Samsung Semiconductor Spotlight

Samsung Semiconductor runs three businesses that compete with and depend on each other. Samsung Memory is the world's largest DRAM producer by volume and one of two leading suppliers of high-bandwidth memory (HBM) for AI GPUs. Samsung Foundry is the second-largest leading-edge foundry by revenue share, but yield problems at SF3 (3nm) cost it major customers between 2022 and 2025. Samsung LSI, the internal chip design division, designs the Exynos mobile SoCs that go first onto each new foundry node, so their success is what external foundry customers look for.

In 2026, Samsung is regaining HBM share with HBM4 after the HBM3e qualification problems of 2024, using a 4nm base die and customer acknowledgment that "Samsung is back." SF2 (2nm gate-all-around, or GAA) is its most important foundry node in a decade, and the Taylor, Texas fab adds U.S. production backed by the CHIPS Act. Samsung is also giving HBM4 base die production priority over external customers for foundry capacity, which limits how much it can offer outside buyers.

Samsung Semiconductor Supply Chain Snapshot

• HBM market position: SK hynix held 53% and Micron 11% in Q3 2025. Samsung has about 30% of NVIDIA Vera Rubin HBM4 allocation, led Vera Rubin HBM4 qualification as of April 2026 and prices HBM4 at parity with SK hynix, against a 30% discount on HBM3e.
• HBM4 technology: competitors use 12nm-class base dies and 1b DRAM, while Samsung uses 1c DRAM (sixth-generation 10nm-class), making it the only HBM4 supplier with both an advanced base die and the latest DRAM generation. Pin speed targets 13 Gbps, hybrid copper bonding supports 16 or more layers, and HBM4E at 4.0 TB/s and 16 Gbps was unveiled at GTC in March 2026.
• DRAM production: 1c yield targets 80% to 90%, and output is rising from about 80,000 wafer starts per month (WSPM). Pyeongtaek P4 converted from foundry to 1c DRAM at 80,000 WSPM and is expanding, GDDR7 is in production for NVIDIA RTX 50 and AMD RX 9000, and V9 NAND is in production.
• Foundry SF2: Exynos 2600 (Galaxy S26 in Europe and Korea) is the lead product. Yield started about 30%, reached 50% to 60% by November 2025 and targets 60% to 70%, and the wafer price is about 33% below TSMC N2 at $30,000. SF2P (second-generation 2nm with improved backside power delivery) has a complete process design kit (PDK) with Exynos 2700 in 2027, SF2Z (full backside power delivery) is due in 2027 and SF1.4 (1.4nm vertical GAA) targets 2028.
• Foundry external customers: Tesla AI6 ($16.5 billion through 2033, SF2P, Taylor anchor customer); NVIDIA Groq 3 LPU (4nm, the first NVIDIA AI platform chip at Samsung); MicroBT and Canaan crypto mining ASICs (SF2); Charbright (4nm, U.S. AI startup); DeepX (2nm, Korean AI startup). Qualcomm 2nm talks are not finalized, and Exynos 2600 is the internal anchor.
• Taylor, Texas campus: Fab 1 costs $25 billion or more, began pilot operations in February 2026 and EUV equipment testing in March 2026. The 1,200-acre campus has a $6.4 billion CHIPS Act grant and is dedicated to Tesla AI6, with a second fab planned for added AI6 demand, within $40 billion of total Texas investment. Co-CEO Jun Young-hyun said 2nm at Taylor was ahead of schedule in March 2026.
• HBM4 base die priority: the Pyeongtaek foundry facility runs 30,000 WSPM, and its lines are above 90% utilization on internal orders.
• OpenAI HBM4 supply: Titan is OpenAI's in-house AI chip for H2 2026, designed with Broadcom and built at TSMC, and OpenAI ranks third in Samsung HBM4 allocation after NVIDIA and AMD. The win positions Samsung for Titan Gen 2 and Gen 3, and Samsung HBM4 output in 2026 exceeds 5.5 billion gigabits.
• Pyeongtaek campus: P1 makes HBM3e and NAND, P2 1b DRAM (converting to 1c), P3 is installing 1c DRAM equipment and P4 converted from foundry to 1c DRAM at 80,000 WSPM. HBM4 sells for more than $600 per 12-high stack, well above external foundry margin per wafer.
• Financial trajectory: Q4 2025 operating profit is expected at about 20 trillion won (up more than 200% year on year), with memory above 16 trillion won, the highest quarterly memory profit in Samsung's history; foundry is still loss-making. HBM revenue is expected to triple in 2026, Morgan Stanley forecasts EPS up more than 150%, and total HBM shipments in 2026 reach 11.2 billion gigabits, about half HBM4.

