SK hynix and the HBM Revolution

0:02 In July 2026, SK hynix raised a record $26.5 billion by listing its ADRs on the NASDAQ, the second-largest offering in US history. The company had always been well-known inside Korea. But now, they are taking a momentous step into the world spotlight. In today's video, how a company long seen as South Korea's semiconductor basket case clawed back from the brink to become an AI memory titan. Few companies in semiconductor history have seen the kind of turmoil over the years like SK hynix has.

0:38 The story of SK hynix can be traced back to the aftermath of the 1997 Asian financial crisis. As part of broader efforts to improve the financial stability of highly leveraged conglomerates, Korea pursued a restructuring initiative known as the "Big Deal," which encouraged the consolidation of overlapping businesses across several industries. The semiconductor industry was one of the sectors included in this initiative. In 1998, Hyundai and LG, then the 2nd and 3rd largest semiconductor companies in Korea, began discussions on combining their semiconductor businesses.

1:18 On paper, the combination offered a compelling opportunity. Together, the two companies could achieve a DRAM market share approaching that of Samsung and gain the scale needed to compete more effectively with Micron and major Japanese semiconductor manufacturers. Following extensive discussions on the structure of the transaction, Hyundai acquired LG's semiconductor business in 1999. The combined company subsequently became one of the world's largest memory manufacturers.

1:49 The transaction also brought on substantial financial obligations. In the years that followed, those financial challenges became one of the defining issues facing the newly combined company. At the time, Hyundai Electronics had approximately $13–14 billion in debt. Unfortunately in the second half of 2000, the industry suffered a classic overbuild and down-cycle after the Dotcom and telecom bubbles busted. Analysts called it the worst DRAM market since 1985, with some prices cut to below the cost of manufacture.

2:27 Semiconductors accounted for approximately 80% of Hyundai Electronics' revenue. In 2000, Hyundai Electronics generated around $1.5 billion in operating profit, but recorded a net loss of about $2 billion due to the financial costs of the significant debt taken on from the acquisition of LG Semiconductor, plus valuation losses. The following year, losses further worsened as DRAM prices sharply declined. With $3.1 billion of debt coming due in 2001, the company scrambled to raise money.

3:02 During this, they officially severed themselves from the Hyundai Group. So in March 2001, Hyundai Electronics took a new moniker: Hynix. Per a press release, the name derives from the combo of "high" and "electronics". Or "hi electronics", stressing the friendliness of its advanced technology. Unfortunately, the name did not alleviate the struggles - which only worsened in the economic turmoil after 9-11. In October 2001, Hynix's creditors - led by Korea Exchange Bank - initiated a comprehensive restructuring of the company.

3:40 After Hynix entered creditor control, the US firm Micron Technology approached to buy its memory chip fabs. After several rounds, Hynix's creditors and executives negotiated a non-binding Memorandum of Understanding to sell Hynix's memory business for 108 million Micron shares. The problem was that the vast majority of Hynix's revenue came from DRAM. Once the memory business was carved out, the debt-heavy units left behind stood virtually no chance of surviving on their own.

4:12 And Micron was not buying that memory business to run it. There seemed no plans to grow the Korean fabs. The point was to take that capacity off line and erase a competitor. The creditor banks holding the majority of the company's debt pushed ahead with the sale. But in late April 2002, Hynix's ten-person board unanimously rejected Micron's offer. Hynix said the offer undervalued the memory business, that the restructuring conditions were unacceptable, and memory prices were recovering.

4:45 There was also significant opposition from the public and other parties, who objected to selling the Korean company to the American company Micron. Even the Hynix executives who negotiated the MOU and initially supported it ended up voting NO. The rejection vexed Micron and the stock price. But Hynix resolved to escape the crisis on their own and survive. It is interesting to ponder what might have happened had they accepted Micron's offer. The following years saw the company sell off whatever it can to pay down debt.

