1) Introduction to 2nm Semiconductor Fabrication

The 2nm semiconductor fabrication process represents a significant milestone in the ongoing miniaturization of integrated circuits, following the 3nm node in the industry’s scaling roadmap. In semiconductor manufacturing, the “2nm” designation is primarily a marketing term rather than a direct measurement of any specific physical dimension, such as gate length or metal pitch. It refers to a generation of process technology that achieves higher transistor density, improved performance, and better energy efficiency compared to prior nodes. This advancement is driven by the need for more powerful and efficient chips to support applications in artificial intelligence, high-performance computing, data centers, mobile devices, and the Internet of Things (IoT). Key players like TSMC, Samsung, and Intel are at the forefront, with TSMC leading in production timelines as of March 2026.

As of early 2026, TSMC has already initiated volume production of its N2 (2nm-class) process in the fourth quarter of 2025, marking it as the most advanced logic technology available. This process introduces fundamental changes in transistor architecture and fabrication techniques to overcome the physical limitations encountered at larger nodes, such as leakage currents and scaling barriers.

Evolution and Key Innovations in Transistor Technology

Traditional semiconductor nodes from 7nm down to 3nm have relied on FinFET (Fin Field-Effect Transistor) structures, where the transistor channel is shaped like vertical “fins” to improve gate control over the current flow. However, as dimensions shrink below 3nm, FinFETs face challenges like increased leakage, reduced drive current, and manufacturing variability. To address these, the 2nm node transitions to Gate-All-Around (GAA) transistors, specifically using nanosheet or nanowire channels.

In a GAA nanosheet transistor, the channel consists of multiple thin, horizontal silicon sheets (typically 3-5 layers, each around 5-10nm thick) stacked vertically. The gate material completely surrounds each nanosheet on all four sides, providing superior electrostatic control compared to FinFET’s three-sided wrap. This design minimizes short-channel effects, suppresses leakage currents, and allows for better scaling while maintaining high performance. For instance, IBM’s prototypes for 2nm use a three-layer silicon nanosheet structure. TSMC’s N2 implementation is the first commercial deployment of this first-generation nanosheet technology.

Core Fabrication Process Steps

Fabricating 2nm chips involves hundreds of steps in a cleanroom environment, but the process can be broadly divided into front-end-of-line (FEOL) for transistor formation, middle-of-line (MOL) for contacts, and back-end-of-line (BEOL) for interconnects. Here’s a detailed overview:

1. Wafer Preparation and Epitaxy: Starting with a high-purity silicon wafer (typically 300mm diameter), epitaxial growth deposits thin layers of silicon or silicon-germanium (SiGe) to form the nanosheet stacks. Alternating layers of Si and SiGe are grown, with SiGe acting as a sacrificial material later removed to isolate the silicon nanosheets.
2. Lithography and Patterning: Extreme Ultraviolet (EUV) lithography is essential for 2nm, using light with a 13.5nm wavelength to pattern features below 20nm. TSMC employs curvilinear masks—lithography masks with curved shapes—to optimize resolution and reduce defects. Multi-patterning techniques may be used for critical layers to achieve the required precision. The process follows a 0.7x scaling factor per generation, enabling transistor densities exceeding 300 million per square millimeter.
3. Etching and Channel Formation: Reactive ion etching (RIE) defines the nanosheet stacks. Selective etching removes the SiGe layers, releasing the silicon nanosheets. This step requires atomic-level precision to avoid damage.
4. Gate Deposition: High-k dielectric materials (e.g., hafnium-based oxides) and metal gates are deposited using Atomic Layer Deposition (ALD), ensuring uniform coverage around each nanosheet. ALD is crucial for sub-2nm nodes, allowing monolayers of material (e.g., 1nm tungsten) to be deposited controllably.
5. Source/Drain Formation: Epitaxial growth forms raised source/drain regions, often with dopants like boron or phosphorus for p-type or n-type transistors.
6. Interconnects and BEOL: Low-resistance redistribution layers (RDL) and super high-performance metal-insulator-metal (MiM) capacitors enhance signal integrity. Copper or cobalt interconnects are used, with air gaps or low-k dielectrics to reduce capacitance. TSMC’s N2 does not include backside power delivery (BPDN) yet; that’s reserved for later variants like A16.
7. Testing and Packaging: Wafers undergo electrical testing, dicing, and advanced packaging like 3D stacking for multi-chip modules.

The entire process demands extreme precision, with tools like ASML’s EUV scanners costing hundreds of millions of dollars each.

Performance Improvements and Metrics

Compared to TSMC’s N3E (3nm) process:

• Performance: 10-15% gain at iso-power (same power consumption).
• Power Efficiency: 25-30% reduction at iso-performance (same speed).
• Density: 1.15x increase for mixed designs (logic + SRAM), achieving over 300M transistors/mm².

