[Part 2] Inside the Semiconductor Value Chain: HBM4 Wars, Foundry Realignment, and Korea's System Chip Renaissance

Executive Summary: While Part 1 established the macro earnings thesis anchored in DRAM's historic margin expansion, the more nuanced — and arguably more actionable — investment opportunity in 2026 lies deeper inside the semiconductor value chain. The HBM4 technology transition is reshaping supplier hierarchies in ways that diverge sharply from the HBM3E cycle; Samsung is reclaiming ground it ceded to SK Hynix, Micron is facing a validation crisis that threatens its recent market share gains, and the NVIDIA Vera Rubin supply architecture is consolidating around a Korean duopoly at 70/30. Meanwhile, Samsung Foundry's 2nm yield recovery — reported at 55–60% by late 2025 — is catalyzing a wave of design-win diversification that is beginning to monetize the overflow from TSMC's saturated advanced-node capacity. The downstream beneficiaries of this foundry renaissance, including Korea's nascent system semiconductor IP and design service companies, remain the most underappreciated leverage point in the entire ecosystem. China's GPU fabless build-out and DRAM indigenization effort introduce a decade-long structural competitive dynamic that investors must model — but do not yet fully price.

Analyst J's Strategic Takeaways — Part 2

• HBM4 Supply Architecture: The shift from HBM3E to HBM4 is restructuring the memory supply chain around two distinct camps — the TSMC-allied "Foundry-Memory Alliance" (SK Hynix, previously Micron) versus Samsung's vertically integrated "Turnkey" model. Intelligence from multiple industry sources as of February 2026 indicates NVIDIA has consolidated Rubin-generation HBM4 allocation to a Samsung/SK Hynix split approaching 30/70, effectively removing Micron from the primary Rubin supply chain in the near term. This is a sharper share shift than consensus anticipated entering the year.
• Contrarian View on Samsung Foundry: The market continues to discount Samsung Foundry as a chronic underperformer relative to TSMC. The contrarian position is that Samsung's 2nm GAA yield trajectory — accelerating from approximately 30% in Q1 2025 to an estimated 55–60% by late 2025 — combined with TSMC's $30,000+ per-wafer 2nm pricing, is creating real cost-driven pull for supply chain diversification. Samsung's Taylor, Texas facility adds a US-domiciled manufacturing option that is becoming strategically non-optional for defense and government-adjacent customers.
• Key Risk Factor: Korean system semiconductor companies (DSPs and silicon IP firms) carry high-single-digit to low-double-digit 2026F operating margins that are structurally dependent on Samsung Foundry's design-win momentum continuing. If Samsung Foundry's customer ramp lags the timeline implied by current domestic consensus estimates, the earnings inflection for this tier of the value chain could be delayed by 4–6 quarters.

The HBM4 Memory War: Who Actually Wins NVIDIA's Rubin Supply Chain

The transition from HBM3E to HBM4 is not an incremental product refresh. It is an architectural redesign of how memory and logic interact, and it is fundamentally reshuffling the competitive standings that defined the last two years of the AI memory market.

At the technical core, HBM4 introduces a logic base die manufactured on advanced foundry nodes — a departure from the commodity DRAM-process base dies that characterized all prior HBM generations. This design bifurcation has created two distinct strategic camps: the "Foundry-Memory Alliance," led by SK Hynix partnering with TSMC to manufacture its HBM4 base dies on advanced logic nodes, versus Samsung's vertically integrated "Integrated Titan" approach, using its own internal foundry to produce both the memory and logic base die. :antCitation[]{citations="06255fb9-31b5-433f-84bb-ea21f079f76b"}

SK Hynix maintained a dominant position with approximately a 62% share of HBM shipments as of Q2 2025, and Goldman Sachs assessed that the company would sustain a total HBM market share of over 50% at least through 2026. UBS further highlighted SK Hynix's position as the first HBM3E supplier for Google's latest TPUs and predicted the company would achieve approximately a 70% market share in the HBM4 market for NVIDIA's next-generation Rubin platform. :antCitation[]{citations="ab813a63-5fbb-4307-8031-607d59a153a0"}

The Micron variable is the most underappreciated development in this transition. Analysis from SemiAnalysis as of February 2026 indicated that NVIDIA's HBM4 supply for the Rubin architecture would consolidate to a SK Hynix/Samsung split approaching 70/30, effectively removing Micron from the primary Rubin supply chain in the near term, with Micron's internal HBM4 base die facing issues with customer validation, pin speeds, and foundry logic precision. :antCitation[]{citations="64eb3f81-0d4f-4575-9b1d-8dd681e50ebb"} This is a material departure from Micron's trajectory — the company had secured approximately 21% HBM market share in mid-2025 based on its HBM3E ramp, and its strategic positioning as a "greener" memory provider had won hyperscaler relationships specifically around power-constrained data center deployments. A near-term exclusion from NVIDIA's primary HBM4 supply chain creates a bifurcation: Micron's general-purpose DRAM business remains robust, but its premium HBM revenue visibility for H1 2026 is materially impaired.