5:19 In 2003, they sold their TFT-LCD panel display business to the Chinese company BOE Technology. BOE later becomes one of the world's biggest display and OLED makers. A year later, they spun off their system IC business, which later became MagnaChip. Today, they are a South Korean maker of power semiconductors. And together with ST Microelectronics, they opened a new fab in the city of Wuxi in Mainland China. To recover their technical reputation, Hynix kicked off a series of successful branding and quality initiatives called the Blue Chip, Prime Chip and Golden Chip grading systems.

6:01 After a brutal 2002 and 2003, things got back on track in 2004 and 2005. Profits rose to an all-time high. The stock climbed from a low of 11 cents per share to about $21. Many creditors got paid back. In September 2005, the New York Times publishes a complimentary article labeling Hynix the "comeback darling" - calling its turnaround a "miracle". Those familiar with K-dramas will probably recognize that as a bad omen.

6:35 2006 was a good year with strong unit prices and anticipated demand from a new Windows OS. But with rising competition, Hynix felt it had to keep up. In July 2007, Hynix and its new CEO Kim Jong-gap announce an aggressive 300-millimeter expansion plan in both DRAM and NAND. They pledged to add a new line each year for the next three years and become the world's largest memory-maker by 2017. However DDR2 DRAM prices were already then in free fall as it becomes clear that Windows Vista isn't triggering a super-cycle.

7:13 And then came the Global Financial Crisis. As the world economy melted down, Hynix reports its worst quarterly loss in eight years, with cash flow dwindling to dangerously low levels. Micron lost a billion dollars that year. Infineon/Qimonda in Germany collapses. Taiwan's DRAM industry gets largely bought up by Micron after the government declines intervention. In late 2008, Hynix - which all this time has still been creditor-owned - goes back onto life support.

7:46 Some months later, the company's banks put a controlling stake in the firm up for sale. SK Group was founded in 1953 as Sunkyoung Textiles. Over the following decades, it expanded from textiles into synthetic fibers, then into petrochemicals, and entered the refining business in 1980. That same year, when the US oil major Gulf Oil decided to withdraw from Korea's first refinery the Korea Oil Corporation, the government privatized the company.

8:19 Sunkyoung was selected to assume management. Why Sunkyoung? One reason was the long-standing relationship that the late Chairman Chey Jong-hyon had built with Saudi Arabia over several years, particularly with Oil Minister Ahmed Zaki Yamani. During the oil shocks of the 1970s, when securing stable crude oil supplies was a national priority, such relationships helped strengthen Korea's access to Saudi crude. As Sunkyoung prepared to enter the refining business, Chairman Chey secured both a commitment from Saudi Arabia for crude oil supply, and a $100 million loan guarantee from a Saudi bank to support the acquisition.

9:01 Energy remains one of SK Group's core business areas today. A decade later in 1994, Sunkyoung bought a 23% stake in Korea Mobile Telecom, a cellular phone service provider, then being sold off by the government. Their winning bid for the stake and the management rights that came with it offered $415 per share, four times the $100 share price at the time of the privatization announcement. Back then, it was seen as an overpay.

9:32 Korea Mobile Telecom was a niche player offering pages and car phones and faced imminent competition. Korea Telecom was the real telephone giant. Today, Korea Mobile Telecom is SK Telecom. And Korea Telecom is "KT". Yeah it tripped me up too. But Sunkyoung's then-chairman Chey Jong-hyon foresaw the mobile revolution coming to South Korea. They invested heavily into building a telecommunications network and made an early strategic commitment to CDMA technology.

10:06 In January 1996, SK Telecom launched the world's first commercial CDMA service. And the telecom firm became another growth engine and example of the late chairman's particular industrial foresight and savvy. Chey Tae-won is the nephew of the group's founder and son of the Chairman Chey Jong-hyon mentioned earlier. After working in the group's strategic office, Chey became chairman in 1998 after his father's passing from cancer that August.

10:37 In January that same year, the company changed its name from Sunkyoung to SK, part of an ambition to become a more global company. Soon after the creditors put Hynix up for sale, Chairman Chey began pondering to buy it. But why? SK had previously entered the semiconductor business. Back in the 1970s when SK was still Sunkyoung, the late Chairman Chey Jong-hyon - valuing the industry's high potential - established Sunkyoung Semiconductor.