Samsung HBM4 Recovery

SK hynix won NVIDIA's HBM3e launch partnership for Hopper and Blackwell by delivering reliable, high-yield parts before Samsung finished qualification. Yield management and qualification delays gave SK hynix a six- to nine-month head start, and Samsung had to price HBM3e about 30% below SK hynix to win any NVIDIA allocation, which erased most of its profit on one of the highest-margin products in memory.

For HBM4, Samsung builds the base die on its own 4nm foundry process, while SK hynix and Micron use 12nm-class nodes. The HBM4 base die handles error correction, training modes, power management and signal integrity at 10 to 13 Gbps, all of which gain from a smaller node with more metal routing layers. Competitors cannot easily match it without committing their own foundry capacity to base die production.

Samsung began the first commercial HBM4 shipments in February 2026, led NVIDIA Vera Rubin HBM4 qualification as of April 2026, and its CEO said in his January 2026 New Year address that customers were saying "Samsung is back." Its 30% of Vera Rubin HBM4 is smaller than SK hynix's 70% but sells at parity pricing. At more than $600 per 12-high stack, that share is worth billions of dollars in high-margin revenue Samsung could not earn on HBM3e.

Memory technology leads rarely last more than a generation. SK hynix trailed Samsung in DRAM a decade ago and caught up, and Micron has held DRAM leads at times. Samsung's longer-term opportunity is custom HBM, where base die logic is designed for a specific customer's AI accelerator. Custom HBM is the highest-margin HBM tier, and Samsung's combined foundry and memory operations give it an advantage there before SK hynix and Micron move their base dies to newer nodes. HBM technology is covered on HBM.

Samsung Foundry from SF3 to SF2

Samsung was the first to produce GAA transistors commercially, at SF3, but could not reach yields that worked for fabless customers. Exynos 2500 on SF3 shipped only in some Galaxy S25 models, so Samsung lost its flagship internal validation, and Qualcomm, Apple, AMD and NVIDIA all chose TSMC N3E, a FinFET process with higher yield, for volume production. Samsung led on transistor architecture and trailed on manufacturing yield.

SF2 is improving faster. Yields rose from about 30% to 50% to 60% by November 2025, faster than SF3, and mass production began ahead of the original schedule. Exynos 2600 in Galaxy S26 phones for Europe and South Korea tests the process across millions of units, a volume SF3 never reached. At about $20,000 a wafer against about $30,000 for TSMC N2, SF2 appeals to AI startups, second-tier smartphone chip designers and customers with strategic reasons to diversify away from TSMC.

• NVIDIA Groq 3 LPU: the language processing unit from NVIDIA's $20 billion Groq acqui-hire, on Samsung 4nm. It is NVIDIA's first inference-focused chip since the deal and makes Samsung Foundry a qualified supplier in NVIDIA's production base, with room to grow as Samsung's processes mature.
• Tesla AI6: a $16.5 billion agreement through 2033 that gives Taylor a captive anchor customer and the volume certainty to justify more investment there. Tesla engineers work on the Taylor line, and Elon Musk committed to walking it personally. The AI6 program is covered on Tesla Terafab Supply Chain.

Samsung Foundry is a realistic TSMC alternative for some customers. North American companies that want to reduce Taiwan exposure can trade ecosystem depth for U.S. production at Taylor and lower prices on SF2P and later nodes. AI chip startups facing long TSMC lead times and lower allocation priority than Apple and NVIDIA benefit from the 33% discount and shorter lead times, as do sovereign AI programs. Apple, Qualcomm, AMD and NVIDIA have years invested in TSMC's PDK ecosystem and depend on its mature yields, and Samsung is not yet a volume alternative for them.

Foundry Capacity and HBM4 Base Die Allocation

Samsung has allocated more than half of its 30,000 WSPM Pyeongtaek foundry facility to internal HBM4 base die production. An HBM4 stack selling for more than $600 with a Samsung 4nm base die earns more per foundry wafer than an external customer wafer at $20,000, and HBM4 demand in 2026 is running ahead of supply while Samsung rebuilds share lost on HBM3e.

Pyeongtaek foundry capacity is effectively closed to external customers while HBM4 demand stays at current levels, which at current pricing means through at least 2027. Tesla AI6 runs at Taylor, and AI startups on Samsung 4nm or 2nm are served from Korean foundry lines not redirected to base die. Samsung puts its own most profitable memory products first, so it cannot match the capacity commitments TSMC offers external customers. Taylor, built for external customers with Tesla as anchor, lets Samsung grow foundry revenue without taking capacity from HBM4.

