Ultra-High-Performance, Low-Power AI Accelerator Chips

The Hardware Foundation of the AI Era

If artificial intelligence is the defining technology of the 21st century, then the semiconductors that power AI systems are the defining hardware. Mission 11 sits at the convergence of Korea's greatest existing industrial strength, semiconductor manufacturing, and its most ambitious future aspiration, becoming a top-3 global AI power. No other K-Moonshot mission leverages as strong a competitive foundation or addresses as critical a strategic imperative.

Korea's semiconductor industry is not merely large; it is structurally dominant in categories that are essential to AI computing. Samsung Electronics and SK Hynix together control approximately 70% of global DRAM production and roughly 50% of global NAND flash output. In the specific category of High Bandwidth Memory (HBM), the specialized memory chips that feed data to AI processors, Korean companies command an even more dominant position that has become the critical enabler of the generative AI revolution.

Mission 11 extends this memory dominance into new territory: the design and manufacture of AI accelerator processors, neural processing units (NPUs), and the full-stack silicon ecosystem required for sovereign AI computing. This is the most capital-intensive, technologically demanding, and geopolitically sensitive of all 12 K-Moonshot missions.

HBM4: The Memory Inflection Point

High Bandwidth Memory has become the single most supply-constrained component in the global AI infrastructure buildout. Every major AI accelerator, from NVIDIA's GPUs to Google's TPUs to the custom silicon developed by hyperscale cloud providers, requires HBM to feed data to processing cores at the speeds that AI workloads demand. Korea's two memory giants are locked in an intense but strategically aligned competition to dominate each successive HBM generation.

Samsung Electronics: HBM4 and the Trust-by-Design Pivot

Samsung commenced mass production of HBM4 in February 2026, marking a significant milestone in the company's campaign to reclaim HBM market share from SK Hynix. Samsung's HBM4 strategy centres on a "trust-by-design" approach, embedding reliability and quality assurance into the manufacturing process at a fundamental level rather than relying solely on post-production testing.

Samsung's HBM trajectory has been defined by the need to close a gap. SK Hynix won the initial HBM3E qualification battle for NVIDIA's H100 and H200 platforms, capturing the majority of early high-margin orders. Samsung's quality issues with early HBM3E production, which delayed NVIDIA qualification, became a defining corporate crisis that triggered significant organizational restructuring within Samsung's memory division.

The company has responded with a comprehensive recovery programme:

• HBM output surge: Samsung is increasing total HBM production output by approximately 50% through late 2026, reallocating DRAM fab capacity from conventional products to HBM
• Quality infrastructure: New testing and qualification protocols developed in collaboration with NVIDIA and other accelerator designers
• Advanced packaging: Investment in hybrid bonding and through-silicon via (TSV) technology to push HBM stacking density beyond current 12-layer configurations
• Foundry integration: Samsung's unique position as both a memory manufacturer and a logic foundry enables co-design of HBM and accelerator silicon on integrated packages

SK Hynix: Defending HBM Supremacy

SK Hynix enters 2026 from a position of commanding strength. The company projects approximately 70% market share in HBM4 for NVIDIA's next-generation Rubin platform, a dominance that reflects both technical leadership and deep customer trust built over multiple product generations.

Key SK Hynix developments for 2026:

• M15X fab opening: SK Hynix's new advanced memory fabrication facility, M15X, is scheduled to begin operations in May 2026, adding significant HBM production capacity in Cheongju
• All 2026 HBM production sold out: SK Hynix has disclosed that its entire 2026 HBM production allocation has been committed to customers, reflecting insatiable demand from AI infrastructure builders
• HBM3E dominance: HBM3E accounts for approximately two-thirds of total HBM shipments, with SK Hynix holding roughly 62% market share in the category
• CES 2026 technology showcase: At CES 2026, SK Hynix demonstrated a 16-layer HBM4 chip delivering 48GB capacity, alongside its conceptual "H3" memory architecture designed for AI workloads beyond current HBM specifications

The NPU Insurgency: Korea's AI Chip Startups

While Samsung and SK Hynix dominate AI memory, Korea's ambitions under Mission 11 extend to the processing side of the AI silicon stack. A new generation of AI chip startups is building Korea's domestic capability in neural processing units (NPUs) and AI inference accelerators, challenging the near-monopoly that NVIDIA currently holds in AI training and inference silicon.

