TSMC A14 vs Intel 14A: The 1.4nm Foundry Race, Updated for November 2026
The two companies racing to 1.4nm have made opposite bets on almost every hard question. TSMC says it will not use High-NA EUV lithography on any node through 2029, and it ships its first 1.4nm-class node with power still delivered from the front of the wafer. Intel puts High-NA scanners and a second-generation backside power network into its 14A node from day one, and has said 14A may not be built at all without outside customers. The TSMC A14 vs Intel 14A contest is no longer a spec-sheet comparison. It is a test of two theories of how to scale.
It matters now because the decision window is open. Intel shipped its 14A 0.9 process design kit in October 2026 and expects customers to commit between late 2026 and mid-2027. TSMC says A14 risk production is on track for 2027.
You will leave with a verified, dated picture of both nodes, a worked cost model, and a decision framework for chip teams.
What this covers: process technology, backside power, High-NA timing, schedules, customers, capacity, Samsung and Rapidus, per-die economics, risks, and recommendations.
Context and Background
Leading-edge logic manufacturing has narrowed to three serious players: TSMC, Intel Foundry and Samsung Foundry, with Japan’s state-backed Rapidus trying to become a fourth. TSMC dominates by every commercial measure. In its third quarter of 2026, TSMC reported revenue of US$46.72 billion and a 66.4% gross margin, with 7nm-and-below nodes making up 79% of wafer revenue, according to its Q3 2026 management report filed with the SEC. High-performance computing, which is mostly AI accelerators and server CPUs, was 68% of revenue that quarter, per post-earnings coverage.
The current production frontier is the 2nm class. TSMC’s N2 entered volume production in the fourth quarter of 2025 and reached 8% of wafer revenue by Q3 2026. Intel’s 18A, its first node with gate-all-around (GAA) transistors and backside power, entered high-volume manufacturing in October 2025 at Fab 52 in Arizona. We covered that generation in depth in our analysis of the TSMC 2nm AI chip competition and in our look at Intel 18A-P and its position against TSMC.
The 1.4nm generation is the next step. “1.4nm” and “14A” (14 angstroms) are marketing names, not physical gate lengths. No feature on either process measures 1.4 nm. The names signal a full node of improvement over the 2nm class, and both companies define that improvement differently.
Three things changed between the original version of this article and November 2026. TSMC published a roadmap through 2029 at its April 2026 Technology Symposium, which pushed A16 volume production into 2027 and added A13 and A12. Intel confirmed High-NA EUV for 14A, delivered its 0.9 PDK, and saw Elon Musk publicly state that Tesla’s Terafab project plans to use 14A. And Samsung pushed its 1.4nm node to 2029, which removes it from the 2028 head-to-head.
My thesis is simple. The 1.4nm race will not be decided by transistor density. It will be decided by whether Intel converts evaluations into signed capacity before TSMC’s A14 ramp locks in the 2028 to 2030 design cycle. The technology gap is now small enough that commercial execution dominates.
TSMC A14 vs Intel 14A: Two Different Technical Bets
Direct answer: TSMC A14 is a 2028 node using second-generation nanosheet transistors, NanoFlex Pro cells and front-side power, made without High-NA EUV. Intel 14A is a 2028 node using RibbonFET 2 transistors, PowerDirect backside power and High-NA EUV. TSMC optimizes for cost and yield predictability; Intel optimizes for early adoption of new tools.

Figure 1: Process lineage for both foundries. TSMC splits its roadmap into a front-side power line for client chips and a backside power line for data center chips. Intel puts backside power and High-NA EUV into a single 14A node.
The figure shows the structural difference. TSMC runs two parallel tracks from N2. One track, A14 and its A13 shrink, keeps power on the front side and targets phones and PCs. The other track, A16 and later A12, adds Super Power Rail backside power for data center chips. Intel runs a single track. 18A introduced backside power, and 14A deepens it, while adding High-NA lithography on critical layers.
Transistors: nanosheets everywhere, with different tuning
Both nodes use GAA transistors, where the gate wraps around stacked horizontal silicon sheets instead of the fins used since 22nm. Wrapping the gate on all sides improves electrostatic control, which reduces leakage at short gate lengths. TSMC calls its version nanosheet; Intel calls its version RibbonFET.