For Samsung, the HBM4 cycle represents a critical redemption narrative. Samsung's HBM4 is built on its 1c (sixth-generation 10nm-class) DRAM process, which the company claims offers approximately 40% improvement in energy efficiency over current 1b-based modules — leveraging its internal foundry to produce both the memory and the logic base die as an integrated, cost-effective solution. :antCitation[]{citations="9fad0eae-f917-4ea3-8546-da80d62689e2"} The company has passed NVIDIA's required pin speed tests for HBM4, positioning it for meaningful volume allocation in the Rubin supply chain. The open question is yield scaling at the 16-Hi stack level, where all three suppliers are racing to reduce individual DRAM wafer thickness to approximately 30 micrometers — roughly one-third the thickness of a human hair — to fit 16 layers within the JEDEC-defined 775 micrometer height standard, a challenge industry observers describe as "formidable." :antCitation[]{citations="fd245f27-72a9-4161-82d3-bd55ceee3d49"}

The 16-Hi HBM4 race is the defining technical contest of the second half of 2026. NVIDIA reportedly requested delivery of 16-Hi HBM memory chips by Q4 2026, with HBM4 16-Hi stacks targeting that timeframe and the standard 12-Hi configurations already entering mass production in early 2026. :antCitation[]{citations="3c7ed5fd-8cb9-479a-8191-2d51486c76d5"} The supplier that achieves stable 16-Hi yields first commands extraordinary pricing power — these stacks will enable NVIDIA's "Rubin Ultra" variants carrying 500GB+ of HBM per package, a configuration with no cost-effective substitute.

Samsung Foundry's 2nm Inflection: The Underpriced Catalyst

While the memory narrative dominates investor attention, the more structurally durable value creation story in Korea's semiconductor ecosystem may be the gradual rehabilitation of Samsung Foundry. The market continues to discount this division — not without justification given its multi-year history of yield challenges and operating losses — but the evidence base for a trajectory change is accumulating in ways that the stock price has not yet reflected.

According to TrendForce data, TSMC dominated the global foundry market in Q2 2025 with a record 70.2% revenue share, while Samsung remained in second place with approximately 7.3% share. However, Samsung's 2nm yield rate climbed to an estimated 55–60% by late 2025, attracting new customers and with Samsung Foundry expected to approach profitability around 2027 once its Taylor, Texas plant begins operations in 2026. :antCitation[]{citations="d4bf994d-aacc-41a4-9a72-1ff8586dec1a"}

The strategic calculus for Samsung Foundry's recovery rests on three distinct pillars that operate in parallel rather than sequentially. First, the 8/5/4nm node tier — where Samsung's process maturity is well-established and yields are predictable — is absorbing genuine overflow from TSMC's congested advanced capacity as Chinese fabless companies and second-tier global hyperscaler ASIC teams seek alternatives. Second, the 2nm GAA node is attracting early design-win commitments from customers willing to accept a moderate yield and pricing tradeoff in exchange for US-domiciled manufacturing through Taylor, Texas and supply chain diversification from TSMC dependence. Samsung has already secured 2nm orders from Tesla and Ambarella, and TSMC's 2nm wafer pricing projected to exceed $30,000 per wafer is creating cost-driven pull toward Samsung's more competitive pricing, which is approximately 33% lower than TSMC's comparable node. :antCitation[]{citations="4b6efb7b-2f1e-47b2-b92c-77b16d3e3893"}

Third — and most underappreciated in global investor analysis — Samsung's HBM4 logic base die integration with its foundry creates a unique turnkey offering for customers seeking co-optimized memory and logic silicon. No other foundry can offer a customer simultaneous access to leading-edge logic manufacturing, HBM packaging, and advanced DRAM on a single supply chain. This integration capability is becoming a meaningful differentiator for hyperscaler custom silicon teams building AI accelerators that require tight memory-logic co-design.