These gains stem from the GAA structure’s better control and innovations like N2 NanoFlex, which allows flexible cell designs for optimized density. Overall, 2nm enables chips that are faster, cooler, and more compact, ideal for AI accelerators and mobile SoCs.

Challenges in 2nm Fabrication

Scaling to 2nm introduces significant hurdles:

• Physical Limits: Quantum tunneling and leakage increase, necessitating GAA to maintain control.
• Manufacturing Complexity: EUV lithography is costly and yield-sensitive; variability in nanosheet thickness can affect performance.
• Cost and Yield: Initial yields may be low (though TSMC reports improvements), with high capital expenditure limiting access to major foundries.
• Materials and Integration: New deposition chemistries (e.g., ALD for heterostructures) and 3D integration are required for reliability.

Current Status and Industry Players

TSMC’s N2 is in volume production at fabs in Kaohsiung (Fab 22) and Hsinchu (Fab 20), with a combined capacity of ~90,000 wafers per month by late 2026. Early adopters include MediaTek, which taped out a flagship SoC on N2P (an enhanced variant) for production in late 2026. Samsung aims for 2nm in 2025, while Intel’s 20A (equivalent) uses similar GAA tech. Japan’s Rapidus targets 2nm by 2027 through partnerships with IBM and others.

Process Design Kits (PDKs) from imec and Synopsys facilitate design pathfinding, enabling academia and startups to prototype 2nm chips virtually.

The Major Players: 2026 Status

The race for 2nm dominance is currently a three-way battle between TSMC, Samsung, and Intel, with a new Japanese entrant, Rapidus, aiming for niche production.

Feature	TSMC (N2)	Samsung (SF2)	Intel (18A)
Status (March 2026)	Volume production ramping up.	2nd-gen (SF2P) ramping up.	In mass production (Panther Lake).
Transistor Type	Nanosheet GAA	MBCFET (Multi-Bridge Channel)	RibbonFET (GAA)
Power Delivery	Front-side (Backside in N2P/1.6nm)	Front-side (SF2)	PowerVia (Backside Power)
Yield (Estimated)	~65–75%	~40%	~55–65%
Key Customers	Apple, NVIDIA, AMD, Qualcomm.	Samsung LSI, Tesla, AMD.	Internal, Microsoft, AWS.

---

2) TSMC’s 2nm process technology

TSMC’s 2nm process technology, officially designated as N2, marks the company’s transition to the angstrom era and its first commercial implementation of Gate-All-Around (GAA) nanosheet transistors. This node represents a full generational advancement over TSMC’s prior 3nm family (primarily N3E), delivering meaningful improvements in performance, power efficiency, and transistor density to support the escalating demands of AI accelerators, high-performance computing (HPC), flagship mobile SoCs, and other compute-intensive applications.

As of March, 2026, TSMC’s N2 has successfully entered volume production (high-volume manufacturing or HVM), having commenced in the fourth quarter of 2025 (specifically announced as starting by December 31, 2025). Production is ramping aggressively across key facilities, including Fab 20 in Hsinchu (Baoshan Science Park) and Fab 22 in Kaohsiung, southern Taiwan. Early yield indicators have been strong, with reports citing logic test chip yields in the 70-80% range and certain SRAM blocks exceeding 90% average yield during the ramp phase. This positions N2 as the industry’s most advanced foundry technology in terms of combined density and energy efficiency at this point in 2026.

The rapid production scale-up reflects extraordinary demand driven by the AI boom, with capacity for 2026 largely sold out or fully booked. Major customers have secured allocations, including Apple (reportedly taking over half of initial capacity for its A20-series chips in upcoming iPhones and other devices), NVIDIA, AMD (for next-gen EPYC Venice CPUs and MI450 AI accelerators), Qualcomm, MediaTek, and others—totaling at least 15 confirmed customers, predominantly in HPC and AI spaces. TSMC is expanding output aggressively, with projections for monthly wafer starts reaching tens of thousands (potentially approaching or matching 3nm levels) by mid-to-late 2026 across multiple fabs, fueled by both smartphone and AI/HPC applications.

Transistor Architecture and Key Innovations

N2 introduces TSMC’s first-generation nanosheet GAA transistors, moving away from the FinFET structure used in all prior nodes from 7nm through 3nm. In this design:

• The channel consists of multiple thin, horizontal silicon nanosheets (typically stacked in 3–5 layers, each ~5–10 nm thick).
• The gate material fully surrounds each nanosheet on all four sides, providing superior electrostatic control compared to FinFET’s three-sided wrap.
• This architecture significantly reduces short-channel effects, suppresses leakage currents (especially at low voltages), and enables better drive current scaling.