11:08 But difficult circumstances at the time caused them to discontinue the business. Chairman Chey Tae-won thought it was time to dive back in. Semiconductors were rising once more and he believed that it can drive a new growth engine. But SK's own executives tried to tear Chey away from submitting a bid. They argued that chips had no synergies with the rest of the SK Group. Hynix had effectively gone bankrupt in 2001 and had been under creditor-led management for over a decade.

11:38 Even in 2011, the company reported an annual net loss, struggling to return to profitability. But Chey felt that the market had changed. After years of intense competition, the industry was consolidating. Germany's Qimonda had collapsed during the Global Financial Crisis, and Japan's Elpida was also in a precarious position. After such a wave of industry consolidation, being amongst the surviving companies can shore up against future market cycles.

12:10 Chey also saw the mobile market starting to take off - led by the emergence and rise of the iPhone and Android. The growth of this new form factor can counter-balance the booms and busts of the PC market. He also put in the due diligence. He personally studied semiconductors - which I find endearing - and consulted with industry people. Memorably, he flew to Taiwan to meet with Morris Chang, TSMC's founder. In a book published earlier this year titled "Super Momentum" (a major source for this video), Chey recalled that Morris' first question when Chey told him about acquiring Hynix was: "Are you going to do foundry?" Chey said no, and Morris relaxed considerably.

12:53 Morris told him that satisfying customers was the single most important capability, saying: "The deeper the downturn, the better you must connect with your customers; and in an upturn, never act as though you are above them." In the end, the chairman felt that he had learned enough to see the same opportunity that his father did with SK Telecom. He was all in - personally going to persuade each board member and directly leading the decision-making process. Even so, internal resistance meant that SK did not submit a bid until the deadline day.

13:28 Several months later in February 2012, SK completed its acquisition of the 21% controlling stake for $3 billion. With that, Hynix became SK hynix. After the purchase closed, Chey traveled to Icheon and met one-on-one with 100 of SK Hynix's executives for 1-2 hours. The big question he asked them was "What should SK hynix do well going forward and how?" The interviews revealed the company's problems as well as its path forward.

14:01 In the Super Momentum book, Chey explained the strategy: "What I really wanted to accomplish when we acquired Hynix was to transform it from a commodity memory producer into a mainstream semiconductor company whose products are indispensable" The lesson of the previous decade was clear: relying solely on commodity DRAM would not allow SK hynix to escape the industry's cyclical nature. In the DRAM market, competitors can be cross-supported by their broader electronics businesses.

14:32 SK hynix needed to make something unique. Chey hand wrote a memo laying out that SK hynix had to leverage its technology chops to add "economic value". To get that, they must work with customers closely from the design stage, as Morris said. And they need products that can be more than just another stick of DRAM. The idea behind HBM dates back over two decades. Bryan Black is colloquially known as the "Godfather of the Chiplet".

15:03 In 2006, he published a landmark paper titled "Die Stacking (3D) Microarchitecture". This paper discussed the advantages of "disintegrating" the microprocessor monolith and reintegrating them as a 3D die stack. Today, it is remembered for paving the way to chiplets. Ergo his nickname. But this paper also broached the benefits of stacking DRAM dies. The idea is simple: Compute performance is proportional to how much data you can pull out of memory and transport to the logic cores, i.e. bandwidth.

15:39 But moving data uses energy and all systems must operate within a particular power envelope. So either we tolerate having less data or starve the logic circuits of power. Both paths give us worse performance. Compute demands for graphics and emerging applications like deep learning were growing. And even the latest standards for graphics-oriented DRAM memory - GDDR5 - cannot keep up. This led AMD to approach SK hynix to co-produce a new industry standard architecture.