Korea's System Semiconductor Ecosystem: The Third Derivative Play

For investors who have followed the memory and foundry narrative for years, the most differentiated opportunity in the current cycle may lie in the tier of Korean companies that most global institutional investors have not yet begun to model: the domestic system semiconductor design service and IP licensing ecosystem.

This segment is experiencing what domestic market analysis correctly characterizes as a "Scale-across" moment — the transition of its primary customer mix from consumer-oriented SoC design work toward high-performance computing, AI server inference silicon, and foundry-adjacent IP licensing. The financial implications of this transition are profound: HPC and AI silicon engagements carry design service fees that are 3–5x the per-project economics of consumer SoC tapeouts, and they generate recurring IP licensing revenue streams that consumer projects do not.

Company	Primary Business	2025A Revenue (KRW bn)	2026F Revenue (KRW bn)	2026F OPM	2027F Revenue (KRW bn)	Strategic Positioning
Semifive	Samsung Foundry DSP / SoC Design	121 (OPM -44%)	259	+6%	449	Primary beneficiary of Samsung Foundry design-win ramp; HPC/AI pivot driving revenue inflection
Gaonchips	Samsung Foundry DSP	69 (OPM -24%)	108	+8%	—	Leveraged to Samsung Foundry order recovery; near-term profitability hinge on 2nm ramp velocity
ADTechnology	Multi-foundry DSP / Turnkey SoC	165 (OPM +2%)	262	+7%	333	Only company in segment currently profitable; multi-foundry model reduces Samsung Foundry concentration risk
OpenEdge Technology	Silicon IP (memory controllers, interconnect)	16 (OPM -180%)	30	-7%	40	Highest-quality IP asset in segment; path to profitability hinges on IP licensing scaling with design-win volume

The earnings trajectories in this table carry a consistent pattern: heavy investment years in 2023–2025 creating operating losses, followed by a sharp revenue inflection in 2026–2027 as the design pipeline converts to shipped silicon. This is precisely the dynamic that made Korean EMS and display equipment companies exceptional investments in the preceding decade — the capital expenditure precedes the revenue, and the market chronically undervalues the revenue step-function when it arrives.

Semifive represents the most direct Samsung Foundry derivative play in the domestic system semiconductor universe. As a dedicated Samsung Foundry design service platform company, its revenue is structurally correlated with Samsung Foundry's design-win momentum across the 8/5/4nm and 2nm node spectrum. Domestic consensus estimates project Semifive's revenue more than doubling from KRW 121 billion in 2025 to KRW 259 billion in 2026, followed by a further 73% expansion to KRW 449 billion in 2027 as the foundry customer ramp matures. The 2026 operating margin is estimated to inflect to approximately +6% — the company's first sustained profitability — at an EV/EBITDA of approximately 213x, a valuation that prices in meaningful execution risk but also reflects the optionality value of being the primary design gateway for Samsung Foundry's most important customer tier.

ADTechnology is the segment's current profitable outlier, having achieved a 2% operating margin in 2025 on KRW 165 billion in revenue. Its multi-foundry model — servicing customers across both Samsung Foundry and TSMC process nodes — provides a structural hedge against Samsung Foundry-specific execution risk that neither Semifive nor Gaonchips carries. Domestic consensus projects ADTechnology's revenue reaching KRW 262 billion in 2026 at a 7% operating margin, expanding further to KRW 333 billion in 2027. At current market valuations, it trades at approximately 2.8x 2026F P/S — arguably the most defensible entry point in the segment given its demonstrated profitability and customer diversification.

OpenEdge Technology occupies the highest-quality strategic position in this tier but the longest path to financial maturity. As a silicon IP company specializing in memory controllers and interconnect IP — technology that is embedded into every SoC designed on Samsung Foundry's advanced nodes — OpenEdge has the highest structural leverage to the foundry renaissance of any company in the domestic ecosystem. Each new Samsung Foundry tapeout that incorporates OpenEdge's memory controller IP generates a recurring license fee that scales with the number of chips shipped. The model is extraordinarily capital-light once the IP library achieves critical mass; the challenge is that the company is currently operating at deeply negative margins (approximately -180% OPM in 2025 on KRW 16 billion in revenue) as it funds the development of its next-generation IP suite. The 2027 inflection to 11% OPM on KRW 40 billion in revenue, if achieved, would represent a business that has crossed the profitability threshold — at which point the royalty-stream valuation methodology typically generates significant upward multiple rerating.

China's Semiconductor Ambitions: The Decade-Long Structural Variable

Any rigorous analysis of the global semiconductor value chain in 2026 must grapple with China's parallel-track build-out across both GPU fabless and DRAM manufacturing — a strategic initiative that is simultaneously the market's most overhyped near-term threat and most underappreciated medium-term structural risk.

