9-fold surge, net profit surpasses Samsung — How did SK Hynix go from “near bankruptcy” to becoming the world’s most profitable chip company?

2008 end, Germany’s Qimonda declared bankruptcy. That was the world’s second-largest DRAM manufacturer, and it collapsed in less than one quarter. When the news came out, the whole industry was asking who would be next. Analysts started listing names; Hynix appeared frequently— that year, it had a net loss of 4.38 trillion KRW, DRAM average prices dropped 35% in a single quarter, barely surviving on bank debt roll-overs. No one could have guessed then that, 17 years later, this company’s single-quarter operating profit margin would surpass Nvidia. In Q1 2026, Hynix’s operating profit margin was 72%, higher than Nvidia (65%), higher than TSMC (54%), the highest quarterly profit margin in global chip industry history. Its stock price increased over nine times that year, with a market value of around $800 billion. In 2025, its annual operating profit surpassed the entire Samsung Electronics for the first time. From a bankruptcy list to the world’s most profitable chip company, what happened in between? **Two Near-Deaths, and an Overlooked Byproduct** Hynix nearly died twice, and survived both times for less-than-glorious reasons. The first time was 2001. Aftershocks of the Asian financial crisis combined with a DRAM price collapse; the company accumulated losses of over 8.3 trillion KRW over four years, debts piled to 10 trillion KRW, and it survived due to a debt restructuring led by Korea Development Bank. It was then subject to anti-subsidy investigations by the US and EU over "illegal state aid." The second time was 2008. Debt roll-over again — too-big-to-fail. Two brushes with death left an overlooked byproduct: Hynix didn't have the capacity to scatter resources like Samsung. Samsung could invest in DRAM, NAND, phone screens, appliances at the same time; Hynix could only concentrate its entire engineering force on its core domain. In 2012, SK Group acquired 21% stake for 3.37 trillion KRW, the largest corporate acquisition in Korean history. For the first time, Hynix had a stable shareholder to support long-term R&D investment. Without this step, subsequent bets wouldn't have been possible. **Betting on a Technology Without a Market** Hynix started studying bandwidth issues earlier than most competitors. In 2006, the company began researching the feasibility of TSV (Through-Silicon Via) technology. At the time, their only conclusion was: processor computational speed keeps growing, but memory bandwidth grows far slower—this “bandwidth wall” would eventually become the biggest bottleneck in computing. When the market would explode, nobody knew. In 2008, Hynix began collaborating with AMD, developing the earliest internal prototype, codenamed "HBM 0". In 2013, they launched the world’s first HBM chip, and JEDEC added HBM to the industry standard that month. In 2015, AMD’s Fiji GPU came equipped with Hynix HBM. The market’s reaction: expensive, niche, unnecessary. GPU and server vendors virtually did not follow. Even tougher was the second generation. The performance specs of HBM2 didn't meet design targets, forcing a total redesign — architecture, process, packaging, all overhauled. Engineers recalled a pervasive internal feeling: “We have already failed.” Hynix didn’t stop. Institutional support played a role. After SK Group’s 2012 acquisition, company decision culture changed noticeably— the book "Super Momentum" on Hynix’s HBM history describes it: engineers’ judgment was prioritized ahead of financial calculation. Wafer fab investments originally shelved due to unclear markets were restarted, HBM R&D budgets weren’t cut during the coldest market years. From 2011-2022, Hynix invested about 86 billion KRW in HBM R&D, and about 1.5 trillion KRW in equipment and plants— most during a period when HBM had virtually no commercial orders. Samsung also invested in HBM— Nvidia’s V100 and some AMD products used Samsung HBM2E. But in 2018, an event quietly changed everything. Jensen Huang secretly visited Samsung, suggesting three collaboration directions: jointly develop advanced HBM, transfer part of foundry orders from TSMC to Samsung, build CUDA software ecosystem together. Each direction meant Nvidia would tie its core supply chain to Samsung. All three were rejected. At the time, Jay Y. Lee was under judicial investigation; Samsung had no one to sit down and discuss decade-long strategy. After being rejected, Jensen Huang immediately contacted Hynix. Tech media TweakTown reported a detail: after rejection, Huang said, "Samsung doesn't have anyone who can discuss long-term strategy with me." The next year, 2019, Samsung management made what now appears a very costly decision: concluding HBM market growth was overestimated, disbanding its internal HBM R&D team, focusing resources back on standard DRAM. Some core engineers soon joined Hynix. That year, Hynix’s HBM patent accumulation was peaking. The rest became industry lore: Hynix obtained Nvidia H100’s exclusive memory contract. **Samsung’s HBM Fails to Meet Standards** May 2024, Reuters exclusively revealed a secret ongoing for over a year: Samsung’s HBM3E chips repeatedly failed Nvidia’s qualification tests, due to overheating and excessive power consumption. The news puzzled markets. Samsung is the world's largest semiconductor company, with advanced manufacturing nodes, so why can't it pass client tests for a critical product? The answer: packaging process. The two companies took radically different routes in HBM stacking methods. Samsung used NCF (Non-Conductive Film) — hot-pressing thin film to bond chip layers. This works fine with few layers, but as HBM stacked from 8-layer to 12, 16 layers, uniformity became harder to control, heat trapped between layers couldn't dissipate, power