Historically, there have been a few notable “beer summits”.
The most famous one in recent American history was in 2009, between former President Obama, then Vice President Biden, Harvard professor Henry Louis Gates Jr., and police sergeant James Crowley to smooth over tense racial relations between the police and the African American community. If you look elsewhere, in 2015, then German Chancellor Angela Merkel held a beer summit (or “Biergipfel”) with Bavarian leader, Horst Seehofer, to settle their differences in refugee policy. In 2018, then Australian Prime Minister Scott Morrison had beers with members of the opposition party to deescalate political infighting. As a life-long hockey fan, my personal favorite is a beer summit in 2010 between NHL Commissioner Gary Bettman and representatives from the players association to ease tension during a collective bargaining negotiation, so fans like me get to watch more hockey.
However, no beer summit is more intimate, or has more dollars and market cap at stake, than the one that Jensen Huang of NVIDIA had last week in Seoul with Samsung’s Chairman, Jay Y Lee, and Hyundai Motor’s Executive Chairman, Chung Euisun.
The day after the now-viral beer (and fried chicken) summit, the South Korean president, Lee Jae Myung, and Jensen – the AI industry’s de facto “head of state” – announced yet another "sovereign AI” deal, this time worth 260,000 of NVIDIA’s most advanced Blackwell GPUs. Using the lower end of one Blackwell chip’s price ($30,000) and removing the complexity of a rack-size product’s higher selling price, like the GB200 NV72, this deal is worth at least $7.8 billion. To further show South Korea’s national commitment to AI, President Lee announced yesterday that his 2026 budget plan will more than triple the country’s investment in AI, befitting of his own claim that this is “the first national budget for the AI era.”
What is more interesting is how these Blackwell chips will be divvied up. Here is the breakdown:
- South Korean government: 50,000
- Samsung: 50,000
- SK: 50,000
- Hyundai: 50,000
- Naver Cloud: 60,000
Now the big question is, can South Korea make the most out of these GPUs? Or at least better than other sovereign countries who’ve signed, or about to sign, similar agreements?
I think South Korea stands about as good of a chance as any sovereign in doing so. Other than the US and China, the southern half of the Korean peninsula has arguably one of the most full-stack technology ecosystems, spanning digital services, consumer electronics, cloud computing, and advanced manufacturing. Its high digital penetration rate, well over 95%, is a well-known phenomenon for many years. Its fiber broadband connectivity rate is one of the highest in the world. The foundation is there for "Sovereign AI with South Korea Characteristics” to be more than a slogan or a newsletter article title.
Some of the GPU allocation decisions seem predictable, if not a bit circular. Of course, Samsung, who is trying to get NVIDIA’s HBM (high-bandwidth memory) business, and SK, who is getting NVIDIA’s HBM business and working hard to keep it, would be more than happy to buy their number one customer’s products. To be fair, injecting some more AI into their own processes, whether it's chip manufacturing for Samsung or wireless connectivity for SK, which has a major telecom arm in addition to memory production, are also decent use cases.
What stood out to me were Hyundai Motor and Naver Cloud, getting as many or more GPUs than the two largest chaebols. This bodes well for South Korea’s overall sovereign AI story.
Hyundai Motor
Physical AI – artificial intelligence deployed in the real (not digital) world – is the next stage of AI deployment. For the long-term AI narrative to not end up as a mirage, physical AI has to work! Unsurprisingly, Lee’s budget speech name checked “physical AI” as a focus of his to grow and foster. It is also the hardest application to get working, which is not to say that the digital AI applications (foundation models, chatbots, agents) are at all easy.
The biggest challenge to getting physical AI working is the lack of quality, use-case-specific data. Data that captures physical world scenarios and use cases are much less plentiful and less available, compared to wisdom-of-the-crowd data on the Internet, e.g. Wikipedia, Reddit threads. (That is also why the “world model” R&D direction, where models simulate the physical world, is getting so much attention lately.) This isn’t to say that the data suitable for training physical AI applications don’t exist. They are, however, locked inside the few large companies that make and ship a lot of physical stuff. One of those companies is Hyundai.
The conglomerate, Hyundai Group, makes a lot of stuff: ships, engines, offshore drilling rigs, roads, bridges, buildings, and the list goes on. Hyundai Motor, the division that is receiving the 50,000 Blackwell chips, makes cars, SUVs, trucks, buses, and happens to also own the humanoid robot startup, Boston Dynamics. In short, Hyundai has the data to train physical AI in many real-world scenarios, if it chooses to apply its know-hows and data that way. Given how digitally connected Korea is, it is a decently safe assumption to make that Hyundai has collected and stored all that data from its manufacturing lines, machineries, and products.