On the GPU fabless side, the domestic Chinese AI chip ecosystem has bifurcated sharply between credible commercial players and government-sponsored pretenders. Haiguang Information Technology (Hygon) and Cambricon are the two companies that have demonstrated genuine revenue traction and a path to profitability within the sector, while Moore Thread, MetaX, and Biren Technology remain in deep operating loss territory with no clear near-term profitability visibility. According to market data, Haiguang posted a 2024 operating profit of approximately USD 361 million on USD 1.273 billion in revenue — a 28% OPM that would be respectable in any semiconductor vertical. Its 2026 revenue is projected to approach USD 3.18 billion at a 31% OPM, driven by the captive Chinese domestic AI infrastructure buildout that is partially insulated from global competitive dynamics by the government's "domestic first" procurement mandate.

Cambricon's trajectory is more dramatic. From a USD 163 million revenue base in 2024 with a -42% OPM, the company is projected to reach USD 2.036 billion in 2026 revenue at a 37% OPM — a growth velocity that reflects the AI chip spending surge within Chinese hyperscalers and CSPs who are denied access to NVIDIA's highest-performance hardware. The critical external perspective here is that Haiguang and Cambricon are not competing with NVIDIA globally; they are providing a domestically acceptable substitute in a market where the alternative is either inferior hardware performance or export control exposure. This regulatory moat sustains elevated valuations — Haiguang currently trades at approximately 105x 2026F P/E — that would be indefensible in a fully competitive market but are partially rational given the captive demand structure.

On the DRAM side, CXMT's IPO ambitions — targeting the Shanghai STAR Market's Kechuang Board — represent the most consequential near-term catalyst for the Chinese memory ecosystem. According to industry filings, CXMT carried accumulated deficits of approximately CNY 40.86 billion (approximately KRW 7.7 trillion) as of mid-2025, driven by annual fixed asset depreciation exceeding CNY 10 billion and R&D spending that surpasses its current revenue base. The company projects breakeven in 2026 or 2027, contingent on ASP normalization and shipment growth. Its "skip-generation" development strategy — moving directly to DDR5 and LPDDR5/5X in parallel rather than iterating through DDR4 — is aggressive but not implausible given the technology transfer embedded in its early partnerships with Western equipment vendors before export controls tightened.

The geopolitical overhang on CXMT's trajectory is material but asymmetric. Equipment export restrictions prevent CXMT from accessing ASML's EUV systems, capping its ultimate density scaling at approximately the 1α/1β DRAM node — competitive with current commodity DDR5 but structurally incapable of producing HBM at competitive cost without technological breakthroughs. This means CXMT's competitive threat is concentrated in the commodity DRAM tier (DDR5, LPDDR5X for mobile) rather than the HBM segment where the cycle's current margin structure is anchored. For a 3–5 year investment horizon, that distinction matters significantly: memory incumbents' HBM margins are insulated, while commodity DRAM margins — currently elevated well above long-run averages — face a structural compression trajectory as CXMT scales.

Market Sizing & Full Value Chain Financial Overview

Segment / Company	Investment Tier	Primary Catalyst (12M)	2026F OPM	Key Risk	Conviction Level
SK Hynix	Core Overweight	HBM4 Rubin allocation (70% share); 16-Hi yield	~76%+ (DRAM div.)	Macro demand compression; 16-Hi yield ramp delay	HIGH
Samsung Electronics	Core Overweight	Q1 earnings confirmation; HBM4 Rubin volume ramp (30% share); Foundry 2nm	~60% (Semi div.)	HBM4 yield lag; Foundry profitability timeline	HIGH
ADTechnology	Tactical Long	Multi-foundry revenue ramp; AI/HPC SoC tapeout wins	~7%	Samsung Foundry execution; customer concentration	MEDIUM-HIGH
Semifive	Speculative Long	Samsung Foundry design-win pipeline conversion	~6%	100% Samsung Foundry dependency; execution risk on revenue doubling	MEDIUM
OpenEdge Technology	Long-dated Option	IP licensing scale; memory controller IP for HBM4 ecosystem	-7%	Extended path to profitability; small revenue base amplifies execution risk	MEDIUM
Haiguang IT (China)	Geopolitical Play	Captive Chinese AI infrastructure spending; domestic NVIDIA substitute	~31%	Export control escalation; regulatory/trade policy	MEDIUM (regulatory)
CXMT (China)	Watch / Monitor	STAR Market IPO; domestic DRAM market share growth	Still loss-making	Equipment access constraints; ASP volatility; governance structure	LOW-MEDIUM

Risk Assessment: The Tail Scenarios That Break the Bull Case

The most dangerous risk in the current semiconductor investment environment is not the one most frequently discussed — it is the interaction effect between multiple moderate-probability tail risks occurring simultaneously rather than in isolation.