TSMC describes A14 as its second-generation nanosheet transistor. It pairs that with NanoFlex Pro, a design-technology co-optimization feature that lets designers mix cells of different heights within a block. Short cells save area and power; tall cells drive more current on critical paths. At its 2025 symposium, TSMC claimed A14 delivers up to 15% higher speed at the same power, or 25% to 30% lower power at the same speed, versus N2, as reported by Tom’s Hardware. The 2026 roadmap table lists about 1.23x logic density and 1.2x mixed chip density versus N2.
Intel’s 14A uses RibbonFET 2, which stacks four nanosheets. Intel claims 15% to 20% better performance per watt versus 18A, which it says translates to 25% to 35% lower power at the same performance, and 1.3x transistor density. It also introduces Turbo Cells, a library approach that mixes fast, power-hungry cells into critical CPU and GPU paths, and a wider range of threshold voltages, according to Tom’s Hardware’s coverage of Intel’s 14A disclosure. Turbo Cells and NanoFlex Pro solve the same problem: give designers per-path control over the speed-versus-area trade-off.
The headline percentages are not comparable across vendors. TSMC measures against N2; Intel measures against 18A. Those baselines differ. Independent analyst teardowns published in 2025 estimated 18A’s high-density logic density meaningfully below N2’s. If those estimates hold, a 1.3x gain on 18A lands roughly level with N2, not with A14. Treat that as a directional conclusion, because neither company publishes comparable cell-level density for these nodes.
Backside power: TSMC splits it out, Intel goes all in
Backside power delivery moves the power supply network to the underside of the wafer, under the transistors. Signal wires keep the front-side metal stack. This reduces voltage drop, frees routing tracks and can improve density. The cost is process complexity: the wafer must be bonded to a carrier, thinned to a few hundred nanometers of silicon and then patterned from the back.
Intel adopted backside power first, with PowerVia on 18A. On 14A it moves to PowerDirect, which contacts the transistor source and drain directly from the back rather than through an intermediate via. Intel calls PowerDirect the main driver of 14A’s gains. Direct contact shortens the current path, which reduces resistance further, but it demands tighter alignment between front and back patterning.
TSMC’s Super Power Rail, introduced with A16, is also a direct-contact backside scheme. TSMC describes A16 as N2P plus Super Power Rail, and claims 8% to 10% higher performance or 15% to 20% lower power versus N2P, with 1.07x to 1.10x density. Crucially, TSMC chose not to put Super Power Rail into the first version of A14. Backside power arrives on the A14 platform only with A12, in 2029.
The reason is economics, not capability. Backside power adds wafer steps and cost, and it complicates heat removal because the thinned silicon now sits between two dense metal stacks. A phone SoC running at a few watts gains little from it. A 1,000-watt AI accelerator with severe voltage drop gains a lot. TSMC is pricing that distinction into its roadmap; Intel is betting that one node can serve both markets.
Lithography: the High-NA split
Extreme ultraviolet (EUV) lithography prints the finest features using 13.5 nm light. Standard EUV scanners use optics with a numerical aperture (NA) of 0.33. High-NA scanners, ASML’s EXE series, use 0.55 NA, which improves resolution and can replace two or three exposures of a standard scanner with one.
Intel confirmed in January 2026 that High-NA EUV “will be part of our 14A process,” according to its CFO as reported by Bits&Chips. Intel completed acceptance testing of ASML’s second-generation EXE:5200B in December 2025. TrendForce reported that the tool reaches 175 wafers per hour with 0.7 nm overlay accuracy. Intel’s October 2026 earnings release also announced a High-NA readiness collaboration with ASML.
TSMC has gone the other way. At its April 2026 symposium, TSMC said it plans no High-NA use for any node through 2029, including A13 and A12. A TSMC executive said the R&D team keeps finding ways to scale without it, adding that “one day they may have to use it,” per Tom’s Hardware’s report on the 2026 roadmap. On the Q3 2026 earnings call, CEO C.C. Wei said TSMC expects to adopt High-NA around 2030, when the tool is mature.
The trade-off is concrete. Multi-patterning with standard EUV adds exposures, masks, etch steps and overlay error budget, but uses fully depreciated, well-understood tools. High-NA reduces steps on the hardest layers but uses scanners reported at more than US$350 million each, with new masks, resists and a halved exposure field. Intel accepts tool risk to cut process steps. TSMC accepts process steps to avoid tool risk.