The result is improved performance at lower voltages (e.g., 0.5–0.6 V), with notable gains in performance-per-watt and reduced standby power (up to ~75% lower in some low-voltage scenarios). N2 also incorporates optimizations such as:

• Advanced high-k metal gate (HKMG) refinements.
• Low-resistance redistribution layers (RDL).
• Super high-performance metal-insulator-metal (MiM) capacitors with over 2x capacitance density versus prior generations, plus 50% reductions in sheet and via resistance for better power stability and efficiency.
• Enhanced EUV lithography techniques, including multi-patterning and curvilinear masks where needed, to achieve the required precision at these dimensions.

Importantly, the base N2 does not include backside power delivery network (BSPDN). TSMC deferred this feature to later variants to prioritize rapid validation of the nanosheet architecture and manage complexity/cost during the initial ramp.

Performance, Power, and Area (PPA) Improvements

Compared to TSMC’s N3E (enhanced 3nm) process, N2 delivers full-node benefits:

• Performance: 10–15% gain at iso-power (same power consumption).
• Power Efficiency: 25–30% reduction at iso-performance (same speed/clock).
• Transistor Density:
  • ~15% increase for mixed designs (logic + analog + SRAM).
  • Up to ~20% higher for logic-only layouts.
• Overall, N2 achieves leading-edge transistor densities (in the 300+ MTr/mm² range for high-density configurations), with exceptional SRAM density reported at around 38 Mb/mm² in some characterizations—providing a notable advantage in cache-heavy designs like mobile SoCs and AI accelerators.

These gains stem from the GAA structure’s better channel control, reduced parasitic capacitance, and process optimizations that allow tighter scaling without excessive leakage penalties. N2 excels particularly in energy-efficient computing scenarios, making it ideal for battery-constrained devices and power-hungry AI workloads.

Production Status, Capacity, and Roadmap Derivatives (March 2026)

• Current Ramp: Volume production is accelerating in 2026, with strong early momentum. Initial output started modestly (thousands of wafers/month) but is scaling rapidly, supported by high yields and fab expansions. Combined capacity from key 2nm fabs is projected to grow substantially through 2026–2027.
• Customer Adoption: Demand has created a supply crunch, with 2026 capacity effectively sold out in many reports. Apple dominates early slots for flagship consumer products (e.g., expected in iPhone 18 series launching fall 2026), while HPC/AI players drive massive volumes.
• Derivatives:
  • N2P (performance-enhanced variant): Builds on N2 with further refinements, including NanoFlex flexible cell design for optimized PPA. Offers incremental gains (e.g., ~5–10% better performance/power in some metrics) using the same design rules. Volume production is scheduled for the second half of 2026.
  • A16 (angstrom-class with Super Power Rail): Integrates backside power delivery (SPR) on an N2P-like base, providing additional 8–10% speed at iso-voltage or 15–20% power reduction. Targets complex AI/HPC designs; volume production also slated for second half of 2026, with products expected in 2027.

Challenges and Strategic Position

Scaling to 2nm involves extreme precision in EUV patterning, nanosheet stacking uniformity, and defect control, but TSMC’s mature GAA experience (building on extensive R&D) has enabled a smoother ramp than some competitors. While density gains are described as more iterative than revolutionary in some analyses (prioritizing power/performance leadership), N2’s energy efficiency and ecosystem maturity position TSMC strongly against rivals like Samsung’s SF2 (modest PPA uplifts, lower yields) and Intel’s 18A (strong BSPDN but smaller ecosystem and later volume scale).

In March 2026, TSMC’s N2 stands as the dominant 2nm-class foundry technology, powering the industry’s shift to more efficient, denser compute. Its rapid ramp and sold-out status underscore the insatiable demand for advanced nodes amid the AI era, with further enhancements via N2P and A16 set to extend leadership through 2027 and beyond. This node enables breakthroughs in flagship mobile processors, next-gen data center CPUs/GPUs, and AI accelerators, sustaining Moore’s Law-like progress through architectural innovation rather than pure dimensional scaling.

---

3) TSMC’s N2P process

TSMC’s N2P process is an enhanced, performance-optimized derivative within the company’s 2nm-class family, building directly on the foundational N2 node. It represents a targeted refinement rather than a full generational leap, focusing on incremental improvements in performance, power efficiency, and design flexibility while maintaining full compatibility with N2 design rules. This approach allows seamless migration for designs already taped out on N2, minimizing redesign effort and accelerating time-to-market for customers.

As of March , 2026, N2P remains in the pre-volume production phase. TSMC has confirmed that volume production (high-volume manufacturing or HVM) for N2P is scheduled for the second half of 2026 (likely starting in the latter part of the year, with meaningful output ramping toward late 2026 or early 2027 for initial products). Early risk production, yield learning, and customer tape-outs are progressing, supported by the successful ramp of base N2 (which entered volume production in Q4 2025). TSMC’s overall 2nm family is experiencing exceptionally strong demand—driven by AI accelerators, flagship mobile SoCs, and HPC—leading to capacity constraints and sold-out allocations through much of 2026 for the broader N2 ecosystem.

