National security advisor Divyansh Kaushik on AI diffusion to the Gulf, Huawei's chip gap, and why America needs an open-source flood strategy

night. And she said she was watching the show and because you were playing the horse sound, her dog was running around looking for a horse. So give give the dog some horse. For all the dogs that are listening, you're now on the hunt for all the dogs. Anyway, uh welcome to the stream. How are you doing?

Doing uh great welcome. Uh thank you so much for joining. Would you mind kicking us off with a little bit of introduction on yourself for for everyone who's listening? Great. Yeah. Uh so Tan Koshik uh I am at Beacon Global Strategies. We're national security advisory firm.

So basically what I do is uh you know work with US industry on how to be better national security partners to the United States. Great. Uh and I I I wanted to have you on to to kind of look at a retrospective on what happened in the Middle East. We saw Jensen Wong go over there.

I think Elon and Sam and Alex Wang and all the big tech CEOs went over.

uh what what is your read now that we have a little bit more distance from how those relationships are changing what we should expect what the actual trade deal came how everything netted out basically yeah I think it's it's a very interesting thing that happened over there because you look at what the Biden administration did in January they they tried to approach AI diffusion uh with like three key strategic objectives they're like we want to prevent offshoring to the Gulf Mhm.

M we want to prevent these chips being smuggled into China through which is mostly a Southeast Asia problem. And the third one like we want to prevent remote access where Chinese entities are just using these chips through the cloud.

Like there are hyperscalers trying to build big data centers for Chinese companies and Malaysia and all. Yeah. Uh using this technology as a tool of economic statecraft I think was a secondary objective not primary.

Whereas I feel like what has happened, what we've seen with the Gulf trip is more of an upside down uh where we're using this technology as a tool of economic state. We're trying to get concessions like we're going to give you these trips, but we want reciprocal investment in the United States.

We want onetoone investment. Y and we want certain security conditions. Now, a lot of that is yet to be cleared out, right?

We haven't finalized, it seems, per media reporting on what the security conditions are, whether they apply to specific entities or whether they apply to the UAE government or the Saudi government as a whole.

Do they get to host Huawei chips or Huawei infrastructure in the same data centers where they're hosting US chips, right?

like there are reasonable concerns that some people have raised and those are things obviously I think the security details will will probably be figuring out uh because uh if you remember Microsoft and G42 did a similar deal at a company level last year in April that was accompanied by an intergovernmental assurance agreement which required G42 to do certain things like not have Chinese hardware in certain number of years uh just you restrict access for military or intelligence purposes like that deal was really well drafted.

There were a lot of things that people pointed out that could be done on top. So I think we'll see how this shapes out in the next coming few weeks how we see the security guarantees coming together.

Uh but overall I think this is in line with what the president has said that we'll give you our technology but what will you give us in return? Mhm. So that's the theme that I expect to uh continue. Art of the deal. Art of the deal. How how important is Llama in this emerging geopolitical narrative around AI.

Uh we're hearing that uh these I I've been calling them like kind of jump ball countries. They're countries that aren't full allies and they kind of could do business with China and that's kind of the geopolitical backdrop for all this.

If they don't do a deal with Nvidia, they're probably going to run Huawei Ascend and Deepseek and Manis on top of that. Is an open-source stack important to some of these countries? Yeah. No, I think there are two questions in there. First, whether China can even export. Mhm.

Which there are actually the the data suggests, you know, China has enough dyes that TSMC shipped to China. Mhm. TSMC shipped to Huawei in violation of US export controls like about 2. 9 million dies enough to produce a million 910 C's about 900 910bs.

Uh now you know that's that's a lot of GPUs but they also have a lot of domestic demand. Mhm. So the question becomes like will they be able to meet their domestic demand let alone export? Mhm. I don't think that they have the capacity to produce it domestically at scale to export.

There was this announcement from Malaysia recently which the minister walked back afterwards where which made for a splashy headline that Ascend data center is coming online in Malaysia. But if you look at the details, it was 3,000 ascents by 2020. Yeah, it's nothing. So they can't meet that. That's that's one thing.

Yeah. The second thing to your point, right, some of these countries have worked with China a lot. Yeah, of course. And well, one of the things of bringing countries into our fold is that we have to work with countries that are in China's fold right now. Yeah. That is a precursor of bringing them into our fold. Yeah.

But you have to be realistic what the risks are and how do we mitigate them.

Some of these countries have been doing joint military exercises in Shen Jang of all places or hosting military bases and Saudi Arabia I believe like a month or two months ago their national telecom champion did a strategic uh announcement with Huawei and so you have to be realistic and put put in place those mitigation agreements but separately to your point on open source I think we should be flooding the zone with American open source at the expense of Chinese models.

Uh I I think we're not paying enough attention to what can happen when a state controlled company to some extent. There are a lot of PLA ties that deepseek has have been scrubbed from the internet.

Now if the state wanted to there was a great anthropic paper last year u not sure if people remember but on sleeper Asians that can be converted into models y if the Chinese state really wanted to you know us infrastructure or get into your phones why wouldn't they just put in sleeper agents that would activate in a year two years yeah if it runs into critical infrastructure it changes the way it reasons.

Exactly. And so we should be pairing ICTS action, right? Like we're doing a lot of export controls on American tech not going to China, but where's the import actions on preventing this Chinese tech from coming into the United States or helping other countries deploy American tech? Like look, come on. Let's be real.