Side-by-side summary
| Dimension | TSMC A14 | Intel 14A | Confidence |
|---|---|---|---|
| Transistor | 2nd-gen GAA nanosheet | RibbonFET 2, four sheets | Vendor-disclosed |
| Design-level feature | NanoFlex Pro | Turbo Cells, wider Vt range | Vendor-disclosed |
| Power delivery | Front-side; backside arrives with A12 in 2029 | PowerDirect backside | Vendor-disclosed |
| High-NA EUV | No, through 2029 | Yes, critical layers | Vendor statements |
| Perf or power vs prior node | Up to 15% speed or 25-30% lower power vs N2 | 15-20% perf/W or 25-35% lower power vs 18A | Vendor claim |
| Density vs prior node | About 1.23x logic vs N2 | About 1.3x vs 18A | Vendor claim |
| Risk production | 2027 | 2027 per Q3 2026 call | Vendor guidance |
| Volume production | 2028 | Production start 2028 | Vendor guidance |
| Named external customers | Not disclosed; broad interest claimed | Tesla intent stated by Musk; others unnamed | Mixed |
The percentages are “up to” figures on tuned libraries. Real designs land lower, and the gap between vendor claims and product results is usually several percentage points.
Deeper Analysis: Schedules, Customers, Capacity and Cost
Technology parity means little without wafers, customers and a cost that buyers will pay. This section tests each node on those three axes and then places Samsung and Rapidus in the field.

Figure 2: Verified and guided milestones from 2025 to 2030. Solid arrows show TSMC and Intel milestones from company statements. Dotted links show Samsung and Rapidus 2nm programs that compete for the same customers.
The timeline shows how tightly the schedules now overlap. Both A14 and 14A target risk production in 2027 and production in 2028. That leaves little room for either company to claim a generational lead on timing alone. The differentiation shifts to yield maturity, capacity and price.
Schedule credibility
TSMC’s schedule record at the leading edge is strong, with one recent exception. A16 was originally guided for volume in the second half of 2026; the 2026 roadmap shows volume in 2027. TSMC framed this as customer-driven: the process is ready, but customers chose their ramp timing. That explanation is plausible, since several large AI customers moved designs to N2P and N2X variants instead.
For A14, the signals are positive. On its Q3 2026 call, TSMC said “risk production is well on track for 2027 and volume production will start in 2028.” C.C. Wei added that he expects A14 to be “a bigger node than our current N2 technologies.” TrendForce, citing Taiwanese press in July 2026, reported that TSMC told investors A14 is ahead of where N2 was at the same stage, with a 256Mb SRAM test vehicle near 90% yield. That SRAM figure is reported, not company-published, so treat it as directional.
Construction supports the timeline. TrendForce reported that the Central Taiwan Science Park administration said the 1.4nm fab is ahead of schedule, with the first building expected before April 2027. Supply chain sources cited in that report suggest pilot production from Q3 2027 and mass production by mid-2028. Those are third-party estimates.
Intel’s schedule has moved more. In April 2025 Intel guided 14A risk production to 2027. Reporting in mid-2026 suggested slippage toward 2028 risk and 2029 volume. On its Q3 2026 call, however, CFO David Zinsner said Intel will “prepare Intel 14A for risk production next year” and is investing “to position 14A to begin production in ’28.” Intel also named Coral Rapids as a 14A server product for 2028. The official line is therefore back to 2027 risk and 2028 production start. “Production start” is softer than “high-volume manufacturing,” and that distinction is worth tracking.
The customer gate
Intel has made 14A explicitly conditional. It told investors in 2025 that it may pause or discontinue 14A and later nodes if it cannot land a significant external customer. Major capacity spending waits for those commitments. That is a rational capital policy, but it creates a chicken-and-egg problem: customers want proof of capacity before they commit, and Intel wants commitments before it builds capacity.

Figure 3: How a 14A design commitment flows. Each PDK release unlocks deeper customer evaluation, and Intel ties volume High-NA tool orders to signed commitments rather than building ahead of demand.
The sequence explains why the PDK version numbers matter so much. A 0.5 PDK lets a customer run early power and area studies. A 0.9 PDK is close enough to final that engineering teams can tape out meaningful test chips. A 1.0 PDK is the production contract between process and design. Intel released 14A PDK 0.9 in October 2026 and CEO Lip-Bu Tan said he expects 1.0 “by Q1 next year,” per the Intel Q3 2026 results and the earnings call.
On customers, the public record is thin but improving. Tan said he is “highly encouraged” by progress with “multiple potential customers on 14A” without naming them. The most concrete public statement came from outside Intel. On Tesla’s Q1 2026 earnings call in April, Elon Musk said, “We plan to use Intel’s 14A process,” for chips made through the Terafab project, as reported by PC Gamer. Intel had announced it was joining Terafab with SpaceX, xAI and Tesla earlier that month. No contract terms, volumes or dates have been published.