MediaTek has already announced successful tape-out of its flagship SoC (potentially the Dimensity 9600 or equivalent) on N2P, with volume production targeted for late 2026. Other major players, including Qualcomm (rumored for Snapdragon 8 Elite Gen 6 variants), are expected to adopt N2P for competitive positioning against Apple’s likely use of base N2 for certain products (e.g., A20-series or M6 chips) to balance cost and performance. This positions N2P as a key node for mobile and mid-to-high-end compute designs seeking marginal but meaningful edges in frequency, efficiency, or thermal headroom.

Transistor Architecture and Core Features

N2P retains the same first-generation nanosheet Gate-All-Around (GAA) transistor structure introduced in N2:

• Multiple horizontal silicon nanosheets (typically 3–5 stacked layers, each ~5–10 nm thick) fully surrounded by the gate on all four sides.
• This provides excellent electrostatic control, reduced short-channel effects, and lower leakage compared to FinFETs used in prior nodes.
• High-k metal gate (HKMG) materials, advanced source/drain epitaxy, and low-resistance interconnect optimizations carry over unchanged.

Key enhancements in N2P include:

• NanoFlex technology — A flexible cell library approach that allows designers to mix-and-match transistor variants (e.g., high-performance, low-power, or high-density cells) within the same design block. This enables fine-tuned optimization of performance-per-watt-area (PPA) trade-offs without changing the underlying process rules.
• Process refinements for slightly better device characteristics, such as improved drive current, reduced parasitic capacitance, or tighter variability control.
• Continued use of advanced EUV lithography (including multi-patterning and curvilinear masks where applicable) and other manufacturing optimizations inherited from N2.

Importantly, N2P does not include backside power delivery network (BSPDN or Super Power Rail). TSMC deferred this major architectural feature to the follow-on A16 node (1.6nm-class, volume production also slated for second half of 2026) to prioritize rapid maturation of the nanosheet platform, manage complexity, control costs, and ensure high yields during the initial 2nm ramp. This decision reflects lessons from the industry’s aggressive scaling challenges, allowing N2 and N2P to focus on validating GAA reliability and ecosystem readiness.

Performance, Power, and Area (PPA) Improvements

TSMC positions N2P as delivering incremental but valuable gains over base N2, while offering stronger uplifts when benchmarked against the prior N3E (enhanced 3nm) node:

• Vs. N2 (base):
  • Performance: ~5% improvement (at iso-power).
  • Power efficiency: ~5–10% reduction (at iso-performance).
  • Density: Minimal to no change (same design rules; any gains come from NanoFlex-enabled optimizations rather than process scaling).
  • These modest uplifts primarily stem from device tuning, better leakage management, and flexible library usage, making N2P attractive for designs pushing clock speeds or battery life without requiring full re-layout.
• Vs. N3E (3nm):
  • Performance: Up to ~18% gain at iso-power.
  • Power efficiency: ~36% reduction at iso-performance.
  • Logic density: ~1.2x increase.
  • Chip density (mixed logic/SRAM/analog): ~1.15x.
  • These figures represent cumulative benefits from N2 → N2P evolution, highlighting why N2P is seen as a compelling upgrade path for customers moving from 3nm-era designs.

N2P inherits N2’s strengths in energy-efficient computing (particularly at low voltages) and excellent SRAM density (~38 Mb/mm² range), making it well-suited for cache-heavy AI accelerators, mobile processors, and edge devices.

Production Status, Capacity, and Customer Adoption (March 2026)

• Timeline: Volume production targeted for H2 2026. Early customer engagement (tape-outs) is active, with yields expected to target high levels (potentially 80%+ goals cited in some reports) building on N2’s strong ramp (70%+ logic yields reported early in 2026).
• Capacity: N2 family fabs (e.g., Fab 20 in Hsinchu, Fab 22 in Kaohsiung) will support both N2 and N2P starting late 2026. Overall 2nm output is scaling aggressively to meet demand, though the node family remains capacity-constrained through 2026.
• Adoption:
  • MediaTek (flagship SoC tape-out complete, production late 2026).
  • Qualcomm (rumored adoption for competitive Snapdragon variants).
  • Potential others in mobile/HPC segments seeking slight edges over base N2 users (e.g., Apple reportedly favoring N2 for cost reasons on some lines).
  • HPC/AI players may lean toward A16 for backside power advantages.

Strategic Role and Challenges

N2P serves as a bridge node in TSMC’s 2nm family strategy: it extends the lifespan and adoption of the 2nm platform by offering quick, low-risk improvements before the more disruptive A16 introduction (with Super Power Rail BSPDN for ~8–10% speed or 15–20% power gains over N2P). This “continuous enhancement” model—similar to past N3 → N3E → N3P iterations—helps maximize returns on the massive capital invested in GAA infrastructure.