consumption spiraled. Hynix used MR-MUF (Mass Reflow Mold Underfill) — injecting liquid epoxy from the side, curing as a block, reducing voids and enabling more uniform heat dissipation. Adding each layer caused less yield drop compared to NCF. At HBM3 level, Samsung’s yield was 10-20%, Hynix’s 60-70%. A 50-point yield gap means, for same wafer input, Samsung's effective capacity is only a third of Hynix’s. This gap can't be bridged quickly by capital investment— it’s due to process-route choice plus years of mass-production experience, not money. Samsung later announced it would consider moving to MR-MUF— a process Hynix has used for over ten years. **Only Four Years: From H100 To Surpassing Samsung** After securing the Nvidia order, numbers began to change rapidly. June 2022, Hynix mass-produced the world’s first HBM3, exclusively supplying Nvidia H100, with a bandwidth of 819GB/s, a 78% increase over previous generations. The H100 drove Nvidia’s market cap from less than $500 billion to $3 trillion— every H100 contains Hynix memory. December 2023, both sides signed an HBM3E priority supply agreement, locking in Blackwell platform (B100/B200). That year, due to industry downturn, Hynix lost 7.7 trillion KRW, but capacity was already set for the next boom. March 2024, Hynix was first to mass-produce HBM3E. TechInsights later tore down Nvidia B200, confirming memory supplier: SK Hynix. 2025, operating profit of 47.21 trillion KRW, first time surpassing all of Samsung. Q1 2026, operating profit margin 72%. HBM market share: Hynix 57-62%, Samsung 22%, Micron 21%. Nvidia sources about 90% of HBM from Hynix. In 2008, when it was nearly dead, no one wrote that this number would ever appear. **5.9 Times—The Market Still Uses an Outdated Framework** Hynix’s current forward P/E is about 5.9x. TSMC 26.5x, Nvidia 24.5x, Micron 11.2x. Both TSMC and Hynix are irreplaceable nodes in the AI supply chain, but one is priced as “AI infrastructure,” the other is still treated as a “memory cycle stock.” This pricing logic historically made sense—DRAM was the classic commodity, prices swung wildly, low valuation justified cycle risk. But HBM isn't DDR5. Each HBM generation specification co-defined with Nvidia/AMD, switching suppliers is nearly like rebuilding production lines, supply is constrained physically, not just by price. If Hynix maintains 60%+ share in HBM4 and relationship with Nvidia stays strong, 5.9x means the market gives it zero structural premium, pricing it as a generic DRAM cycle. When this disconnect is corrected depends on the trigger. Failure conditions are real: If Samsung overtakes HBM share in HBM4 phase, or AI capital expenditures contract systemically, the moat assumption needs a rethink. If Samsung’s HBM4 gets certified promptly and share shifts, the market’s first reaction to Hynix is likely short-covering rather than trend-shift selling— structure wobbles but doesn’t overturn, read: cyclical pullback, not valuation framework collapse. Conversely, if Hynix keeps leading in HBM4, the first step in repricing will likely come from overseas investors more familiar with “AI infrastructure,” not local funds. TSMC’s shift from cycle stock to platform stock followed this path. **Samsung HBM4 Mass Production Will Be Next Story’s Prologue** Samsung announced mass production of HBM4 starting February 2026, engineering samples sent to Nvidia for testing September 2025. If certified and supplied in Q2, Samsung may claim about 30% share, and Hynix’s “near-exclusive” status will loosen— but loosening and overturning are two different things. The real variable is Samsung’s process choice. For HBM4, it's using TCB (Thermal Compression Bonding), betting next on Hybrid Bonding—a technical direction for HBM beyond 16 layers. But current hybrid bond yields are only around 10%, mass production schedule remains uncertain. Only when hybrid bonding matures will Samsung pose a structural threat to Hynix— until then, it's still early. Hynix HBM4E samples are expected to ship to Nvidia in late 2026; this will be the check-point for whether it can continue to dominate the next generation. The pace of Nvidia Rubin platform rollout will also influence demand, any delay narrows the window for the whole chain. These signals arrive in batches in 2026; together, they sketch out if, and when, the structure will loosen. But the whole framework itself—who can be Nvidia’s main memory supplier—was not even a conceivable question back in 2008. That bankruptcy list from 2008 now seems like a strange prophecy— survivors gained extra years to keep betting; those who fell exited forever. Qimonda died, Hynix was forced to focus by debt— two paths split, for even less than inspiring reasons. Seventeen years later, Hynix’s single-quarter profit margin surpassed Nvidia’s. The oddest thing about this: at the time, nobody could see this outcome, not even Hynix itself. The “bandwidth wall” hypothesis was correct, but the engineers who started TSV research in 2006 didn’t know that wall would appear around 2023 via AI compute boom, nor that this moment would mesh perfectly with the rift between Nvidia and Samsung. Forced focus, rivals’ management vacuum, AI demand erupting at the right time—all three happened at once. Missing any one, this story wouldn’t reach here. The next company to walk this path is likely, at this moment, quietly filling up their patent database in some unknown niche, with no market and no orders. Risk Warning & Disclaimer The market involves risk, investments should be cautious. This article does not constitute personal investment advice and does not consider the individual investment goals, financial situation, or needs of any particular user. Users should consider whether any opinions, views, or conclusions herein fit their circumstances. Invest accordingly, at your own responsibility.