Other names remain in the reported category. The New York Times reported in June 2026 that Apple is considering 14A for some M-series and lower-end A-series chips. Analyst notes have suggested Apple wins on 18A-P and 14A. Several outlets reported that AMD, Nvidia, Microsoft, Google and others are evaluating 14A. Evaluation is not a design win. A company evaluating a second source gains negotiating leverage with TSMC whether or not it commits.
TSMC’s book and capacity
TSMC does not need to announce A14 customers. Apple has been TSMC’s lead customer at every node since N7, and AI accelerator vendors are migrating to N2-family and A16 nodes. TSMC says N2-family capacity will more than double in 2027, and it holds a 2026 capital budget of US$60 billion to US$64 billion. The company also said 2026 to 2028 capex will be “even more significantly higher” than the prior three years.
That spending sets the competitive bar. Intel guided 2026 gross capex to about US$20 billion, with 2027 “up strongly.” Intel Foundry revenue was US$6.7 billion in Q3 2026, but external foundry revenue was only US$411 million, much of it from a nonrecurring program. Intel Foundry still posted an operating loss of about US$1.06 billion that quarter, per its press release. The financial gap matters because capacity is what turns a competitive node into a competitive product. Our analysis of hyperscaler AI capex economics shows where that demand comes from.
Geography is Intel’s strongest argument. Intel says Fab 52 in Arizona can exceed 10,000 18A wafer starts per week at full ramp. 14A development runs in Oregon, and Arizona’s Fab 62 is unassigned and targeted for readiness around 2028. Intel’s Ohio site has slipped to 2030 to 2032 per the February 2025 schedule, and Intel cancelled its Magdeburg and Wroclaw projects in July 2025. TSMC, meanwhile, said its second Arizona fab will begin volume production in Q3 2027, slightly ahead of schedule, but its leading nodes stay in Taiwan first.
The wider field: Samsung and Rapidus
Samsung Foundry has stepped back from the 1.4nm race. According to Tom’s Hardware’s August 2026 Samsung roadmap review, Samsung moved SF1.4 to 2029 and will keep it on standard EUV, reserving High-NA for a 1nm-class node around 2030. Samsung’s near-term focus is 2nm: SF2 powers the Exynos 2600, SF2P targets the Exynos 2700, and SF2Z adds backside power in 2027. Reported first-generation 2nm yields were near 55% in April 2026, which is an estimate, not a Samsung figure. Its biggest anchor is a US$16.5 billion Tesla contract for AI6 at Taylor, Texas, which reportedly slipped about six months.
Rapidus is earlier still. Its 2nm pilot line in Hokkaido began operating in April 2025, a 2nm GAA prototype met target electrical characteristics in July 2026, and volume production is targeted for 2027 to 2028. Rapidus plans 1.4nm process development from 2026 and mass production around 2029. As of mid-2026, Tom’s Hardware reported that no customer had signed a volume agreement, although more than 60 companies were in talks.
Neither Samsung nor Rapidus changes the 2028 contest. Both matter as price pressure and as second-source options for designs that do not need the absolute leading edge.
Decision matrix for chip buyers
| Use case | TSMC A14 | TSMC A16 then A12 | Intel 14A | Samsung SF2 family |
|---|---|---|---|---|
| Flagship phone SoC, 2028 | Strong fit | Over-specified | Possible if Apple-type deal lands | Price option |
| AI accelerator compute die | Weak; no backside power | Strong fit | Strong on paper; High-NA field limits | Limited ecosystem |
| Server CPU tile | Good | Good | Good; Intel’s own Coral Rapids | Limited |
| US-made leading edge, 2028 | Arizona lags Taiwan | Arizona lags Taiwan | Strongest claim | Taylor, 2nm class |
| Supply chain risk hedge | Low diversification | Low diversification | High diversification | Medium |
The matrix is my assessment based on the public facts above, not a vendor ranking. It shows that the two nodes are less direct competitors than the “A14 vs 14A” framing suggests. A14 competes for phones and PCs. 14A competes with both A14 and A16 at once.
Worked example: what yield and field size do to die cost
The following numbers are illustrative, chosen to show mechanisms, not to estimate either foundry’s real prices or yields.
Take a 600 mm² AI compute die on a 300 mm wafer. The standard gross-die approximation is wafer area divided by die area, minus an edge-loss term: 70,686 / 600 − (π × 300) / √(2 × 600). That gives about 118 − 27, or roughly 90 candidate dies per wafer.