Challenges mirror those of base N2: extreme patterning precision, nanosheet uniformity, defect control, and high wafer costs. However, deferring BSPDN to A16 reduces early complexity, contributing to smoother yield ramps.

In March 2026, N2P positions TSMC to capture additional share in the fiercely competitive mobile and mid-range HPC markets during late 2026–2027, complementing the high-volume base N2 ramp (dominated by Apple, NVIDIA, AMD) and setting the stage for A16’s backside power revolution in AI/HPC workloads. This layered 2nm-family approach underscores TSMC’s leadership in sustaining performance scaling amid physical limits.

---

4) TSMC’s A16 node technology

TSMC’s A16 node (officially branded as A16™ technology) represents the company’s first angstrom-class (sub-2nm) logic process, marking a pivotal step in the industry’s shift toward extreme scaling for AI, high-performance computing (HPC), and other power-hungry workloads. Announced in 2023 and positioned as the successor to the N2 family, A16 integrates advanced nanosheet Gate-All-Around (GAA) transistors with TSMC’s proprietary Super Power Rail (SPR) backside power delivery network (BSPDN). This combination addresses key bottlenecks in power distribution and signal routing that become critical at these dimensions, enabling superior performance-per-watt, reduced IR drop (voltage loss), and higher overall efficiency.

As of March , 2026, TSMC’s A16 is in the risk production phase, with early wafer starts, yield learning, and customer prototyping underway. Volume production (high-volume manufacturing or HVM) remains on track for the second half of 2026 (likely starting late in the year, with meaningful ramp toward the end of 2026 and into 2027 for initial commercial products). This timeline aligns with TSMC’s consistent roadmap updates from late 2024 through early 2026, including confirmations at events like the Open Innovation Platform (OIP) forums. Fab preparations are progressing rapidly in Taiwan (e.g., expansions at Fab 20 in Hsinchu and Fab 22 in Kaohsiung), with future capacity extensions planned for U.S. facilities like Fab 21 in Arizona (targeting A16-capable production around 2028–2030).

Demand for A16 is exceptionally strong, driven by the insatiable needs of AI accelerators and HPC designs. Early reports indicate NVIDIA as the lead (and potentially initial sole) customer, planning to leverage A16 for next-generation GPU architectures such as Feynman (expected around 2028, following Rubin/Rubin Ultra on prior nodes). Other potential adopters include AMD, OpenAI-linked designs, and select HPC players, though consumer/mobile segments (e.g., Apple) may skip directly to follow-on nodes like A14 (1.4nm-class) due to cost and timeline considerations. A16’s high wafer costs and complexity make it particularly suited for premium, high-margin AI/HPC chips rather than broad smartphone volumes.

Transistor Architecture and Key Innovations

A16 builds on the first-generation nanosheet GAA foundation from N2/N2P but incorporates second-generation refinements for better drive current, variability control, and leakage suppression:

• Nanosheet GAA Transistors — Multiple thin, horizontal silicon nanosheets (stacked layers, typically 3–5) fully surrounded by the gate on all four sides. This provides exceptional electrostatic control, minimizing short-channel effects and enabling aggressive scaling.
• Super Power Rail (SPR) — TSMC’s advanced BSPDN implementation, which routes power lines to the backside of the wafer. Unlike simpler via-based approaches (e.g., Intel’s PowerVia), SPR uses direct backside contacts to the transistor’s source and drain. This virtually eliminates frontside power routing congestion, reduces parasitic resistance, and allows denser signal interconnects on the frontside.
  • Benefits include significantly lower IR drop, improved power stability under high current loads, and enhanced frequency headroom—critical for AI GPUs and accelerators with dense power grids.
• Additional optimizations: Refined high-k metal gate (HKMG) stacks, low-resistance interconnects, advanced EUV lithography (including High-NA preparations where applicable), and process controls for yield and uniformity.

The base A16 does not introduce complementary FETs (CFETs) or other post-GAA architectures; those are reserved for future nodes (e.g., A14 or beyond). Instead, A16 emphasizes mature GAA + sophisticated BSPDN integration for faster risk-to-volume transition.

Performance, Power, and Area (PPA) Improvements

TSMC positions A16 as delivering tangible, full-node-like gains over its immediate predecessor, N2P (the performance-enhanced 2nm variant without BSPDN):

• Performance: 8–10% improvement at the same supply voltage (iso-Vdd).
• Power Efficiency: 15–20% reduction at iso-performance (same speed/clock).
• Density: 7–10% increase in logic density (some reports cite up to 1.1x overall chip density improvement, depending on design mix).
• Cumulative vs. earlier nodes (e.g., N3E 3nm): Significantly larger uplifts when factoring in GAA + SPR synergies.

