would love to buy a huge, huge number of these chips. But then you have the companies like Huawei, smic, the people who are trying to create the alternative ecosystem to Nvidia in China. And they mostly care about what's good for them. They mostly don't care about what's good for China. And what's good for them is banning Nvidia, right, because that gives them a captive market and near term revenue. And it also encourages the important and powerful customers of China to invest in them, both from a buying their product thing, but also more importantly perhaps is building applications that leverage their software ecosystem, et cetera, et cetera. So that is what Huawei wants. And the Chinese government is in the middle to adjudicate between these two constituencies. And so we don't know where China is ultimately going to come down. It is clearly the case that at least for right now, they're saying, no, you can't buy these H200 chips, quote, unquote, unless necessary. And what exactly that means, I don't know. One interpretation of this behavior of blocking the chips at customs is the idea that this is continuing to negotiate, right? So when China blocked the imports of the H20s, that could be seen as a bargaining maneuver where they basically said, we're not going to accept these lower quality chips. And that successfully persuaded the Trump administration to, quote, unquote, up the ante and offer H2 hundreds a much more attractive chip from a financial return on investment per capability basis. So one interpretation is that China's just doing that again, again, right? That they are refusing to allow these H2 hundreds into the country in the hopes that America will once again up the ante and not sell just H2 hundreds, but also the Blackwells and ultimately the Vera Rubin. And I want to say that Nvidia has claimed that the Vera Rubin next generation chip is going to, on a performance per dollar basis for AI applications, going to be something like 10 times better than the Blackwells, which were already considerably better than the H2 hundreds. So, so China might be saying, hey, you know, you need to make us a real offer here. And what all we're going to accept is the latest and greatest stuff. It could also be a negotiating strategy around these restrictions for the H2 hundreds, which, you know, as the CNAS analysis points out, only allows around 900,000 H200 equivalents in the near term, at least. And so they could be Saying, you know, we don't want it if you're only going to sell us 900,000. We would want it if you're going to sell us like 5 million. And that's actually enough to, to move the needle so that we can build five AI companies at the frontier or whatever. So that's one interpretation of events. Another interpretation of events would be that Huawei is just winning the argument domestically. Right. And I think that's plausible. It's certainly consistent, right, with the customs blocking, the maneuvers. And Xi Jinping was pro self reliance when America was willing to sell everything to China in terms of it. I mean, that was reflected in made in China 2025 back in 2015. Right. That was a path to self sufficiency. And that was, you know, China's first salvo. And it was not in response to some American provocation. That was just what they wanted as their policy. And then the final thing I think could be that, uh, it's not just that, you know, China agrees with Huawei and SMIC in the argument they're making, but that the Chinese government is like being fooled by Huawei and smic. And so you could actually interpret, and again, I'm speculating here, but you could interpret remarks like those of the Alibaba executive as trying to basically say, hey, Xi Jinping, we know Huawei is telling you that they can cover all of China's domestic needs like immediately, but they're, they're lying to you, like they can't do that. Like we actually are suffering. We need you to help us out here and allow us to buy more of these Nvidia chips. So like, that's the question. If, if it is indeed the case that China's government is agreeing with the Huawei faction, are they agreeing because they know the painful trade offs and they accept that pain, or are they agreeing with Huawei because they're being misled by Huawei and they don't even recognize that this trade off is exists.
- [27:23]
Sadie McCullough
So Greg, what do other experts in the field suggest as other reasons that China might be blocking the imports of H200 chips?
- [27:31]
Greg Allen
Well, I think the interpretations that I've given are the main ones that are out there. But I do want to give credit to another friend of the podcast, Leonard Heim, who we've had on here before, and he posted on X about China, quote, they don't have an accurate picture of their own AI chip production capabilities. They've invested billions. Of course they think the fabs are working. I've that SMIC and Huawei have a hard time telling them what's going on. So that's a. That's sort of an elocution of the ignorance theory of China's decision making. There's one other post on X from Sam Hammond which is pretty funny, so I have to read it because it's pretty good. So he said, quote, we're being saved from the mistakes of Boomer US politicians with unrealistically long AGI timelines by the mistakes of Boomer Chinese policymakers with unrealistic realistically long AGI timelines. I thought that was pretty funny. One thing, just because this is like the only podcast audience where I need to make this distinction. But when he says Boomer, he means as in baby Boomer, as in old people, not as in Boomer versus Doomer, AI optimist versus AI pessimist. He's basically saying that both folks think that AI is going to be transformative over a longer time horizon than is real. And, and as he said, we're being saved from the mistakes of our side by the mistakes of their side, which I thought was a. Was a funny way of characterizing it.
