Korea memory in 2026: Samsung versus SK Hynix, NVIDIA qualification, and the HBM share war

The HBM profit pool and how it concentrated #

High bandwidth memory was a niche product in 2022 and a dominant input to AI accelerator economics by 2024. TrendForce sized the HBM market at roughly USD 16 billion in 2024 and projected USD 33 billion in 2025, more than a doubling driven by the Hopper to Blackwell transition at NVIDIA, parallel AMD MI300 and MI325 ramps, and the first material non NVIDIA buyers in Google TPU and AWS Trainium. HBM rose from roughly 10 percent of total DRAM revenue in 2023 to above 30 percent by 2025. On a profit basis the share is higher still because HBM gross margins have run roughly two to three times the conventional DRAM blended margin.

Vendor share is the variable that matters. TrendForce reported fourth quarter 2024 HBM bit shipments at roughly 53 percent SK Hynix, 38 percent Samsung, and 9 percent Micron, a near reversal of the conventional DRAM ranking. The share inversion routes the AI memory rent to SK Hynix rather than to the larger national champion Seoul has organized industrial policy around for two decades.

Qualification timing was destiny #

The HBM market is not a price market. It is a qualification market. NVIDIA, AMD, and the cloud accelerator buyers run multi quarter sample, reliability, thermal, and yield qualification cycles before any HBM stack is approved for a specific GPU SKU. Once a supplier is in, share allocation across that SKU's life is largely set. SK Hynix understood this earlier than competitors and committed capacity to HBM3 and HBM3E ahead of confirmed demand, a bet that turned the number two DRAM player into the structural leader of the AI memory cycle.

SK Hynix qualified HBM3E 8 high at NVIDIA in the first quarter of 2024 for the Hopper H200 ramp, and HBM3E 12 high through the third quarter of 2024 for Blackwell B200 and GB200 NVL72 systems shipping in volume into 2025. Micron used a focused HBM3E effort to qualify in the third quarter of 2024 and lock a meaningful second source slot at NVIDIA, particularly for the United States manufactured supply that is increasingly preferred under CHIPS Act and export control conditions. Samsung qualified HBM3E 8 high only in September 2024, several quarters behind schedule, and HBM3E 12 high qualification remained pending into the first quarter of 2025 according to Reuters Seoul and Financial Times reporting.

Causes of the Samsung delay were thermal and power on the surface, with higher thermal output and yield issues forcing design iteration through 2024. The deeper cause was organizational: Samsung Memory, Foundry, and System LSI shared engineering and capital plans while SK Hynix made focused HBM bets, and foundry losses on the 3 nanometer ramp pulled management attention away from the memory roadmap at the moment HBM became the most strategic product line in the industry.

The Samsung response: foundry separation, succession, and HBM4 catch up #

Samsung's response began with structure. In March 2024 the foundry business was separated from Samsung System LSI in a broader reorganization, acknowledging that the captive foundry plus design plus memory model had become an obstacle to focused execution. The separation gave the foundry a dedicated profit and loss line, clarified accountability for 3 nanometer and 2 nanometer roadmaps, and freed the memory business to make HBM specific capacity and packaging decisions without trading off against foundry capital needs. The Lee Jae yong succession, culminating in his February 2024 acquittal and return to active strategic oversight, was the parallel governance event that allowed the reorganization to move at speed.

On the product side Samsung is running an HBM4 catch up plan. HBM4 12 high samples are scheduled through the fourth quarter of 2025 with mass production targeted for the third quarter of 2026, while SK Hynix has guided to first HBM4 samples in March 2024 and mass production by end 2025. The gap is roughly two to three quarters, narrower than HBM3E but still material in a market where the qualification window for the next NVIDIA generation closes on a known schedule. Samsung is also accelerating the move to a logic die under HBM4, raising the strategic value of having an in house foundry that can deliver competitive 4 or 3 nanometer base dies on schedule.

The financial gap is the table below. Samsung memory and foundry combined operated at quarterly losses estimated near KRW 1 trillion through the trough of 2024, against an SK Hynix that posted KRW 19.8 trillion of revenue and KRW 8.1 trillion of operating profit in the fourth quarter of 2024 alone, a 74 percent year over year revenue increase driven almost entirely by HBM. The question is whether Samsung recovery happens in the HBM4 cycle or only in HBM4E and HBM5, because a second consecutive miss would change the long term valuation of the memory business and the political economy of Korean industrial policy.

Yongin, CHIPS Act, and the geography of HBM capacity #

Korean industrial policy has organized around a single answer to the AI memory transition: build the Yongin semiconductor cluster. MOTIE approved the Yongin general industrial complex with a planned investment of approximately KRW 122 trillion, roughly USD 89 billion at first quarter 2025 exchange rates, anchored by SK Hynix with Samsung capacity adjacent at Pyeongtaek and Hwaseong. The cluster is structured to host four SK Hynix fabs and a supplier ecosystem of materials, equipment, and packaging firms, with first fab operations targeted in the second half of the decade.

Outside Korea, both companies took CHIPS Act awards. SK Hynix received approximately USD 458 million of grant funding plus loan support tied to its Indiana advanced packaging and HBM facility, with roughly USD 6.4 billion of total project investment committed. Samsung secured approximately USD 6.4 billion of CHIPS Act grant funding for its Taylor, Texas fab, oriented primarily at advanced foundry rather than HBM but central to the integrated AI accelerator supply chain. Both awards are subject to milestone based disbursement.