Pyeongtaek Campus

Samsung's Pyeongtaek campus in Gyeonggi Province, South Korea, is the world's largest semiconductor manufacturing complex by floor area. Four mega-fabs, P1 to P4, are running, and P5 is planned for 2028. The campus combines 1a, 1b and 1c DRAM, HBM3e and HBM4, V9 NAND, foundry lines now partly serving HBM4 base die, and advanced packaging. Like TSMC's Hsinchu complex, it concentrates equipment service, specialty chemical supply and engineering talent, which makes each fab more resilient and cheaper to run than a standalone site.

P4 was planned as an advanced foundry fab and has been redesignated as Samsung's main 1c DRAM fab, at 80,000 WSPM now and growing toward the company's 200,000 WSPM 1c target by the end of 2026. HBM4 requires 1c DRAM and is profitable enough to justify moving foundry capital into memory. The move limits external foundry growth at Pyeongtaek while Samsung Foundry needs outside customers to rebuild credibility after SF3. Samsung's memory fabs are listed on DRAM Fabs.

Samsung Supply Chain Bottlenecks and Risks

• SF2 yield gap with TSMC: the risk is to foundry competitiveness. SF2 yields 50% to 60% against more than 90% for mature TSMC N2, so Samsung's cost per good die stays higher despite the $20,000 against $30,000 wafer price, and large fabless companies that model yield risk and PDK maturity may not find the discount enough. Yield rose from 30% to 50% to 60% in about six months, and the 60% to 70% target for end of 2025 points to 80% or more by mid-2026; SF2P is designed with yield as a primary goal, Taylor production from H2 2026 adds yield data, and foundry profitability is expected about 2027.
• SK hynix HBM lead: the risk is to HBM market share. SK hynix keeps about 70% of Vera Rubin HBM4 despite Samsung's faster qualification, backed by more than three years of HBM production at M16 in Icheon and customer trust built during its HBM3e exclusivity, and one generation of base die advantage will not close the gap. HBM4E and custom HBM are where the 4nm base die matters most, NVIDIA's choices will decide whether Samsung grows toward 40% to 45% or holds near 30%, and the exclusive OpenAI Titan HBM4 win shows Samsung can win where custom base die performance counts.
• Foundry and memory capacity allocation: the risk is to Samsung Foundry capacity. More than half of Pyeongtaek foundry capacity goes to HBM4 base die, capping external foundry revenue for as long as base die margin beats external foundry margin. Taylor adds external-customer capacity outside the memory competition, P5 could go to external customers if HBM4E is served elsewhere, and a full fix would need dedicated external foundry capacity or outsourced base die production, which is unlikely.
• Taylor, Texas ramp: the risk is to Samsung Foundry credibility. Fab 1 targets end-2026 operations, possibly early 2027, after two delays from an original 2024 target; a further slip would delay Tesla AI6 and with it FSD, Cybercab and Optimus, and Taylor is Samsung's main proof of reliability for winning Qualcomm and other large customers. Pilot operations began February 2026 and EUV equipment testing in March 2026 on schedule, the co-CEO reported 2nm ahead of schedule in March 2026, and the CHIPS Act grant and Tesla engineers on site add pressure.
• Pyeongtaek concentration: a tail risk. Most of Samsung's DRAM, HBM, foundry and advanced packaging sits at one site in Gyeonggi Province, exposed to earthquake risk (lower than Taiwan), industrial accidents and proximity to North Korea, and a major disruption would hit global DRAM, HBM and Samsung Foundry at once. Hwaseong gives some diversification within Korea and Taylor gives foundry diversification abroad, and SK hynix's planned HBM packaging in Indiana reduces Korea concentration for the industry but not for Samsung.
• External foundry ecosystem depth: the risk is to customer acquisition. The SF2 PDK trails TSMC N2 in electronic design automation (EDA) tool characterization, IP library availability and design rule documentation, so designs take more risk and longer to reach production, and Qualcomm's reported reluctance to return reflects its experience of Samsung execution. Exynos 2600 volume in Galaxy S26 validates the ecosystem, each Samsung LSI product deepens SF2 PDK and IP support, and Design Solution Partners have SF2P guidelines; the two to three year estimate is for parity with N3E, and N2 parity is further out.

Related Coverage

NVIDIA Supply Chain Spotlight | AMD Supply Chain Spotlight | Intel Foundry Supply Chain Spotlight | Tesla Terafab Semiconductor Supply Chain | AI Data Center Spotlight | ASML Supply Chain Spotlight | China Semiconductor Bifurcation | HBM: High Bandwidth Memory Chips | Memory & Storage Chips Overview | Semiconductor Fab Clusters