Now apply a simple Poisson yield model, Y = e^(−D0 × A), where D0 is defect density per cm² and A is die area in cm². At D0 = 0.10, yield is e^(−0.6), about 55%, which gives about 49 good dies. At D0 = 0.20, yield falls to about 30%, or 27 good dies. At an assumed wafer price of US$30,000, cost per good die moves from about US$610 to about US$1,110.
The lesson is that a doubling of defect density costs more than any plausible difference in wafer price between the two foundries. A new node with a 15% cheaper wafer but twice the defect density is a worse deal for large dies. That is why mature-yield predictability is TSMC’s real product, and why Intel’s improving 18A yield trajectory matters more to 14A buyers than any density figure.
Field size adds a second effect. A standard EUV exposure field is 26 mm × 33 mm, or 858 mm². A High-NA field is half that: 26 mm × 16.5 mm, or 429 mm². Our 600 mm² die does not fit in one High-NA exposure. If the die includes any High-NA layer, it must either be stitched across two fields or split into chiplets. Stitching adds design rules and yield risk at the seam. Splitting adds die-to-die interconnect area and packaging cost, a trade-off we work through in our guide to UCIe chiplets and HBM4 packaging.
Most large AI accelerators are already multi-die designs, so the half-field limit is manageable. It still forces earlier floorplanning decisions on 14A than on A14 or A16, where full-field exposures remain available.
Packaging is part of the node now
No leading AI chip ships as a single die on a bare package. TSMC’s CoWoS remains the default; Wei said TSMC is in production at 5x reticle size and working toward 9x and 14x. Intel offers EMIB and Foveros, and on the Q3 call Wei acknowledged that “our competitor offers a very competitive solution.” Intel can also package dies made elsewhere, including at TSMC. That gives Intel a foothold with AI customers even if they never use 14A wafers, and it is one way an evaluation turns into a relationship.
Trade-offs, Gotchas, and What Goes Wrong
The most common analytical mistake is comparing vendor percentages across different baselines. A 1.3x density gain on 18A and a 1.23x gain on N2 are not comparable numbers, because 18A and N2 start from different densities. Always normalize to a shared reference, or treat the comparison as qualitative.
The second mistake is treating customer evaluations as wins. Large chip buyers routinely evaluate a second foundry to improve their TSMC pricing and to satisfy board-level supply chain risk reviews. Until a customer names a product, a node and a volume, it is a lead, not revenue. The Tesla statement is the strongest public signal, and even that has no published contract terms.

Figure 4: A practical node selection tree. Product type decides the first branch; backside power need and supply chain requirements decide the rest. Each terminal node carries a different engineering risk.
The tree highlights specific failure modes.
High-NA immaturity. Every new lithography generation has surprised its first users with mask, resist or pellicle problems. High-NA adds anamorphic optics, which magnify differently in two directions, and therefore new mask-making rules. If a critical High-NA layer underperforms, Intel can fall back to standard EUV multi-patterning, but that changes design rules late and costs schedule.
Backside power thermals and debug. Backside power puts metal under the transistors, so heat must travel through a thinned silicon layer and a bonded carrier. Hot spots can get worse, not better. Failure analysis also gets harder, because classic backside optical probing through the silicon is blocked by the power network. Teams moving to A16, A12 or 14A need new test and debug flows.
SRAM stagnation. SRAM bitcells have barely shrunk since 5nm across the industry. AI accelerators devote large die areas to SRAM, so the headline logic density gains overstate real die shrink for those designs. Expect effective area gains on AI dies well below the vendor logic figures.
Capacity mismatch. A great node without capacity is not a product. If Intel waits for commitments before scaling 14A tool orders, and customers wait for capacity before committing, the 2028 window can pass. The Section 48D US investment tax credit requires construction to begin by December 31, 2026, which adds a hard deadline to Intel’s capacity decisions.
Concentration risk. On the other side, choosing TSMC for everything leaves a design exposed to a single region. That is a board-level risk, not an engineering one, and it is the main reason Intel’s 14A pitch resonates with US hyperscalers and governments.
Practical Recommendations
For most phone and client SoC teams, A14 is the default 2028 choice. It reuses N2-era design flows, avoids backside power complexity, and its A13 shrink offers an IP-compatible path into 2029. Teams that cannot wait should look at N2P and N2U.
For AI accelerator and HPC teams, the real TSMC comparison is A16 and A12, not A14. If your design is voltage-drop limited, backside power is worth its cost now. Plan multi-die floorplans that would also survive a High-NA half-field, so you keep the option of moving a die to 14A later.