16:13 HBM stacks multiple DRAM memory dies on top of a logic "base die". This base die is fabricated using a logic process node and hosts the physical circuitry necessary to drive bits to and from the processor. In the center of the memory dies, we have a large region with thin vertical interconnects called Through Silicon Via, or TSVs, tunneling through the silicon die. The TSVs connect to microbumps, which are tiny soldered interconnects bonding the stacked DRAM dies together.

16:47 So if the TSV travels through the die then the microbump travels between dies. The design philosophy is that many slower channels will beat a few fast ones. A flagship graphics card using GDDR5 like Nvidia's GeForce GTX TITAN Black has 384 "lanes" moving at a blazing 7 gigabits per second. Total bandwidth thus is 2,688 gigabits or 336 gigabytes per second. By comparison, the HBM1 stack has eight memory channels - two in each of the four stacked dies - with each channel having 128 "lanes".

17:27 A GPU chip uses four HBM stacks so that gets you a total of 4,096 lanes. By itself, each HBM lane moves at just 1-2 gigabits per second, but since HBM has so many more lanes, each chip gets multiples higher overall bandwidth. You can imagine the Titan's GDDR5 channels as like the whitewater rapids of the Colorado River - the water blowing through fast and furious. HBM is like the Amazon River - a slow, lazy flow but very deep and miles wide.

18:02 The former looks cooler, but the latter moves far more water per second. And since the TSVs are substantially shorter than the alternative of long wires connecting to off-die DRAMs, we get better resistance and power consumption. HBM entered the market with AMD's new Radeon R9 Fury X GPU. At about 1,011 square millimeters, it is a big, ambitious chip positioned at the highest end of the market. It integrated 22 different dies from multiple vendors.

18:34 Bryan Black said at the time that it was just the beginning: "We will see more HBM in more AMD products. We didn't develop the technology just for this, we developed it for the future" To ensure this, SK hynix and AMD submitted the HBM spec to JEDEC, the leading consortium for standardizing semiconductor memories. The standard was adopted in 2013 - providing the foundation for a broad product ecosystem through collaboration with competitors.

19:06 Unfortunately, while most acknowledge that the Fury X GPU was a pioneer... And that its new HBM made it the fastest, most powerful single-chip GPU on the market, the card did not sell in high volumes. Production challenges and yield issues meant it was a struggle to get enough in the market. Chairman Chey later recalled meeting with AMD CEO Lisa Su afterwards and proposing several new initiatives - like maybe a new CPU. But what Dr.

19:37 Su and AMD most wanted was to continue developing HBM. From a purely financial perspective, more HBM was not an obvious investment at the time. The profits on this niche thing hadn't been large - SK hynix might not have even made back a profit on the line equipment - but Chey recognized that this is what their partner AMD wanted. So he kept the HBM development going despite the uncertain financial case for doing so. It was a strategic investment driven by close customer collaboration and a long-term vision focusing on something beyond just immediate profit.

20:16 Competition also might have had something to do with it. Just six months after the Fury's launch, Samsung Electronics announced that they too were producing HBM. Theirs was the second-generation iteration, HBM2. At that time, the HBM market was small, with the number of customers limited. If SK hynix doesn't deliver their HBM2, then Samsung wins that generation and perhaps the whole category. The race was on.

20:48 SK hynix showed off their first HBM2 wafers in March 2015 - and said that the chips would be ready next year. Presentation slides laid out a timeline with the anticipated release date in the middle of 2016. They also displayed a test wafer at Nvidia's big 2015 GPU Technology show. But internally, the development team was facing challenges and working hard to overcome them. HBM2 doubles performance, in part by raising the raw per-pin transfer speed and bandwidth per stack.

21:21 But after HBM1 failed to sell as well as initially hoped, the team worked to deliver better performance at a lower cost. So they gambled on going straight from a high 20-nanometer process node (2x) to a low 20-nanometers process node (2z). So skipping the mid 20s-nanometer 2y generation node. Smaller nodes means smaller dies means more dies from each wafer. QED cheaper dies. In late December 2015, ETNews reported that both Samsung and SK hynix were in the final stages of HBM2 development with high volume manufacturing beginning in the first half of 2016.