- [29:01]
Sadie McCullough
That is a good way to wrap up this section of the podcast, too. And I'm sure that we will be talking about export controls many more times in the near future, so stay tuned. But for now, let's transition to talk about the Pentagon's AI strategy. So, on January 9, the Pentagon published its Artificial Intelligence Intell Intelligence Strategy for the Department of War, which Secretary Pete Hegseth had distributed to senior Pentagon leadership. So to start us off, could you give us an overview of what this memo includes?
- [29:32]
Greg Allen
Well, first, I have to just remind folks that when I was in the government, my final job before I left was Director of Strategy and Policy at the Joint Artificial Intelligence Center. And so I had inherited one DoD AI strategy. I had drafted the white paper that we submitted to the National Defense Strategy authoring folks, for what should the updated be AI strategy be? This was all years ago, of course, so you can imagine I was deeply, deeply interested in this document called the New Artificial Intelligence Strategy for the Department of War. It's sort of the direct heir to work streams that I was deeply involved in and at times led when I was in the Department of Defense. So this is a short document. It's only six pages long, and it is, to put it mildly, bullish on the opportunities that AI presents to the Department of War. It really says that AI is absolutely critical to the future of military power and that the Department of War needs to have a, quote, wartime footing and mindset in adopting artificial intelligence as fast as possible. And strategy documents like this, I want to emphasize, sometimes matter a lot and sometimes are completely irrelevant. And the amount of attention that they get in the press doesn't necessarily tell you how much they're going to matter. One other thing that I think is worth pointing out here is that there is a classified annex to this document. It is referenced in the unclassified version. It talks about the classified annex. And so all that we're going to be discussing, you know, on this podcast, is in the unclassified version of the memo. It is possible that what is in the classified annex is, you know, just reiterating most of this stuff with a little more spice and flair and specific relevance to classified activities. It is possible that it is a much more substantial and direct document and that most of what matters is in the classified annex. I am not privy to which of those scenarios is true. I'm just sort of telling you how the relationship between classified and unclassified strategies sometimes work in the DoD. I will say that I once met a senior leader in the Department of Defense who did matter for AI stuff in the Department of Defense who was completely unaware that there was a classified annex to the first Department of Defense AI strategy. And so that shows you that for a strategy to matter, people have to know about it, to read it, and to feel accountable to it. And I think there's a lot of mechanisms in this document and processes that it sets in motion that to me, basically guarantee this is to be a quite significant document. I mean, to put it mildly, there's a lot of threats in this document to people inside the Department of War who do not get with the program on adopting AI as fast as possible. Let me give you just one example of one of those threats. Here we go. So in the section titled AI Native War Fighting, it said, we must put aside. Sorry, this is a quote. We must put aside legacy approaches to combat and ensure we use this disruptive technology to compound the lethality of our military. Exercises and experiments that do not meaningfully incorporate AI and autonomous capabilities will be reviewed by the Director of Cost Assessment and Program Evaluation for Resourcing adjustment. Now, for anybody who doesn't understand do bureaucratic language reviewed by the Director of Cost Assessment and Program Evaluation for resourcing adjustment means getting rid of you. Yeah, like, those are the people who cut your budget. They basically say, like, you stink. Your program stinks. We're going to cut your budget, and it's our job to Write, like, the reasons for why your budget gets cut. So he's saying exercises and experiments that are not using AI enough, that are not using autonomy enough. You're going to be put on the naughty kids list for a budget cut. So I promise you, everybody who runs exercises in the Department of War is now saying, oh, gosh, how am I going to put AI into my exercises? You know, in a meaningful way? I do not want to be on the chopping block for that kind of a reason. There's other stuff that's in there that I think is really noteworthy. So one section titled Data Access gives new authorities to the Chief Digital and