Thus, Hyundai has all the raw ingredients to make physical AI work. It would be a shame if it does not have enough advanced GPUs to do so. Jensen’s beer and fried chicken summit with Chung solved that problem. What’s more, because Boston Dynamics works deeply with US institutions like the Army, whatever breakthrough Hyundai Motor makes will eventually flow to its humanoid subsidiary, as well as its customers in America.
A hallmark of successful sovereign AI, in my view, is AI technology that has a distinct national favor and is desirable and exportable elsewhere. Physical AI courtesy of Hyundai Motor has a real shot of nailing this combo.
Naver Cloud
If there is another Korean tech company that is more under the radar, yet has more hidden potential to deploy sovereign AI, that would be Naver Cloud. It is the cloud computing division of Naver, the ubiquitous search engine and SuperApp, akin to Google. Naver Cloud is its third-party cloud business, just like Google Cloud Platform.
Naver Cloud is decidedly smaller than its hyperscale counterparts from both the US and China. Yet, it dominates Korea’s vibrant but cloistered digital economy. Its biggest competitor, Kakao, which makes the equally ubiquitous Kakao Talk messaging and payment app, among many other things (comparable to WeChat) did not make it into the good graces of Jensen’s GPU allotment. And for good reasons. Three years ago, Kakao suffered a data center disaster that took its SuperApp down for 10 hours, its co-CEO at the time resigned in shame, and the country’s president at the time launched an investigation and questioned Kakao’s monopolistic power. (For more on that disaster, I wrote about it three years ago.)
Today, Kakao is still the dominant messaging and payment app of the country. The company is doing fine. It does have a cloud business under the sub-division, KakaoCloud. But it is hard to recover from a data center disaster of that magnitude and put forth a competitive 3rd party cloud offering, leaving Naver Cloud as the only Korean cloud with scale on the field.
Why did Naver Cloud receive 10,000 more GPUs than all the other much larger chaebols? Chances are, it is because of Naver Cloud’s dominant position in Korea and its growing global footprint and potential to sell GPU-based AI services elsewhere. The company is running cloud regions in Japan, Germany, Singapore, and the US west coast. It plans to build more regions in Thailand, Taiwan, Vietnam, and the US east coast. Moreover, it is building a 500 megawatt-sized AI factory in Morocco to deliver sovereign AI capacity to that part of the world.
Right now, the most reliable way to deliver AI services is still via old-fashioned cloud computing with capacity redundancy and (hopefully) solid disaster recovery backup plans. Naver Cloud has shown that it can operate a cloud at scale and is investing to increase that scale. It has a lot of empty data center shells waiting to be filled by GPUs. An extra 10,000 Blackwells seem more than reasonable.
Sovereign AI with Attitude
It was Europe that first came up with the notion of "sovereign AI” in 2019. The idea was a defensive one – applying the inherent power of sovereignty to control AI technology from spreading and growing too fast in Europe, as driven by American and also increasingly Chinese technology.
Few expected the term to be harnessed by the NVIDIA CEO and turn it into a powerful and effective sales pitch, spreading AI faster and to more places than anyone could imagine, including Europe itself. On the heel of his multibillion dollar beer summit in South Korea, Jensen inked another sovereign AI partnership, this time with Germany’s Deutsche Telekom to build an AI industrial cloud that’s worth $1.2 billion and runs up to 10,000 Blackwell GPUs.
The fact that the German cloud is getting only one-sixth the GPU allocation as the Korean cloud is rather telling, but a topic we will reserve for a future post. What is most telling is the chutzpah and aggression that Jensen is pushing sovereign AI – turning a defensive posture into a potent, offensive go-to-market strategy.
The jury is out on whether or not these sovereign AI investments will pay off as the new factories of knowledge and intelligence that every country needs to innovate and stay competitive, as Jensen confidently prophesized. Or will they become second-rate palaces, satisfying the vanity of kings and presidents, marginally cool to look at, but ultimately a waste of national resources.
As for South Korea in particular, the set up is stronger than most, though the jury is also out on whether the Korean flavor of sovereign AI will be as popular, both domestically and internationally, as K-Pop or Korean fried chicken. But the determination that Jensen is showing in pushing sovereign AI everywhere is clear. That determined attitude can perhaps be best summed up by the closing lines of my favorite Korean hip-hop song (Rosario, by Epik High feat. CL & ZICO):
“Out of my way
I am a legend and I'm here to stay
Did it my way. Did it my way
밟기만 하면 다 길이 됐네 (Where I set foot becomes the way)
I paved the way, I paved the way
For everyone that is pavin' the way
말이 많네 (말이 많네) (You talk too much)
Ain't no one givin' a fuck what you say”