These metrics stem primarily from SPR’s power delivery advantages (reducing resistance and enabling tighter routing) combined with transistor tweaks. A16 excels in scenarios with complex signal wiring and high power demands, offering better thermal/power headroom than N2P while maintaining compatibility with N2-family design rules (facilitating IP reuse).

Production Status, Capacity, and Roadmap Derivatives (March 2026)

• Current Phase: Risk production active; early yields and process learning progressing toward HVM targets. No public yield figures released yet, but TSMC emphasizes faster maturation due to lessons from N2 ramp.
• Volume Ramp: H2 2026 start, with products expected in market from 2027 onward (e.g., NVIDIA Feynman GPUs potentially debuting late 2027/early 2028).
• Capacity Buildout: Massive investments (tens of billions) in Taiwan fabs for N2/A16 family; U.S. expansions (Arizona) will add A16 capability later.
• Derivatives: A16 serves as the BSPDN entry point; follow-ons like A14 (1.4nm-class, full-node scaling) are planned for further density and efficiency gains, potentially in 2028+.

Challenges and Strategic Position

A16 introduces manufacturing complexities: backside wafer thinning, precise alignment for SPR contacts, and defect control in nanosheet stacking. TSMC’s deferral of BSPDN from N2/N2P to A16 allowed focused maturation of GAA first, contributing to smoother ramps. Cost remains high, limiting broad adoption to AI/HPC initially.

In March 2026, A16 positions TSMC strongly in the angstrom era, countering Intel’s 18A (which entered volume production earlier with PowerVia BSPDN but trails in some density metrics and ecosystem scale). TSMC’s approach—more advanced SPR integration, higher projected density in certain cells, and vast customer base—gives it an edge for AI dominance, though Intel claims competitive performance in normalized metrics. A16 sustains TSMC’s leadership by enabling the extreme efficiency needed for next-gen AI accelerators, bridging to even denser future nodes while capitalizing on explosive HPC/AI demand. This node underscores the shift from pure dimensional scaling to architectural innovation for continued progress.

---

5) Samsung’s 2nm process technology

Samsung’s 2nm process technology, branded as SF2 (first-generation 2nm), along with its enhanced derivative SF2P (second-generation 2nm), marks Samsung Foundry’s continued push into the angstrom era using Gate-All-Around (GAA) transistor architecture. Samsung was the first company to commercially deploy GAA transistors at the 3nm node (via its MBCFET implementation), and SF2 extends this foundation to achieve higher transistor density, better power efficiency, and performance scaling for demanding applications like mobile processors, automotive semiconductors, high-performance computing (HPC), and AI/edge devices.

As of March, 2026, Samsung Foundry has achieved mass production of its first-generation SF2 process, which began in the fourth quarter of 2025 (late 2025). This milestone is validated by the high-volume production of the Exynos 2600 mobile SoC, Samsung’s flagship application processor for select models in the Galaxy S26 series (primarily non-Ultra variants in certain markets, including South Korea). Yields on SF2 have improved significantly, reaching the 50% range (with some reports citing 50-60% during ramp-up), sufficient for commercial shipments despite being lower than leading competitors. No major critical defects have been widely reported in Exynos 2600 production, indicating stabilization.

Samsung is actively promoting SF2 and transitioning focus to SF2P, with guidelines sent to design partners and customers to prioritize the enhanced variant. SF2P volume production is targeted for late 2026 (second half of the year), building on SF2’s foundation with further refinements. The company expects strong order growth—over 30% year-on-year in 2nm-related business—driven by internal use (e.g., future Exynos chips like the rumored Exynos 2700 on SF2P), the confirmed Tesla AI chip deal (potentially AI5/AI6 variants), and ongoing discussions with major U.S. and Chinese customers (including Qualcomm rumors for Snapdragon variants). Samsung is also ramping capacity at facilities like the Taylor, Texas fab (nearing full operational status around mid-2026) to support localization and broader adoption.

Transistor Architecture and Key Innovations

Samsung’s SF2 employs MBCFET (Multi-Bridge Channel FET), a nanosheet-style GAA transistor:

• Multiple horizontal silicon “bridges” or channels (typically 3–4 layers per device) are stacked vertically.
• The gate fully surrounds each channel on all four sides, delivering superior electrostatic control, reduced short-channel effects, lower leakage currents, and improved drive current compared to FinFETs.
• This enables aggressive scaling while maintaining reliability at sub-2nm dimensions.

Key process features include:

• Advanced EUV lithography (with multi-patterning and pellicle solutions for defect reduction).
• Optimized high-k metal gate (HKMG) stacks.
• Refined source/drain epitaxy and low-resistance interconnects.
• Compatibility with prior SF3 (3nm GAA) designs, allowing IP reuse and easier migration.

SF2P introduces incremental optimizations (e.g., better variability control, drive current tuning, and process refinements) without changing core design rules, enabling flexible adoption.