Export controls reshape the geography again. The Bureau of Industry and Security rule issued in December 2024 brought HBM2e and above under license for export to China, including foreign direct product controls that capture HBM produced abroad if it incorporates United States technology. Samsung historically had meaningful HBM exposure to Chinese accelerator customers including Huawei and the broader Cambricon and Biren ecosystem, while SK Hynix had aligned earlier toward NVIDIA and the United States hyperscaler channel. The control regime accelerates a structural reallocation of HBM capacity away from China as a destination, which is favorable for SK Hynix's existing customer mix and creates a transition cost for Samsung that compounds the qualification timing issue at NVIDIA.

Macro and the inference demand pull #

Korean memory exports were the most visible macro story of the AI cycle. MOTIE customs data showed semiconductor exports up 43.9 percent in 2024, with memory the dominant contributor. Won weakness through the first quarter of 2025, with the spot rate reaching roughly KRW 1,470 per United States dollar in political turbulence following the December 2024 short lived martial law declaration, added to won denominated revenue while pressuring imported equipment and materials costs. The Bank of Korea response, a measured rate cutting cycle into 2025, supported domestic credit conditions for the Yongin buildout while leaving the won weak relative to the dollar.

Demand acquired a new vector with DeepSeek V3 and R1 in early 2025 and the inference cost decline that followed. Memory bound inference workloads, where the binding constraint is HBM bandwidth rather than logic FLOPS, became the dominant economic question for cloud providers. Frontier reasoning models with long context windows and large key value caches consume HBM bandwidth at a rate that scales close to linearly with sequence length, pushing HBM allocations into a structural shortage through 2025. That shortage was the binding constraint on NVIDIA Blackwell volumes through several quarters and on AMD MI325 ramps in parallel.

Strong export earnings, HBM shortage, and a procurement environment where GPU OEMs cared about supply assurance more than headline price gave SK Hynix exceptional pricing power. The same combination penalized Samsung twice: missing the price up cycle on HBM3E and facing a more expensive catch up because qualified suppliers held price discipline. By the time HBM4 arrives in volume in 2026, the question is whether three qualified suppliers and disciplined NVIDIA dual sourcing compress HBM margins back toward conventional norms or whether AI demand growth keeps the rent pool intact.

Implications for OEMs, memory makers, foundries, and policymakers #

For NVIDIA, AMD, and the cloud accelerator OEMs the lesson is that HBM dual sourcing must become a first class procurement objective rather than a fallback. Single source dependence on SK Hynix concentrates risk in one corridor of Gyeonggi province. The procurement implication is to invest more buyer engineering in supplier qualification, publish credible volume commitments to second and third sources, and design GPU bill of materials with HBM stack interchangeability as a hard constraint.

For SK Hynix the question is how to extend the lead. The answer is to use the rent pool to fund HBM4, HBM4E, and packaging and base die integration with TSMC ahead of competitors, and to use the Yongin buildout to lock in capacity competitors cannot replicate inside the qualification windows of the next two NVIDIA generations. The risk is complacency, the same risk that converted Samsung from leader to number two over the HBM3 cycle. For Samsung the question is sequencing: the recovery path runs through executing the foundry separation, hitting the HBM4 12 high mass production target in the third quarter of 2026 within plus or minus one quarter, and rebuilding customer trust that qualification delays eroded.

For TSMC the HBM transition raises the strategic value of base die foundry work, because HBM4 places a logic die fabricated on an advanced node under the memory stack. Samsung Foundry has an opportunity to use captive base die supply to differentiate Samsung HBM, conditional on the foundry roadmap delivering on schedule. For Seoul the priority is to keep the Yongin cluster on schedule, preserve United States customer access, and manage export controls so Korean firms retain the qualified hyperscaler channel without losing the legitimate non controlled portion of the China memory market. For United States policymakers the priority is to convert CHIPS Act disbursements into actual operating capacity in Indiana and Texas on schedule, because the alternative is that the AI accelerator supply chain remains structurally dependent on a single Korean corridor for its highest value memory input.

Sources #

• TrendForce, HBM market reports and vendor share estimates 2024 to 2025
• SK Hynix, fourth quarter 2024 earnings release and investor presentation
• Samsung Electronics, investor relations earnings releases and reorganization disclosures
• Micron Technology, fiscal 2024 and 2025 earnings releases on HBM3E qualification
• NVIDIA Corporation, annual report and quarterly filings on supplier and component disclosures
• Reuters Seoul, coverage of Samsung HBM3E qualification at NVIDIA, September 2024
• Financial Times, reporting on Samsung memory and foundry restructuring 2024
• Ministry of Trade, Industry and Energy, Republic of Korea, Yongin semiconductor cluster designation
• United States Department of Commerce, CHIPS Act preliminary memorandum of terms for SK Hynix Indiana
• United States Department of Commerce, CHIPS Act preliminary memorandum of terms for Samsung Texas
• Bureau of Industry and Security, December 2024 export control rule on HBM and advanced computing items
• Bank of Korea, monetary policy decisions and exchange rate statistics