Rebellions: From Startup to National Champion

Rebellions has emerged as the most prominent Korean AI chip startup, achieving a trajectory that positions it as a potential national champion in sovereign AI silicon. The company's momentum in 2025-2026 has been remarkable:

• National Growth Fund investment: ₩250 billion from Korea's National Growth Fund, the largest government-backed investment in a Korean AI chip company, signalling state-level strategic commitment
• Series C funding: $253 million raised in Series C, valuing the company at approximately ₩2 trillion (roughly $1.5 billion)
• Rebel chip mass production: Rebellions' Rebel AI inference accelerator enters mass production in 2026, targeting deployment in Korean data centres and sovereign AI computing infrastructure
• IPO trajectory: Samsung Securities has been appointed as IPO underwriter, with a public listing anticipated within 12-18 months, subject to market conditions

Rebellions' strategic positioning targets the inference side of AI computing, where power efficiency and cost-per-inference are more important than raw training throughput. This is a deliberate strategic choice: NVIDIA's dominance is most entrenched in AI training, while the inference market, which is growing faster as AI models move from development to production deployment, offers more opportunity for alternative architectures.

FuriosaAI: The NPU Specialist

FuriosaAI has carved a distinctive position in the Korean AI chip ecosystem through its focus on purpose-built NPUs for AI inference at the edge and in data centres. The company's trajectory reflects both technical ambition and strategic independence:

• 2026 deployment target: 20,000 NPU units targeted for delivery in 2026, a significant scale-up from pilot deployments
• Series C Bridge: $130 million raised in Series C Bridge funding, achieving unicorn valuation status (over $1 billion)
• Strategic independence: FuriosaAI notably rejected an acquisition approach from Meta Platforms, choosing to remain independent and pursue its vision of building a Korean-sovereign AI chip capability
• Architecture differentiation: FuriosaAI's Warboy and Renegade NPU architectures are designed for high throughput-per-watt AI inference, targeting workloads where NVIDIA GPUs are overprovisioned and energy-inefficient

DeepX and the Broader Startup Ecosystem

DeepX completes what Korean technology media has termed the "AI semiconductor trinity" alongside Rebellions and FuriosaAI. DeepX focuses on ultra-low-power AI chips for edge devices, including smartphones, IoT devices, and autonomous systems. The company's DX-M1 processor targets on-device AI inference with power consumption measured in milliwatts rather than watts.

Beyond these three leaders, Korea's AI chip startup ecosystem includes Sapeon (spun out of SK Telecom, focused on data centre inference), Nota AI (model compression and efficient inference), and several earlier-stage ventures targeting specialized AI silicon niches. The ecosystem's depth suggests that Korea's NPU ambitions are not dependent on any single company's success.

The ₩33 Trillion Government Semiconductor Package

Mission 11 operates within the context of the Korean government's broader ₩33 trillion semiconductor support package, one of the largest state-backed semiconductor investment programmes globally. This package, coordinated through MSIT and the Ministry of Trade, Industry and Energy, encompasses:

The package reflects Korea's recognition that semiconductor leadership requires state-level coordination of capital, infrastructure, talent, and regulatory frameworks. Individual companies, even conglomerates as large as Samsung, cannot sustain the investment trajectory required to remain competitive against state-backed competitors in the United States (CHIPS Act), the European Union (European Chips Act), Japan (semiconductor revival programme), and China (National Integrated Circuit Fund).

The Export Control Dimension

No analysis of Korea's AI chip mission is complete without addressing the US export control regime that has fundamentally reshaped the global semiconductor landscape. Korean companies operate at the intersection of US technology restrictions on China and China's demand for advanced memory and semiconductor products.

VEU Status Revocation and Its Aftermath

In August 2025, the US Department of Commerce revoked the Validated End-User (VEU) status that Samsung Electronics and SK Hynix had previously held for their China-based semiconductor operations. The VEU designation had allowed these companies to operate their Chinese fabs with a degree of regulatory streamlining. Its revocation meant that Samsung's Xi'an NAND flash facility and SK Hynix's Dalian DRAM and Wuxi packaging operations now require individual export licenses for equipment and technology shipments.

The practical impact has been significant:

• Operational uncertainty: Individual licensing introduces delays and unpredictability into maintenance, equipment upgrades, and capacity expansion at Chinese facilities
• Investment planning: Long-term capital expenditure planning for Chinese operations has become substantially more difficult, with some expansion projects effectively frozen
• Technology restrictions: Advanced process nodes (beyond certain thresholds) cannot be deployed at Chinese facilities, limiting the types of memory products that can be manufactured there
• Customer relationships: Chinese customers increasingly question the reliability of supply from Korean manufacturers operating under US licensing constraints

Both Samsung and SK Hynix have responded by accelerating domestic Korean and non-China international capacity expansion, a shift that aligns with Mission 11's objective of strengthening sovereign semiconductor manufacturing capability but imposes significant short-term costs and operational disruption.