Artificial Intelligence Officer. This office is the direct successor to the Joint Artificial Intelligence center where I worked. It was formed out of a merger with the JAKE and a few other organizations. And the section here is really interesting because it says that the CDAO can demand access to data sets from anyone in the Department of War, effectively, as long as they hold the relevant security clearances for the data. And it says that the people who have those. Well, actually, Let me just read the whole quote. Military departments and components will establish, maintain and update federated data catalogs, exposing their system interfaces, data assets and access mechanisms mechanisms across all classification levels as mandated by the department's May 2021 memorandum creating data advantage. They will deliver their current catalogs with all available updates to the CDAO within 30 days of the date of this memorandum. The Undersecretary of War for Intelligence and Security will ensure intelligence data receives parallel treatment with exploitation pathways established within the same timeframe. The CDAO is authorized to direct release of any DOW data cleared users with valid purpose consistent with security guidelines, effective immediately. Denials of CDAO data requests must be justified to the Undersecretary of War for Research and Engineering within seven days, who will remediate or escalate to the Deputy Secretary. So what does that mean? It basically says, hey, every single agency in the Department of War, you have to. To go inside and figure out, like, what data you have, and you have to publish it in this catalog that basically says what we have, what it might be useful for, and how to get access to it. And if the CDAO ever asks you for access to it and you don't give, give it to them within. Yeah, within seven days they get to go to the Undersecretary of Defense, Research and Engineering, which is one of the, you know, probably five, six most powerful people, the Department of War, and basically say, you know, this is why we're not giving our data to the cdao and if the Undersecretary of War for Research and Engineering doesn't like your explanation, then it gets kicked up to the Deputy Secretary of War, who can overrule anybody in the entire department. So it basically says, say yes when the CDAO asks you for your data, and if you say no, so they get to escalate very high in the food chain very fast to force you, and you're gonna, you know, pick a fight that you're probably gonna lose unless you have very, very, very good reasons. Now, when I was in the dod, this was a huge pain point for us, right? No data, no AI. So if you have, you know, an awesome idea for how you can use AI to, for example, analyze satellite reconnaissance, photograph well, if you don't have a huge, awesome database of satellite reconnaissance photographs, you're definitely not creating an awesome AI associated with that capability. So we had to go beg, plead, and try and get access to the data sets that mattered to us. And sometimes people told us yes, sometimes people told us no, sometimes people told us yes, but didn't make good on their yes. And we had to come crawling to the Chief Information Officer, who was, you know, powerful, but not, like, super duper powerful. Certainly not as powerful as, like, the Secretary of the Army. And it was just not, you know, a fun thing being told no over and over and over again. This memo basically says, I expect the CDAO to get the data they need. I expect them to get it fast, I expect them to get it easily, and anybody who's not making it easy for them is again going to be on my naughty list, which, as I prove elsewhere in this memo, people who are on my naughty list, not good. Yeah, suffer, suffer real consequences and punishments. So that's one big part of this. On data access, there's another thing that is, like, really music to my ears just from, like, my pain and suffering. Something that I lived when I was at the Department of Defense, something I've written about since I have come to csis, which is just all the blockers that make life hard, hard when you are trying to do AI in the DoD. Data access is one of them, but another one is around, like the authority to operate process, which is targeted at a real problem, which is, when you put stuff on the DoD network, how do you know that it's cybersecure and not, you know, secretly includes a package of malware from Russia or China or wherever? So the authority to operate process is designed to say, say, like anything that's going to touch the DOD network has to go through, you know, a 10,000 year review of all of its code to ensure that it's safe. And I think that process is truly awful and on the one hand is targeted at a legitimate problem, but on the other hand, the cure is