For buyers with US-manufacturing requirements, defense programs or a mandate to dual-source, 14A deserves a serious evaluation in the PDK 0.9 to 1.0 window. Ask for signed capacity, not just a PDK. Ask how Intel will handle a High-NA slip. Consider a staged path: package with EMIB or Foveros first, then move a non-critical die to 18A-P, then commit a lead die to 14A.
For everyone, model cost per good die, not cost per wafer. A fab digital twin approach, like the one in our semiconductor fab digital twin reference architecture, makes yield assumptions explicit and testable.
A short checklist before any 1.4nm commitment:
- Normalize vendor PPA claims to one baseline and one cell library.
- Model yield across a range of defect densities, not a single point.
- Check whether any layer uses High-NA, and what that implies for die size.
- Confirm packaging capacity alongside wafer capacity.
- Get capacity commitments in writing, with volumes and dates.
- Keep an IP-portable fallback node in the plan.
Frequently Asked Questions
Which is better, TSMC A14 or Intel 14A?
Neither is clearly better on public data, because each vendor measures gains against its own prior node. TSMC A14 offers a lower-risk path with front-side power and no High-NA EUV, and it suits phones and client chips. Intel 14A offers backside power and High-NA from day one, which may suit high-power chips, plus US manufacturing. TSMC has the stronger yield and capacity record. Intel has the more aggressive technology stack. For most buyers, the deciding factors are capacity, yield maturity and supply chain policy rather than density.
When will TSMC A14 and Intel 14A enter production?
TSMC says A14 risk production is on track for 2027, with volume production starting in 2028. Intel’s CFO said on the Q3 2026 call that Intel will prepare 14A for risk production in 2027 and position it to begin production in 2028. Earlier 2026 reporting suggested Intel’s schedule could slip toward 2029 volume, so the 2028 target depends on customer commitments and High-NA readiness. TSMC’s fab construction for its 1.4nm site is reportedly ahead of schedule.
Does TSMC A14 use High-NA EUV or backside power?
No to both, for the base A14 node. TSMC said at its April 2026 symposium that it plans no High-NA EUV use on any node through 2029, and its CEO pointed to around 2030 for adoption. A14 uses front-side power delivery. TSMC’s backside power technology, Super Power Rail, debuts on A16, with volume now expected in 2027, and returns on the A14 platform as A12 in 2029. This split lets TSMC charge for backside power only where data center customers need it.
Who are Intel 14A’s customers?
Intel has not officially named any external 14A customer as of its October 2026 earnings. CEO Lip-Bu Tan said Intel is encouraged by multiple potential customers. Elon Musk said on Tesla’s April 2026 earnings call that Tesla plans to use Intel 14A for Terafab chips, but no contract terms were published. Media reports say Apple, AMD, Nvidia, Microsoft and others are evaluating 14A. Intel expects customer decisions between the second half of 2026 and the first half of 2027.
What is the difference between TSMC Super Power Rail and Intel PowerDirect?
Both are backside power delivery networks that move power wiring beneath the transistors and connect directly to the transistor source and drain. Intel’s earlier PowerVia, on 18A, used an intermediate via; PowerDirect on 14A contacts the devices directly. TSMC’s Super Power Rail on A16 is also a direct-contact design. The practical differences lie in integration details, thermal behavior and which products get them: Intel uses backside power across 14A, while TSMC reserves it for A16 and A12.
Where do Samsung and Rapidus fit in the 1.4nm race?
Both sit behind the 2028 race. Samsung moved its SF1.4 node to 2029 and will keep it on standard EUV, focusing in the meantime on its SF2 family and backside-power SF2Z in 2027. Rapidus targets 2nm volume production in 2027 to 2028 and 1.4nm around 2029, but had no signed volume customer as of mid-2026. They matter mainly as price pressure on TSMC and as second-source options for designs that do not need the newest node.
Further Reading
- TSMC 2nm and the AI chip competition covers the N2 generation that both 1.4nm nodes build on.
- Intel 18A-P foundry analysis versus TSMC explains the bridge node Intel customers will use before 14A.
- Hyperscaler capex and AI compute economics shows the demand that funds both roadmaps.
- Hyperscaler custom AI silicon looks at the custom chips most likely to test a second foundry.
- TSMC’s 2026 process roadmap through 2029 from Tom’s Hardware.
- Intel’s Q3 2026 financial results for the 14A PDK and foundry figures.
By Riju — about
