원문 보기
Imagine a situation where your company has only one client . You have to put up with the unit price hike, wait even if delivery is late, and have no choice but to say yes if they say they will pay next month . Because there is no alternative. This is exactly the situation for non-tech companies around the world right now .
NVIDIA's market share in the AI accelerator market is approaching 90% . OpenAI, Meta, and Google all cannot run AI without NVIDIA GPUs. However, the price of NVIDIA's flagship AI GPU, the H, is between $ 30,000 and $40,000 per unit, which is over 40 million won in Korean currency. Since you have to put hundreds of thousands of these into a single data center, the price is astronomical.
The problem is that even if you place an order, you do n't know when the stock will arrive . It is the paradox of monopoly. If a supplier has a monopoly, you can consider the buyer to have no bargaining power . The same goes for U.S. Big Tech companies that boast a market capitalization of trillions of won.
So, Big Tech companies are moving to create alternatives . However, the company found as an alternative has a dramatic history. In 2014, a woman named Shio joined a semiconductor company in Silicon Valley just before it went bankrupt . She was the first female Sio in the history of Silicon Valley semiconductors .
The company she joined was AMD, which had previously divided the graphics industry . However, amidst Nvidia's dominance, it was gradually becoming a setter, and at the time of the intervention, it was a company on the verge of bankruptcy with a stock price of only $2. It ended up becoming the company that everyone in the industry was talking about as a beer snack after work because they said it was going to close down .
Lisa Su revived that failing company with her exceptional management skills . In 2025, the stock price rose 77% in just one year and even beat Nvidia's return. However, AMD, led by Lisa Su who turned a 2-core product into a $200 one, still has a single-digit market share in the AI accelerator market.
Lisa Su herself said that her goal is a double-digit market share . It means it's not even 10% yet . The reason we cannot determine exact figures is that the pace of AI development is faster than the rate at which data is being generated, so we are providing only estimates. However, OpenAI signed a contract worth an estimated $80 billion to $100 billion and Meta a $100 million deal with this single-digit company .
Combined, it amounts to nearly 300 trillion won. Why would they pour money into a company with a market share of less than 10% ? It is because I was that scared of staking everything on Nvidia alone . Google is self-sufficient by manufacturing its own TPUs, while OpenAI and Meta have opted for supply chain integration.
By placing an order with AMD, OpenAI signed a 6GW AI accelerator supply contract with AMD in October 2025. Industry estimates range from $80 billion to $100 billion, which is approximately 115 trillion to 144 trillion won. However, to win this contract, AMD gave up to 10% of its corporate assets to OpenAI.
They gave warrants for the right to buy 160 million shares at 1 per share . AMD's stock price is currently about $200, and they let me buy it for 1 t . By simple calculation, it amounts to giving away shares worth about $30 billion at market value for almost free while selling $100 billion worth of chips .
It's practically the same as selling it at a 70% discount . In February 2026, a $100 million contract was also signed with Meta with the same structure. Also includes a 10% stake warrant. If the 6GW supply over 5 years is combined, the total is 12GW, worth approximately 300 trillion won. Why did they go to such lengths?
In economics, this is called incentive design . Those of you who work, you know about stock options, right? It's offering stock to key talent if they stay here long . It's the same principle. But this time it's the opposite. AMD, the seller, told OpenAI, the buyer, "If you buy from us, we will give you a stake." From AMD's perspective, if they don't win this contract, they will be trapped in Nvidia's shadow forever.
Even if it meant sacrificing my own interests, securing Big Tech clients was the only way to survive . Now, but Big Tech didn't stop there. The second strategy is more radical. I will make it myself instead of buying it from someone else. Have you heard of SC? Simply put, it is a custom-made semiconductor.
If the GPU is an all-rounder chef who knows how to cook everything, then it is a master craftsman who makes amazing sushi. If you only perform specific tasks like AI development, it is more power-efficient and cheaper than a GPU . Meta announced its own AI chip roadmap just two days ago . They plan to install a total of four types of MTI—300, 400, 450, and 500— in their data centers by the end of 2027 .