probably worse than the disease and we need a different flavor of cure. And here's what the memo says about the, the blockers that you encounter that I've written about here at csis. So quote, wartime approach to blockers. We must eliminate blockers to data sharing, authorizations to operate, test and evaluation and certification, contracting, hiring and talent management and other policies that inhibit rapid experimentation and fielding. We must approach risk, trade offs, equities and other subjective questions as if we were at war. To this end, I expect our CDAO to act as a wartime CDAO and to work with the Chief Information Officer to fully leverage statutory and delegated authorities to accelerate AI capability delivery, including cross domain data access and rapid ATO reciprocity. On behalf of PACE pushing leaders across the department, the undersecretary of War for research and engineering will establish a much monthly barrier removal board with authority to waive non statutory requirements and escalate blockers for immediate resolution. That is so interesting to me because it basically means like the way that I've often described it is if you are trying to water your garden, right, you have a bunch of flowers that you would like to bloom of exciting little AI initiatives. Initiatives. Well, there's two ways you can get more water to your flowers. You can either crank up the water pressure or if there is a kink in your hose that is blocking, you know, water from getting into it, you can unkink the hose. And in the previous administrations, I think the overwhelming emphasis has been on increasing the water pressure, like more money, more staff. And I think here what they're really identifying in a direct way is that like we need to unkink the hose. We need to identify what are all the things that make doing AI in the government so hard, so complicated, so slow, even beyond the difficulties of just AI technology. Specifically, what makes it hard when you're in the DoD context. We need to identify those blockers and we need to not, you know, just throw out any of the concerns that they raise. Right, like the ATO process is raising very real cybersecurity concerns. Those have to be be addressed somehow. But the somehow by which they are addressed cannot be worse. You know, the cure cannot be worse than the disease. And they're saying that that needs to be evaluated on a wartime footing, which I think is so interesting because what we've seen in the war in Ukraine, a lot of the progress that they've made on digitization, a lot of the progress that they've made on AI transformation, it's not like Ukrainian brains are 10 times larger than those in America. It's just that they're on a war footing and they have a very different approach to risk. And you can see that the, the CDAO and the Department of War as a whole is now being directed to treat it like they're in wartime and evaluate risk with that mindset, which I think, gosh, if that doesn't make stuff go faster, what will?
- [43:18]
Sadie McCullough
Yeah, absolutely. So next, let's get into the details of the seven PSPs.
- [43:23]
Greg Allen
Yeah.
- [43:23]
Sadie McCullough
Which should be the most concrete outcome of this document. So can you first tell us what PSPs are are, and then can you walk us through each of them?
- [43:31]
Greg Allen
Yes. So this is what they call pace setting projects. PSPs and the way they're summarized is, quote, that will demonstrate the accelerated pace of execution, focus and ethos. We need to stay ahead. The PSPS will also serve as tangible outcome oriented vehicles for rapidly completing our build out of the foundational AI enabled enablers, infrastructure data models, policies and talent needed to accelerate AI integration across the entire department. So basically what they're saying is they've got these seven flagship projects. Worth noting that the classified annex, maybe what it does is it elaborates on these seven projects. Maybe what it does is it adds seven other projects. Right. That are not discussed in this unclassified summary. I have no way of knowing. But in the seven projects as they're described here, it's pretty light in terms of what it says. So it names them. You've got swarmforge, Agent Network, Ender's Foundry, Open Arsenal Project Grant, Genai Mil and Enterprise agents. And they're broken into these three buckets. Warfighting, intelligence and enterprise. So I'll talk about what we've got here, but what we've got here is not a ton. So Swarmforge, this is all they say about it. Quote, competitive mechanism to iteratively discover, test and scale novel ways of fighting with and against AI enabled capabilities. Combining America's elite war fighting units with elite technology innovators. What is that?
- [45:06]
Sadie McCullough
Don't give you much there.