Notably, base SF2 and SF2P do not include backside power delivery network (BSPDN). Samsung plans to introduce this in the follow-on SF2Z variant targeted for 2027, aiming to reduce IR drop, enhance power stability, and improve density in high-current scenarios like AI accelerators.

Performance, Power, and Area (PPA) Improvements

Samsung positions SF2 as an evolutionary step from its second-generation 3nm (SF3 Gen2):

• Performance: Approximately 5% gain at iso-power.
• Power Efficiency: Around 8% improvement.
• Area/Density: About 5% reduction (or higher in optimized layouts).

These are relatively modest compared to some competitors’ generational jumps, reflecting Samsung’s strategy of building incrementally on its early GAA experience. Independent analyses often rank SF2 lower in overall PPA than peers, particularly in SRAM density and high-performance scenarios.

SF2P targets more substantial uplifts over base SF2:

• Up to 12% better performance.
• 25% lower power consumption.
• 8% smaller area.

These enhancements make SF2P more competitive for HPC and mobile flagship designs, with Exynos 2700 (potentially for Galaxy S27 series) as a key early candidate.

Transistor density remains competitive (around 200–231 MTr/mm² in logic-heavy configurations), though generally trailing TSMC’s N2 in SRAM and mixed-cell metrics.

Production Status, Capacity, and Customer Adoption (March 2026)

• SF2: In mass production since late 2025; yields stabilized at ~50%+; flagship product is Exynos 2600 (powering millions of Galaxy S26 units in select regions).
• SF2P: Pre-volume phase; mass production ramp targeted for late 2026; early tape-outs and promotions active.
• Capacity: Scaling across Korea and U.S. (Taylor fab tests/operations advancing); focus on automotive (2027 extension) and HPC.
• Adoption:
  • Internal: Exynos 2600 (current), Exynos 2700 (upcoming on SF2P).
  • External: Tesla (confirmed for AI chips), Qualcomm (rumored testing/collaboration), potential U.S./Chinese clients.
  • Growth target: 30%+ YoY in 2nm orders, leveraging TSMC capacity constraints.

Challenges and Strategic Position

Samsung faces ongoing hurdles typical of aggressive scaling:

• Yield Maturity: SF2 yields (50% range) lag behind TSMC N2 (higher 60–80% in early reports) and Intel 18A, impacting cost and volume scalability.
• Customer Ecosystem: Limited third-party traction compared to TSMC’s broad dominance (Apple, NVIDIA, AMD); pricing remains competitive to attract wins.
• BSPDN Delay: Absence in SF2/SF2P puts it behind Intel’s PowerVia-enabled 18A for power-hungry designs.
• Competition: TSMC leads in density, maturity, and ecosystem; Intel edges in raw performance metrics per some benchmarks.

Despite challenges, SF2’s commercial success with Exynos 2600 demonstrates progress, and SF2P’s refinements position Samsung for a potential rebound in 2026–2027. Strategic bets on Tesla, Qualcomm discussions, U.S. localization, and aggressive order targets aim to capitalize on AI-driven demand and any TSMC supply tightness. Samsung’s layered roadmap (SF2 → SF2P → SF2Z with BSPDN) emphasizes steady evolution over revolutionary leaps, sustaining its role as a viable alternative in the fiercely competitive 2nm-class foundry landscape as of March 2026.

---

6) Intel’s 18A process node

Intel’s 18A process node (18 angstroms, equivalent to a ~1.8nm-class technology) is Intel Foundry’s leading-edge logic manufacturing platform as of March, 2026. It represents the successful culmination of Intel’s “five nodes in four years” (5N4Y) roadmap announced in 2021, delivering the company’s first high-volume implementation of two groundbreaking architectural innovations: RibbonFET gate-all-around (GAA) transistors and PowerVia backside power delivery network (BSPDN). These features position 18A as a direct competitor to TSMC’s N2 family and Samsung’s SF2/SF2P in the sub-2nm/2nm-class generation.

As of March 2026, 18A is in high-volume manufacturing (HVM), having entered volume production in late 2025 (with meaningful ramp through Q4 2025 and into 2026). This marks a significant milestone, as Intel has fulfilled its internal production commitments and begun shipping commercial products. The node powers Intel’s Core Ultra Series 3 (codenamed Panther Lake) client processors, which launched at CES 2026 and became broadly available starting in January 2026. Panther Lake serves as the flagship showcase for 18A, delivering impressive efficiency and integrated graphics performance in laptops, AI PCs, and emerging handheld/edge devices. Additional products, such as data center variants (e.g., Xeon 6+ configurations), are also ramping, with reports of strong early demand exceeding initial supply for Panther Lake.