Compared to the initial model, the MTI 300, the MT500 has 25 times faster computational performance and 4.5 times faster memory processing . Up to 512GB of HBM memory is also installed. If the gap is this wide in just two years, it is only taking up the speed of NVIDIA's generational replacement . Wait, Meta is giving AMD $100 billion and building a house too .
yes that's right. I'm doing both . It means not putting all the eggs in one basket . That is the essence of supply chain responsiveness . There is also a company called Broadcom . Open AI is a company that designs customized AI chips for a fee, and its AI revenue alone in the first quarter of this year was $8.4 billion.
It surged by 106% compared to the previous year. The Broadcom CEO said that AI revenue alone will exceed $100 billion by 2027. To summarize, it's like this. Big Tech's MBdia independence projects are two-way. One, pour money into AMD to foster Korean suppliers . Two, he makes the chip himself . Because it is dangerous to rely entirely on one place .
But... There is a crucial twist here. Whether you are buying a GPU, making SLC, or whatever, there is a component that is absolutely necessary . Nothing works without this . I told you that Nvidia dominates about 90% of the GPU market. He looks invincible, doesn't he? But there is one thing that even Nvidia can never make directly .
It is exactly the HBL ancient reverse explosion. If HB is the brain of the GPU, then HBM is the desk where that brain works. No matter how brilliant your brain is, you cannot work without a desk . In economics, this type of relationship is called a complement. No matter how good a printer is, it's useless without ink.
That is exactly the relationship between GPUs and HBM . However, there are only three companies in the world capable of manufacturing this HBM . Samsung Electronics, SK Nix, Micron. With reports emerging that Samsung Electronics became the first in the world to mass-produce HBM4 and Micron was pushed out of the initial supply of Nvidia's next-generation HBM4 due to failing to meet specifications, the atmosphere suggests that the Samsung-Hynix duopoly is effectively solidifying.
This is the pricing power of bottleneck resources. If all roads have to pass under a single bridge, the owner of that bridge sets the toll . Korea is currently on that bridge . I will show you the numbers. It contains 450 AMD next-generation AI accelerators and 12 HBM 40s. There are 12 pieces in a set.
However, to meet the 12GW promised to OpenAI Meta, it is estimated that between 450 million and over 10 million of these are needed . Since there are 12 HBMs out of 450, even with a simple calculation, more than 100 million HBMs are needed . The combined HBM production capacity of Samsung Electronics and SK Hynix was set at approximately 300,000 units per month by 2025.
Even with the factory running at full capacity right now, it takes about 3 years to produce all of this volume. So, Sam Maltman flew to Korea in person in October 2025 . I met with Samsung Electronics Chairman Lee Jae-yong and signed a Letter of Intent (LOI), the content of which is a request for up to 900,000 HBM units per month .
They asked for three times the current production capacity. After this news came out, Samsung Electronics' stock price rose to 200,000 and SK Hynix's to 1 million . Here is the key reason why Lisa Su is coming to Korea on March 18 . SK Hynix has its orders tied up with Nvidia. SK Hynix has also secured the supply of HBM 4 for NVIDIA's next-generation accelerator, Vera Rubin .
Then, Samsung Electronics is the only place where AMD can secure a large quantity of HBM4 . After all, Samsung is already supplying HBM 31 as its largest partner, and Samsung was the first in the world to start mass production of HBM 4 . The important thing here is that starting with HBM 4, the game changes.
In the past, the structure was such that memory companies would manufacture the products, and GPU companies would purchase and use them. However, starting with HBM 4, memory companies and GPU companies must design together from the development stage. It is a technology alliance where we don't just get made and live with it, but build together from the very beginning .
Reports have emerged that Samsung has become the first in the industry to pass the HBM Sacol test for AMD Mi 45. It means we are already making it together . This is a huge opportunity for Samsung as well. NVIDIA fell behind SK Hynix in the HBM market, didn't it? But on AMD's side, Samsung became the number one partner.