Nobody in Kenya is going to pay $200 a month to access advanced AI models. We should be deploying American stack there. We should be incentivizing companies with like loan programs from the development finance corporation, XM bank, whatnot. Yeah, kind of a new AI belt and road strategy.

Something like that makes a ton of sense. Um, yeah. Do you do you think we'll see something to the to the effect of a stuckset within AI in the next 10 years, 15 years? I'd be surprised if there are not zero day attacks that have already been exploited. Yeah, that we don't know yet. Yeah, that makes sense.

um might uh there's been a lot of talk in around Nvidia's earnings around the shift from uh training to inference and we saw with Deepseek uh some really really creative engineering around uh distilling models and switching to I think FP8 and floatingoint uh and just a whole bunch of optimizations to to get a really great result out of less significant hardware and so uh I wonder if there's is there any risk that the uh as pre-training plateaus, we wind up with a scenario where the distillation process continues and inference shifts onto lagging edge semiconductors and all of a sudden the the the gap between TSMC's leading edge 3 nanometer and China's ability to produce in mass at 7 nanometer even beyond 10 nanometer beyond um that that actually changed the dynamic.

Yeah. And I think uh you know right now we have the lead there. Uh it's it's possible that that happens. And the answer I would tell you that many people in the US government would give you is we should not be complacent about it. Mhm. Right.

Uh we should not be providing components to Chinese companies to build semiconductor manufacturing equipment there to produce these semiconductors. Right. Yeah. Wait.

So, so what is your take on on uh the the chip ban and this idea that like you kind of either need to be all in or all out like it's either don't just worry about the chips but also worry about the entire supply chain or is there is there a creative argument that you could potentially make where there is value to the American semiconductor and and I guess western semiconductor stack ASML TSMC uh actually having an a a benefit in China.

It it feels very counterintuitive, but I was just remembering a decade ago, everyone in technology was making the argument that Facebook and Google should be allowed to operate in China, and now we're making the argument that the other tech company should not be allowed to operate in China.

And I I I kind of agreed with both of those, but they do seem at odds. Walk me through that argument. Yeah, I think look, uh there are choke point technologies, right, where we do not want uh the PRC to gain access.

There are multiple arguments for why you should not give them but the effectiveness really depends on whether it's a chokeoint technology or not. ASML's machines when we export controlled uh extreme ultraviolet lithography. Yeah. Uh then well uh China has not been able to produce anything domestically.

they are still stuck on 7 nanometer uh dual node patterning multiattering right uh and I think that set them back by 7 to 10 years at least. Mhm. But so export controls are not about the point that they will never develop that technology. Yeah. It's buying time and space so that we can extend our lead. Sure.

And that is what matters.

So if we say tomorrow decide you know further tools should not go to China or other semiconductors not should not go or as it happened uh EDA software should not be sold we also need like some promote agenda to really extend our lead uh yeah so Huawei Ascend versus Nvidia um is this a quality or quantity battle how like what is the shape of that differentiation I think it's both uh quality versus yields uh you know they're at about 20% or so uh when producing the dies for a sense so cost yeah it's extremely in you know inefficient and so without state subsidies uh the latest Huawei server that came out it is using technology that Nvidia was using 7 years ago wow right it's putting more GPUs on the same server to come at the same amount of memory and uh interconnect bandwidth and whatnot Yeah.

Now, yeah, can they have a state controlled economy, they can do a lot of this. Mhm. It's not sustainable at the end of it. Mhm. Uh so that's one thing.

Uh I I think there are a couple other things with regards to should we be okay for instance if it turns out that some many of these technologies like look to be real China is a an economy that has an or it is a system of government with an explicit civil military fusion strategy. Mhm.

Every research and development project that happens for civilian purposes is and will be used for military purposes. Mhm.

The vice president of research for Shingua said that explicitly like so are we okay with them developing technology that may be used to kill American soldiers or threaten America's allies in the South China Sea or you know take over the Taiwan Strait.

So those are questions that are you know they're very nuanced questions and hard to answer but the normally but in this scenario I would say perhaps not perhaps it's not okay for us to benefit PLA's modernization. Mhm. So Huawei can do what it is doing right now. I don't think it can scale.

Uh I think the US export controls to much extent are working. The question is how do we extend our lead further? So reviewing this year so far, do you think we're in a better position than we were last year or are we trending in the correct direction?

I guess yeah I think overall we would you know depends on what how do you define the strategy here right the strategy is to win I mean yeah the goal to win the strategy is to stop their development of the technology altogether that's not going to happen right technology races never stay one the Soviets learned that y so the Chinese have learned that too but if the the idea is to maintain gain a lead that they come second, we come first.

Yeah, I think we're going we're headed in the right direction so far. Now, China has tried to weaponize like to try to mimic what BIS does with export controls through their own export controls, but yeah, their export control system is just like 5 years old. Their law passed in 2020.

The nuts and bolts rules for dual use items, they did not arrive until December of last year. So they're very much in an amateur state over there uh with regards to this. So the US does have a lead on that as well. Very cool. Well, thank you so much for joining. This was fantastic. Super insightful.

Uh we'll have to have you back and talk talk more soon. Awesome. Take care. Have a great rest of your day. We'll talk to you soon. And next up we have Aaron Gin, the GPU whisperer himself to take us further into the the discussion around uh AI