22:04 But when the product specifications changed at the last second, SK hynix had to abruptly pivot the design and things fell apart. TSV speeds fell to just 500 megabits per second - a quarter of the 2 gigabits per second target - primarily due to high capacitance. The plan had been to ship two HBM2 product variants - one with 2 gigabits per second and another with 1.6. But each new quarterly update pushed back the release. The higher end 2 gigabit product was eventually scrapped.

22:36 The rushed design change also meant no time to optimize packaging, which led to accidents. Dies were cracked during the stacking process. Or incorrectly cut out of their wafers. On average, HBM2 yields were about 50%, which is horridly low. So even after it was released, the product did not pass customer qualification. Then-CEO Park Sung-wook visited one customer to discuss the matter and received what was described as "bitter criticism".

23:08 Throughout the whole year-and-a-half long delay, the HBM2 team's top design executive was changed three times. The final product would not be delivered until 2018. And even after shipping it, misfortune poured on their parade. In the late 2010s, a major demand driver for HBM was crypto mining. After HBM2 finally made it to production, the company invested over $400 million into a new packaging fab called P&T4 to produce it. But then in 2018, the crypto market plunged into a new winter.

23:42 Bitcoin prices crashed 75% to the $5,000 range. Demand for mining GPUs and HBM crashed along with it. SK hynix's new packaging facility encountered challenges. The HBM2 experience left the team with a dilemma and a lesson: Cost or Performance? Choose one. You can't have it all. So as they embarked on the third generation of HBM - called HBM2E - in mid-2018, they decided to go all in on performance.

24:16 They doubled the number of TSVs and added thermal "dummy" bumps to accommodate temperature-induced physical stress. But the biggest change involved how they packaged the HBM die stack. Which, in my opinion, is pretty risky. Packaging HBM already adds 100 more process steps than with ordinary DRAM and takes twice as long. There is a lot here that can go wrong. And not to say that anything in the world of semiconductors is a cakewalk, but packaging is particularly galling because you only package good chips.

24:48 So a failure at packaging is truly fumbling the bag. As I said earlier, chips connect via an array of microbumps. A solder cap is applied on top of a copper pillar inside each bump. We then put the dies together and heat it. The solder cap melts - joining the microbumps and bonding the dies. After bonding, there remains these gaps of about 10-50 micrometers between the bumps. Such gaps cannot be left empty. To fill in those gaps, we add a material called "underfill".

25:20 The underfill protects from outside moisture and provides electrical isolation. Most importantly, it buffers against physical stress caused when die materials expand or contract at differing rates in different temperatures. After the bonding and underfill steps, we finally encapsulate the whole chip inside a protective mold in a step called "overmold". This completes the packaging. For HBM1 and HBM2, SK hynix used a method called "Thermal Compression With Non-Conductive Film", or TC-NCF, to do the bonding and underfill steps.

25:57 You can imagine it as like making a smash burger. Before bonding, we apply or "laminate" the eponymous non-conductive adhesive film - often made from a blend of epoxy and acrylic - onto a die. We then align those dies, stack them, and press down hard while also heating it to a high temperature. SK hynix applied up to 40 kilograms of pressure so imagine dropping a big sack of rice onto an HBM. When it works, TC-NCF combines the bonding and underfill steps into one step.

26:30 The use of a film rather than a paste offers excellent insulation - reducing the risk of electrical short-outs. It also produces a nice, even layer relatively free of voids. Which matters when stacking things, because uneven layers get you the Leaning Tower of Pisa. But TC-NCF can also be tricky. Doing it right requires the Goldilocks amount of various factors like bonding force, bonding temperature, temperature ramp-up, and how much time is given to the solder to bond.

27:03 (Dwell time) The biggest concern however is that TC-NCF only bonds one die to one die at a time. But HBM stacks five or nine dies - the four or eight DRAM dies plus the one logic base die. So SK hynix must heat and compress the stack multiple times sequentially. It is like making a triple-decker smash burger by smashing on each patty one at a time. This not only takes a long time, but probably something will go wrong.