Here is one more card. We won the Tesla H last July. $16.5 billion. It was a supply contract worth approximately 23 trillion won . The business unit that had been racking up trillions of won in deficits has started to recover. If AMD GPUs are also manufactured at Samsung Foundry, they are selling HBM and producing chips as well.
It is essentially a structure where one company makes money in two markets simultaneously . But Lisa Su is n't meeting only Samsung. I'm also meeting Naver CEO Choi Soo-hyun. Why is Naver meeting Lisa Su? We are short on GPUs. The CEO of Naver Cloud said that they received 60,000 Nvidia GPUs but it wasn't enough .
But if you rely solely on Nvidia, you'll be dragged around regarding both the price and the schedule . So Naver tried Intel Gaudi 2 but failed. Then, I tried using AMD, and it was verified at the Sejong Data Center. There is something even more interesting here. There is a trend called Sovereign AI, which aims to process domestic data using domestic AI.
The basic concept is that data concerning national security cannot be stored on foreign servers . Countries like Saudi Arabia and Thailand are showing interest . Naver is entering this market with Hyperclobyx, but to export it, hardware is required. It is a structure where Naver AI software and AMD hardware are bundled together and sold overseas .
If CEO Choi Soo-hyun's meeting with Jensen Huang last year was an occasion to share the grand philosophy of Sovereign AI, this meeting with Lisa Su is the implementation phase to diversify the hardware that will actually realize that philosophy . Now, let's take a look at the whole picture. In the third week of March, Jensen is at the GTC keynote in San Jose, and the whole world is watching Nvidia.
However, at the same time in Seoul, Lisa Su is securing HBM4 for Chairman Lee Jae-yong and proposing an infrastructure alliance to CEO Choi Soo-hyun . This must be a coincidence. In economics, this is called strategic timing . While Nvidia shines on stage, AMD is quietly completing its alliance offstage.
However, there are two huge barricades standing in front of this rosy scenario . The first is a software fortress called CUDA. It’s not just the chip capability that makes Nvidia scary. The key point is that AI developers around the world have been writing code for CUDA, NVIDIA's proprietary language, for the past 15 years .
To use an analogy, it is like saying that no matter how high-performance or cheap a new smartphone is, no one will buy it if the apps I use every day do not run on it. If developers refuse to port code written in the NVIDIA language because it is too troublesome and difficult, there is a risk that it will become a pile of scrap metal that only consumes 300 trillion won worth of chips .
The second is Samsung's number issue. It is true that Samsung was the first in the world to begin mass production of HBM4. But honestly, the current yield is in the 50 to 60 percent range. It means that nearly half of what is produced is thrown away . This waterway is structured so that the more you produce, the more you lose .
Nevertheless, the reason Samsung pushed ahead with the move was not to make money right now, but because they were in a hurry to be the first to have their name on the NBI AMD supply chain. The strategy is to prove technological prowess and secure a dominant position, then boost sales to recoup profits.
Samsung has reset its target to raise it to 85% by the end of the year, and the industry assesses that this is a realistic level . However, given that SK Hynix is aiming for over 80%, if we fail to narrow this gap, we could lose the position even after securing it because we cannot meet the volume on time.
Will Samsung's yield be able to handle the massive volume of over 100 million units of the 12GW HBM 4 promised by AMD ? Can you break through the software fortress called Schuli ? There is still a mountain to climb that is too high. However, one thing is certain: regardless of the outcome of this war, all sides are seeking out Korea .
Nvidia asked SK Inix for supplies, AMD asked Samsung for supplies, and Semoltman even flew in personally and asked for 900,000 units per month . No matter who makes the GPU or designs the SLC, ultimately nothing works without HBM, and two out of the three places that manufacture HBM are in Korea. Remember just one thing from today's story .
In an irreplaceable position, the other person comes to you first in an instant. Korean semiconductors are in that position right now. It is the same in your career and in your business. Creating a structure where they cannot function without me is the best negotiating power. When Lisa Su arrives in Seoul on March 18, the next chapter of this war opens. Let's watch together. Please subscribe and like.
댓글 0