- [45:08]
Greg Allen
Yeah, so I think there's a few things that stand out to me. Number one is swarm. When people say swarm in a military context, they usually mean swarming drone capabilities. And that could be aerial, that could be naval, that could be land, and then iteratively discover, test and scale novel ways of fighting with and against AI enabled capabilities. So that means like what about our swarm versus their swarm? And then it also says combining America's elite war fighting units. You know, some people, when they say elite warfighting units, they mean Special Forces forces. Right? The people who are the best shots and can run the fastest and blah blah, blah, blah. Like Navy SEAL Team 6. Right. Means elite. And then other parts of the military would take offense to that characterization and they're saying like what you mean like our best fighter squadrons are not elite because they're not special Forces. And so we don't even know who is doing this Swarm Forge. It could be referring to special forces, could be referring to anything else. All right, but, but it almost certainly means robotics really is what it means. Really. It's like actual systems that do stuff in war fighting. The next one is Agent Network unleashing AI agent development and experimentation for AI enabled battle management and decision support from campaign planning to kill chain execution. So we hear Agent Network. I assume that they're using Agent in the same way that it is used by most of the generative AI companies, which is, we're talking like large language model models type chat bot stuff. But they don't just chat with you, they actually take actions on your behalf to make you more efficient. And so what might that mean in campaign planning to kill chain execution and management, battle management and decision support? I assume it means that like, okay, we're not just going to be doing all of that stuff by clicking boxes and doing stuff, we might be able to tell an agent, go do some something, and it will actually implement the various steps associated with that, getting that reliable to the degree of reliability that is required for most things involved in the use of force. I'll tell you, I certainly don't think AI agents are there right now in terms of the reliability required. But this is the Department of War trying to figure out what the future looks like and get there as fast as possible. So they'll struggle with those reliability concerns and think about how to integrate agents into mission critical contexts. I think that's going to be a tough challenge. Ender's Foundry accelerating AI enabled simulation capabilities and SIM dev and SIM ops feedback loops to ensure we stay ahead of AI enabled adversaries. Well, this is obviously all about simulation associated with warfighting, which I think if you think about how, how AI has mastered various types of games like chess, like Go, because the simulation of that game so closely resembles the reality of that game. Well, that's already coming, you know, to war fighting. Relevant capabilities. It was a few years ago that DARPA ran a grand challenge where an AI autopilot defeated really good fighter pilots from the US Air Force Force in flight simulators. So the question is, what if you could expand that simulation from being a dog fighting type scenario to being like at a campaign level where we're simulating the whole thing? Well, maybe you could come up with new operational strategies and tactics developed through these simulations as executed by the AIs. Now, one challenge I want to point out here, anytime you're introducing these types of simulation capabilities, is the degree of difference between your simulation and reality. Especially if you're using reinforcement learning, which is a really popular machine learning paradigm. When you're talking about these kinds of applications, any differences between your simulation and reality are probably going to be exploited by your AI system. And that might be. And similarly, any differences might prevent the system from discovering novel strategies. So what do I mean mean by that? Let's take the latter case first. If your simulation does not allow soldiers to get shovels and dig trenches, then your AI is never going to invent trench warfare. Right? So like, the, the innovativeness of the AI is limited by the capabilities of the simulation environment. The, the flip side of that is if there are differences in reality, for example, like there are unrealistic degrees of accuracy that are possible in the simulation that maybe are not real world realistic, then the system might optimize for that. So, example would be in the DARPA dog fighting simulation. Part of the reason why the AI fighter pilot was winning so reliably is that it had sort of become a perfect aimer of its guns. And so it could look at you head on and like, bullseye you from very, very, very far away. Now that's based on the physics of the simulation where he was able to run that experiment like a million times. And how close do the, you know, bullet accuracy, physics error, modeling, blah, blah, blah, blah, blah of that simulation resemble reality? If it's even at all a little bit impersonal, perfect, then maybe your aim bot, when you deploy it in the real world is not going to be that accurate. And so if you have tactics that are based upon that being possible because the simulation told you it was possible, you're going to have a disaster when you meet the real world environment. So I'm excited about this, but I'll just say there's a lot of stuff that makes doing this kind of simulation work hard, but it's definitely exciting enough that the department of War should be exploring it, but I just like there's a lot of challenges associated with that kind of work. Okay, number four. Open arsenal. Accelerating the tech int to capability development pipeline. Turning intel into weapons in hours, not years. Gosh, I really struggle to know what the heck they're talking about there. My first guess would be something related to cybersecurity, but I don't know. When you turn intel into weapons, it's hard for me to see, you know, how you would if it's not a cyber kind of a story. But maybe I'm missing something. Project grant number five. Enabling transformation of deterrence from static postures and speculation to dynamic pressure with interpretable results. No idea what they're talking about