27:32 And if you are making a nine-die stack, one failed TSV connection means nine dies discarded. Worse yet, HBM2E increases the number of TSVs and thermal dummy bumps on the die as compared to HBM2. This means needing to apply yet more pressure to ensure that all the bumps bond. All that pressure and heat means more cracked dies and quality issues. It doesn't take an Einstein to recognize that this won't scale. More importantly, even if SK hynix succeeds then what they end up with is not significantly different from what existed.

28:09 Wouldn't that end up the same way as it always had before? So SK hynix needed to go off the beaten path. And fortunately, the packaging industry did have something. In 2019, SK hynix tried a method called "mass reflow". This involves first stacking all the DRAM dies, then melting and bonding their microbumps in just one step. No more making the smash burger sequentially. This worked. But then what about the underfill step to fill the empty gaps between the microbumps?

28:41 At first, the team explored a method called Capillary Underfill, or CUF. Together, Mass Reflow and CUF make MR-CUF. Capillary underfill is a well-known technique in packaging. After mass reflow, we inject the underfill material as a liquid at the chip's edges. Then capillary action brings that liquid in through all the gaps. Once done, we cure the liquid with heat to solidify it. MR-CUF requires repeating the capillary action and curing for each die in the stack, one after the other.

29:17 Unfortunately this process takes way too long and after half a year, SK hynix decided not to proceed with it. The team then turned to a newer method then making waves in the mobile chip packaging industry called Molded Underfill, or MUF. Those guys adopted it because they were frustrated with the same throughput issues. MUF combines the underfill and overmold encapsulation steps using a single, special mold material. In vanilla MUF, an Epoxy Molding Compound or EMC mold material is delivered as a resin with tiny filler particles mixed into it.

29:54 The EMC flows not only in between the gaps but also over the whole chip. Once fully distributed, we heat-cure it. Voila, both underfill and overmold encapsulation, no more waiting for capillary action. It just makes sense... if it works. But the risk of creating voids in between the bumps is high. And MUF is a flip-chip technique only previously used for small, single dies. Never for eight-tall DRAM die stacks.

30:25 SK hynix needed to figure this out on their own. Thus, SK hynix embarked on this new packaging technique called MR-MUF. Since no equipment or material then existed for it, SK hynix had to co-develop everything from scratch. The molding equipment, for example, was produced with Japan's Towa. The EMC material was developed in partnership with the Japanese company NAMICS, and required some fifty iterations. SK hynix also had to convince the customers to accept this new, unfamiliar method.

30:59 Unlike commodity DRAM, HBM is highly customized. If MR-MUF flops and the HBM delays, then the customer cannot just buy a replacement off the shelf. When first told about MR-MUF in early 2019, potential customers - perhaps traumatized by the HBM2 debacle - expressed skepticism. Can this actually work? But it did and SK hynix successfully scaled MR-MUF for HBM2E, announcing it in August 2019.

31:30 Qualification tests found SK hynix's HBM2E performing insanely fast in bandwidth. Customers were pleasantly surprised, and specced that HBM2E into their GPUs. Two years later, SK hynix announced their next-generation HBM3 with 8-layer and 12-layer versions capable of 819 gigabytes per second. But did people really need such a product that badly? Before ChatGPT, the HBM market segment was tiny - just 1.5-2% of the larger DRAM market.

32:07 Few wanted these expensive, performance-overkill products - especially after the crypto winter killed demand for mining GPUs. At some point, Samsung's management probably began wondering why they were investing so much into this. In the late 2010s, industry people seemed to agree that the future was in serving AI and machine learning models at the edge. Internet of Things devices, 5G wireless and such. So at a critical point, Samsung chose to focus on HBM-PIM (Processing-in-Memory) instead of HBM3.