The HBM Ecosystem: Beyond Memory Chips

HBM's importance extends beyond the memory chips themselves. The HBM value chain encompasses advanced packaging technologies, testing equipment, substrate materials, and integration expertise that collectively represent a deep technology moat:

• Through-Silicon Vias (TSVs): Vertical interconnections through silicon dies that enable the stacking of multiple DRAM layers. Korea's TSV manufacturing capability is among the most advanced globally.
• Hybrid bonding: Next-generation die-to-die bonding technology that increases interconnect density and reduces power consumption. Both Samsung and SK Hynix are investing heavily in hybrid bonding for future HBM generations.
• Advanced substrates: The organic and silicon interposer substrates that connect HBM stacks to logic processors. Korean substrate manufacturers are expanding capacity to serve the HBM ecosystem.
• Testing and qualification: AI-grade memory requires testing protocols far more stringent than conventional DRAM. Korea's semiconductor equipment ecosystem is developing AI-specific testing infrastructure.

This ecosystem depth means that Korea's HBM advantage is not a single-company or single-technology lead but a systemic capability embedded across dozens of companies and thousands of engineers. Replicating this ecosystem elsewhere would require a decade or more, even with unlimited capital.

Samsung Foundry: The Logic Dimension

Samsung's position as the world's second-largest logic foundry (behind TSMC) adds a unique dimension to Mission 11 that no other country's semiconductor strategy can replicate. Samsung Foundry's ability to manufacture both memory and logic chips enables co-optimization of AI accelerators with their memory subsystems, a capability that is increasingly important as AI architectures move toward tighter memory-compute integration.

Samsung Foundry's AI chip manufacturing capability faces significant challenges. The company's 3nm GAA (Gate-All-Around) process has encountered yield issues that have limited customer adoption relative to TSMC's competing N3 process. Samsung's aggressive pursuit of 2nm and beyond, with volume production targeted for 2027, represents a critical inflection point: success would position Samsung as a credible alternative to TSMC for leading-edge AI silicon, while continued yield challenges could relegate Samsung Foundry to a secondary role in AI chip manufacturing.

For Mission 11, Samsung Foundry's trajectory matters because sovereign Korean AI chip companies, including Rebellions and FuriosaAI, need a domestic foundry option for manufacturing their designs. Dependence on TSMC for fabrication of Korean-designed AI chips would represent an incomplete version of semiconductor sovereignty.

Competitive Landscape: The Global AI Chip Race

Korea's Mission 11 ambitions exist within a fiercely competitive global landscape:

Korea's competitive strategy under Mission 11 is not to challenge NVIDIA's GPU supremacy directly. Instead, it pursues a three-pronged approach: dominate the memory layer (HBM), build a viable domestic foundry alternative (Samsung), and develop specialized inference accelerators (Rebellions, FuriosaAI, Sapeon) for the sovereign AI computing market. This strategy leverages Korea's existing strengths while targeting market segments where NVIDIA's dominance is less entrenched.

Risk Assessment

Geopolitical Risk

The escalating US-China technology competition creates both opportunities and threats for Korean semiconductor companies. The opportunity lies in capturing market share from restricted Chinese competitors. The threat lies in being caught between US export controls and Chinese market access, a position that forces Korean companies to make binary strategic choices about geographic allocation of manufacturing capacity and technology development.

Technology Risk

Samsung Foundry's yield challenges at advanced nodes represent the most significant technology risk within Mission 11. If Samsung cannot achieve competitive yields at 2nm, Korea's sovereign AI chip manufacturing capability will remain constrained by dependence on TSMC.

Market Concentration Risk

Korea's semiconductor industry is concentrated in two conglomerates. Operational disruptions at either Samsung or SK Hynix, whether from natural disasters, geopolitical events, or corporate crises, would have outsized impact on global AI infrastructure supply and Korean national economic output.

Assessment and Outlook

Mission 11 operates from the strongest competitive foundation of any K-Moonshot mission. Korea's HBM dominance is not merely a market position but a deep ecosystem capability that competitors will struggle to replicate within the K-Moonshot timeframe. The 2026 HBM4 cycle, with Samsung and SK Hynix collectively supplying the vast majority of global AI memory requirements, represents a position of structural advantage that few national technology strategies can claim.

The mission's success will ultimately be measured not just by HBM market share retention but by whether Korea can translate memory dominance into broader AI silicon sovereignty. The Rebellions and FuriosaAI trajectories are encouraging, and the ₩33 trillion government support package provides substantial fiscal backing. However, the foundry challenge, Samsung's need to achieve competitive leading-edge yields, and the export control environment's evolution remain the critical variables that will determine whether Mission 11 achieves its full ambition or settles for a dominant but incomplete position in the global AI silicon stack.

What is beyond doubt is that Mission 11 is where Korea's K-Moonshot ambitions are most credible, most immediately impactful, and most globally consequential. The silicon that powers the AI revolution is, to a remarkable degree, Korean. Mission 11's objective is to ensure it stays that way, and to extend that advantage from memory into the full spectrum of AI computing hardware.