The macro inflation shock scenario deserves a more rigorous probability weight than the market is currently assigning. Multiple domestic market analysts draw explicit parallels between the current environment — Middle East conflict driving crude oil prices, inflation reacceleration, and Federal Reserve uncertainty about H2 2026 rate trajectory — and the 2022 setup that destroyed memory valuations despite apparently robust fundamentals entering the year. The key asymmetry versus 2022 is that the demand mix is more institutionally anchored (hyperscaler capex versus consumer electronics), but the equity multiples entering this potential stress scenario are substantially higher than 2022 entry levels.

The Turbo Quant / algorithmic efficiency cascade risk is real but requires a nuanced reading. Google Research's KV-Cache compression demonstrating 6x memory reduction without accuracy loss is one data point in what is likely a sustained multi-year trend of algorithmic efficiency improvements in inference workloads. The aggregate effect on memory demand is ambiguous in direction but is clearly negative for the per-model memory intensity assumption. If the rate of efficiency gain exceeds the rate of deployment expansion over the next 12 months — a plausible if not probable scenario — HBM demand growth could undershoot consensus estimates by 15–20%. This would compress the premium attached to HBM pricing power without eliminating it.

The Samsung Foundry execution risk is the most company-specific downside scenario and is consequential primarily for the Korean system semiconductor ecosystem. TSMC currently commands approximately 86% of the advanced foundry market, underscoring the depth of the competitive gap that Samsung must close. :antCitation[]{citations="1a2b495f-e213-49da-813f-7eb933d38163"} Samsung's 2nm yield trajectory — improving from roughly 30% to 55–60% over 2025 — is encouraging but still meaningfully below TSMC's reported 90%+ yields at comparable nodes. Any reversion in yield progress, whether driven by process challenges or equipment access issues, would materially delay the design-win conversion that underpins Semifive, Gaonchips, and ADTechnology's 2026–2027 earnings inflection assumptions.

Strategic Outlook: The 12–24 Month Playbook Across the Full Stack

The semiconductor sector's investment thesis in 2026 is compelling precisely because it spans multiple layers of the value chain simultaneously — each with distinct return profiles, time horizons, and risk characteristics that complement rather than replicate each other.

At the memory layer, the core position is straightforward: the current earnings cycle is historically unprecedented in both magnitude and duration visibility, and the March sell-off created a re-entry opportunity that partial-cycle analysis would identify as attractive. According to the World Semiconductor Trade Statistics, the global semiconductor market is projected to grow more than 25% year-over-year in 2026, approaching approximately $975 billion, with the memory segment growing at approximately 30% and the total 2026 memory market potentially exceeding $440 billion. :antCitation[]{citations="4d902ba4-e4b6-4f2b-ba8f-dc37cf7e9dcd"} The critical April earnings confirmation event will either validate or challenge this framework within weeks.

At the foundry and ecosystem layer, the investment is a medium-term thesis with a 12–24 month realization horizon. Samsung Foundry's 2nm trajectory, Taylor facility ramp, and design-win momentum will determine whether Korea's system semiconductor DSP and IP companies deliver on their 2026–2027 earnings inflection promises. The companies that navigate this transition successfully — particularly those with multi-foundry agility and defensible IP libraries — will exit this cycle as structurally larger and more profitable businesses than they entered it.

At the China layer, the investment calculus is bifurcated by regulatory regime. Domestic Chinese AI chip companies with genuine revenue traction in the captive Chinese market represent a geopolitically complex but fundamentally sound earnings story — one that deserves monitoring by global investors even if direct accessibility varies by jurisdiction. CXMT's DRAM indigenization trajectory is a 3–5 year structural headwind for commodity DRAM pricing that should inform position sizing on that tier even as near-term HBM margin structures remain insulated.

The sector remains overweight. The confirmation event is imminent. Position accordingly.

---

Disclaimer: The information provided in this article is for informational and educational purposes only and does not constitute financial, investment, or trading advice. Investing in the stock market involves risk, including the loss of principal. All investment decisions are solely the responsibility of the individual investor. Please consult with a certified financial advisor and conduct your own due diligence before making any investment decisions.