Yields have shown steady improvement: estimates from mid-to-late 2025 placed them around 55%, with monthly gains of ~7% leading to projections of 65-75% by early 2026. Intel reports defect density below 0.40 D0 (a healthy level for early HVM), and the company emphasizes predictable ramp progress. Full industry-standard yield maturity (comfortable commercial profitability, typically 80%+) is targeted for 2027, with cost-appropriate levels expected by end-2026. Fab 52 in Chandler, Arizona (a key U.S. site), is fully operational for 18A HVM, supporting domestic production resilience.

Intel has shifted strategy on 18A for external foundry customers. Initially considered more internal-focused (with 14A prioritized for broader foundry promotion), recent progress—including Panther Lake success—has led to renewed external interest, particularly in the 18A-P (performance-enhanced) variant. Inbound inquiries are noted, though ecosystem scale still trails TSMC’s dominance.

Transistor Architecture and Key Innovations

18A introduces RibbonFET, Intel’s branded GAA transistor:

• Horizontal silicon ribbons (multiple stacked layers, fully gated on all four sides) replace FinFETs.
• This provides superior electrostatic control, reduced leakage, better short-channel effect suppression, and improved drive current scaling.

Paired with PowerVia BSPDN:

• Power rails route to the wafer backside, separate from frontside signal interconnects.
• Reduces IR drop (voltage loss), minimizes frontside routing congestion, improves power stability under high loads, and enables 5-10% better cell utilization/density.
• Offers up to ~4% iso-power performance uplift from reduced resistance alone.

Other enhancements include:

• Advanced high-k metal gate (HKMG) stacks.
• Optimized interconnects and EUV lithography for precision.
• Design-rule compatibility with prior nodes for IP reuse.

18A does not yet incorporate complementary FETs (CFETs) or other post-GAA structures; those are planned for future nodes like 14A.

Performance, Power, and Area (PPA) Improvements

Intel quotes 18A vs. its prior Intel 3 node:

• Performance: Up to 15-25% gain at iso-power (same power) or iso-voltage (e.g., ~25% higher frequency at 1.1V).
• Power Efficiency: Up to 36% reduction at iso-performance (same speed).
• Density: Over 30% scaling (chip-level), with high-density (HD) standard-cell transistor density around 238 MTr/mm² and SRAM cells at ~0.021 µm² (HDC) or 0.023 µm² (HCC).

Independent analyses (e.g., TechInsights) rank 18A highest in raw performance among 2nm-class nodes (ahead of TSMC N2 and Samsung SF2), crediting PowerVia’s advantages in high-frequency/HPC scenarios. However, TSMC’s N2 leads in transistor density (e.g., ~313 MTr/mm² HD cells and superior SRAM density at ~0.0175 µm²), enabling more compact cache-heavy designs.

Vs. competitors (approximate 2026 views):

• Vs. TSMC N2: Intel edges in performance (e.g., higher clocks/efficiency in normalized benchmarks), but trails in density (TSMC ~1.3x denser in some HD cells). N2 lacks BSPDN in base form (deferred to N2P/A16).
• Vs. Samsung SF2/SF2P: 18A leads in performance; Samsung trails in yields (~40-50%) and PPA consistency.
• Vs. TSMC A16: A16 (H2 2026 HVM) adds Super Power Rail BSPDN on N2P base, potentially closing Intel’s structural gap with competitive or superior density/efficiency.

18A excels in power-hungry, frequency-driven applications (AI PCs, HPC), while density-focused designs may favor TSMC.

Production Status, Capacity, and Roadmap (March 2026)

• HVM Ramp: Active since late 2025; Panther Lake in market (strong demand, supply constraints noted).
• Capacity: Primarily U.S.-based (Arizona fabs like Fab 52); supports internal products with growing foundry potential.
• Derivatives: 18A-P (performance-tuned) sees inbound interest; 18A-PT variants planned. Follow-on 14A (1.4nm-class) in development (risk production ~2028), with capacity tied to customer commitments.
• Adoption: Internal dominance (Panther Lake, upcoming Xeon); external traction building (inbound for 18A-P); no major third-party wins announced yet (ecosystem maturing).

Challenges and Strategic Position

Challenges include:

• Ongoing variability and yield maturation (not yet at full commercial comfort).
• High initial costs and slower ramp vs. TSMC’s scale.
• Ecosystem catch-up (TSMC’s broad customer base remains dominant).

Strategically, 18A validates Intel’s IDM 2.0 vision: reclaiming process leadership in key metrics (performance, BSPDN first-mover), enabling U.S.-based resilient manufacturing, and positioning the foundry for growth amid AI demand. Panther Lake’s success (efficiency, graphics, battery life) bolsters credibility, with 18A as the inflection point for Intel’s comeback. While trailing TSMC in density and volume scale, 18A’s architectural edges and early HVM status make it a strong contender in the 2nm-class race as of March 2026, sustaining momentum toward 14A and beyond.

---