there. Sounds interesting though. Okay, number six. Genai. Democratizing AI experimentation and transformation across the department by putting America's world's leading AI models directly into the hands of our 3 million civilian and military personnel at all classification levels. I this is pretty straightforward. This is office worker type support from AI co pilots with sort of state of the art generative AI. That's something that the military has already been experimenting with with nipper GPT and other capabilities making these types of things available. They, they clearly state that they want to do that DOD wide. And one other thing that is very interesting is that it's explicitly says the goal in this strategy that they want them to be state of the art models. They want the, the models that are used in the government to be no older than I think it's 30 days from whatever state of the art is in the commercial world. So if OpenAI releases GPT6 tomorrow, then theoretically GPT6 should be available on Department of War networks not later than 30 days. Boy, that feels ambitious. Yeah. The cybersecurity certifications associated with doing that I suspect are going to be a formidable challenge which hopefully will lead them to get a lot better at AI based automatic verification of cybersecurity. Which is something we should be investing in AI anyway. But anyway, that kind of dream of keeping everything state of the art is commendable, but it's really tough to do that in a classified networking environment. Okay, Enterprise Agents. This is the seventh and last pacesetting project. Building the playbook for rapid and secure AI agent development and deployment to transform enterprise workflows. I think it's just the exact same thing. So whereas number six is about putting LLM chatbots on every desk in the Department of War, Enterprise agents is about. As the commercial world moves towards agentic AI, the Department of War should do that the same. And I think we're in a world right now where agentic AI is already a very real capability that is delivering a lot of value to programmers and coders. And the expectation is that that's going to increasingly be the case for all manner of office work type jobs jobs. And the Department of War wants to be on the front of that trend, not a laggard in that trend. So Those are the 7 PSPs. There's no stated budget for these projects, but it does seem like we should expect to see that in the President's budget request in the not too distant future. One other thing that I think is interesting is that all the services, so you know, Army, Navy, Air Force, Marine Corps space and all the components and all the agencies are directed to identify fast follow projects associated with these pace setting projects. So the pace setting projects are going to be handled at the office of the Secretary of Defense level, presumably led in most cases by the cdao. But these fast follow projects basically means like everybody across the Department of War. You need to explain how you're also trying to do do this thing that is being led by the presumably cdao. Wow. You know, that's quite a directive. Right. Like even the janitorial people have to figure out. Right. How they're going to use agentic AI. That's pretty interesting stuff.
- [56:05]
Sadie McCullough
Yeah. And so I'm interested to hear, based on your experience at the Jake and after you have analyzed this new, this new strategy, how does this strategy compare to past AI plans that the Pentagon has released?
- [56:18]
Greg Allen
Well, it's, it's unambiguously more bold, it's unambiguously more forceful. It unambiguously views the situation as more urgent and higher priority and important. I mean, when I was in the Department of Defense and even before I was in the Department of Defense, AI was a trendy topic. It got a lot of media attention. When senior leaders in the DOD would testify before Congress, they were often asked about it. But inside the department where there's a lot of sort of calcified bureaucratic resistance to doing anything that isn't exactly what I hoped to be doing today, we encountered a ton of resistance. And you can definitely tell the frustration that people like I experienced and my successors have experienced. This memo is designed to take a bazooka to the challenges that we faced and to make life hard for anybody who wants to slow those people down in the Department of Defense. So that's really interesting. I do think we're going to have to wait until we see what the funding story is, especially for these pace setting projects, the Department of War can do some stuff right away using what's called reprogramming authority authority, which basically means you can move money from bucket A to bucket B, assuming the congressional committees, you know, bless that kind of activity and they can just, you know, write a letter blessing it. They don't have to pass a law, but there's limits as to how big your reprogramming authority budgets can be. What this document says is vague but meaningful. Quote the means we will employ to pursue this strategy will continue to encompass our substantial program funding and workforce focused on AI across the services and components. We will also use the timely financial resources provided by Congress in the form of one big beautiful bill, along with expanded budget withhold, joint acceleration, reserve flexibility to catalyze our accelerated pace of military AI integration in the immediate term. What does that mean? It means that they're going to tell people they have to do it within their existing budget authorities in most cases, I think, and then at some point they're going to try and figure out how to get more money, presumably from Congress in the next, not too distant future.
- [58:45]
Sadie McCullough
Well, I guess we'll find out and we'll keep talking about it on this podcast, but I think that does it for today, Greg. So thanks as always for breaking down those two big news stories for us.
- [58:55]
Greg Allen
Yeah, thanks. It was a big, big week for me. Two, two topics I'm very passionate about.
- [59:00]
Sadie McCullough
Yeah, well, I really enjoyed it and found it really helpful. Hopefully the listeners did too and can't wait to chat more on a news roundup in the future. But thanks so much.
- [59:09]
Greg Allen
Thank you Sadie. 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 Mand. See you next time.