32:44 The concept sought to address the von Neumann bottleneck by integrating a programmable AI computing engine directly into memory. The concept has its strengths, but unfortunately required significant software effort and - as an "edge inference" solution - was more tuned for speech recognition or machine vision rather than Large Language Models. Thusly on November 30th, 2022 - the day of ChatGPT's release - there was just one company providing the high-performance HBM3 critical for the Nvidia GPUs that everyone suddenly wanted: SK hynix.

33:22 It was a moment for validation for SK hynix and Chairman Chey, who continued to support the HBM development effort for high performance compute despite its uncertain near-term and future profitability. They took a hit from the market for such, but the ChatGPT boom of 2022 showed that those choices - and its long term customer-focused strategy - paid off. Considering the craziness going on right now, it is easy to forget how long it took for the ChatGPT boom to bloom.

33:53 In 2022, the semiconductor industry was suffering a post-COVID hangover. Both DRAM and NAND prices drastically fell. The quarter that ChatGPT came out, the company turned a $1.5 billion operating loss - its first in 10 years. Three more quarterly losses would follow. But by Spring 2023, things on the HBM side were starting to get interesting. ChatGPT was mushrooming and customers were requesting more HBM, as soon as possible.

34:24 Every time Chairman Chey met with someone, they would ask him to move up the development schedule even more. He joked being afraid to meet with anyone else. SK hynix scrambled for more HBM capacity. The aforementioned Package and Test Fab 4 - the one that initially went underutilized after the Crypto Winter - is now running at full tilt, but even that was not enough. With no time to build new fabs, SK hynix rushed hundreds of pieces of equipment into an empty cleanroom on the second floor of their Cheongju M15 fab, originally planned for NAND.

34:59 This front-end space was rejiggered to package HBMs in a span of just six months. The real leap of faith however was the Fab M15X expansion. In September 2022, a few months before ChatGPT, SK hynix announced a fab expansion in Cheongju next to M15. Soon afterwards though, the heavy financial losses from the post-COVID downturn caused SK hynix to put the project on hold. They were already doing a new fab at the Yongin area, so the expansion seemed imprudent.

35:33 But Yongin won't be done until 2027. By early 2024, HBM demand was growing faster than anyone had projected. So as SK hynix was busy converting M15's vacant second floor to back-end HBM packaging, the company also officially restarted the M15X project with a $4 billion investment, pivoting it to producing DRAM. The first chips from that new fab expansion are scheduled to hit this year - and I reckon it cannot come any sooner.

36:06 SK hynix's July 2026 ADR listing is a pretty significant step in Asian business history. The stock price is going to go wherever it is going to go. But I think the bigger thing is that historically few Korean companies have tried to step outside of Korea and into the US markets like this. SK's savvy entrances into the energy and telecom businesses during the last Chairman's time turned them into Korea's fifth largest conglomerate. Chairman Chey's purchase of SK hynix has made them the second largest.

36:38 It's time to look beyond Korea's borders. So the listing is a step towards a more globally impactful company. They have already started building a $4 billion advanced packaging fab in the US state of Indiana. And Chairman Chey's making appearances at the GTC 2026 and AI conferences. Right now, DRAM prices are high, and that is unfortunate. With supply booked out for the next two years - and end demand for more intelligent models continuing to surge - more memory fabs are needed.

37:11 And SK hynix is going all in. They announced a mid to long-term plan to invest 1.1 quadrillion won (how many times do I get to say that number on this channel?) on its clusters in Cheongju and Yongin plus a new one in Korea's southwestern region. Maybe a US cluster in the future? A guy can dream. Few companies have swung this hard between crisis and success. After years of losses, SK hynix was nearly written off. SK bought it anyway over many internal and external objections.

37:43 It made HBM and for years, HBM sat as this niche product few cared about. But the company kept working on it, even when the market gave no sign that it will pay off. In an interview for Super Momentum, Chairman Chey recalled promising his father that he would build a company that can meet or surpass Japan's top tier companies. Thanks to the HBM Revolution, SK hynix has achieved that. All right, everyone. That's it for tonight. Thanks for watching.

38:11 Subscribe to the channel, sign up for the Patreon, and I'll see you guys next time.