The Change
TSMC's 2nm (N2) node is in severe shortage, with all three production fabs fully booked at 78 to 104 week lead times, pushing new order fill into 2028 - while simultaneously, CoWoS advanced packaging capacity is sold out through mid-2026, and SK Hynix has publicly confirmed its entire 2026 HBM supply is already allocated. Three independently constrained layers of the AI chip supply chain are hitting their ceilings at the same time.
Why the Market Has Not Fully Priced It
The consensus narrative runs roughly as follows: TSMC is supply-constrained, Nvidia benefits from priority access, Apple gets the largest consumer allocation, and the shortage normalizes by 2027-2028 as TSMC's $52-56 billion capex cycle completes. Sell-side models broadly anchor 2nm wafer pricing at $25,000-27,000 per wafer and assume packaging constraints loosen as TSMC's AP expansion reaches 130,000 CoWoS wafers per month by late 2026.
Three elements of this consensus are incomplete.
First, Apple's approximately 48-50% share of 2026-2027 N2 wafers is being consumed by consumer devices (iPhone A20, Mac M5/M6), not AI. Apple's volume does not compress when AI demand grows. Every wafer Apple runs through Fab 20, Fab 22, and Fab 21 Phase 3 is a wafer unavailable to Nvidia competitors, inference chip startups, and hyperscaler custom ASIC programs. The constraint is not just Nvidia versus its AI rivals - it is all AI chip programs competing for the slice of N2 that Apple leaves behind.
Second, CoWoS packaging and HBM memory are independently sold out in ways that are not resolved by the wafer capacity expansion. Nvidia has locked approximately 50% of TSMC's total advanced CoWoS capacity. SK Hynix's 2026 HBM supply is fully allocated. These are not consequences of the 2nm wafer shortage - they are parallel constraints with separate lead times. Even if a competitor secures 2nm wafer allocation, it still faces a packaging queue and an HBM sourcing problem that extends the total chip delivery timeline to 2027 or later.
Third, the per-wafer ASP at 2nm is already running above consensus model assumptions. Transactional pricing for N2 wafers is reported at $30,000 and rising. Against consensus models pricing wafers at $25,000-27,000, the gap represents approximately $1.5-2.0 billion in annual EBITDA not in current sell-side estimates for TSMC - based on approximately 90,000 wafers per month at late-2026 capacity, a $3,000-5,000 ASP delta, and TSMC's approximately 55% incremental gross margin on advanced node capacity.
Evidence
1. All three 2nm fabs are fully booked at 78-104 week lead times - a historical first.
Fab 20 (Hsinchu, Taiwan), Fab 22 (Kaohsiung, Taiwan), and Fab 21 Phase 3 (Phoenix, Arizona) are the three TSMC fabs currently producing 2nm geometry chips. As of Q1 2026, all three are operating at the maximum quoted lead time range of 78 to 104 weeks. At 104 weeks, an order placed today does not enter production until early 2028. For comparison, node transition lead times have historically not exceeded 52 weeks in the TSMC customer booking dataset. The 78-104 week reading has no precedent in publicly available customer reporting from prior node cycles (5nm, 3nm, 4nm). The constraint is not acute and temporary - it is at the outer limit of what TSMC's booking system can quote.
2. Apple's N2 allocation is a structural floor, not a variable.
Apple holds approximately 48-50% of 2026-2027 N2 wafer allocation, down slightly from prior nodes but still the largest single-customer share by volume. Wedbush semiconductor analysis and independent foundry allocation tracking services confirm this. Apple's consumer device production schedule is set 12-18 months in advance and is not elastic to AI chip demand signals. The remaining N2 capacity - approximately 50-52% of the total - is divided among Nvidia (by far the largest revenue customer at approximately $33 billion in 2026 TSMC revenue versus Apple's $27 billion), AMD, MediaTek, Qualcomm, and a range of hyperscaler custom ASIC programs. The effective AI chip allocation pool at 2nm is not "TSMC minus Nvidia" - it is "TSMC minus Apple minus Nvidia," which is a materially smaller number than the headline capacity ramp implies.
3. CoWoS packaging and HBM are parallel constraints that don't resolve with wafer supply growth.
TSMC CEO C.C. Wei confirmed in Q4 2025 guidance that CoWoS capacity "remains sold out through 2025 and into 2026." Nvidia separately confirmed that CoWoS assembly capacity is "oversubscribed through at least mid-2026." TSMC is expanding CoWoS from approximately 35,000 wafers per month in late 2024 to a target of 130,000 wafers per month by late 2026 - a 271% increase. Even at 130,000 wpm, Nvidia's locked 50% share equates to 65,000 wpm for Nvidia's own AI accelerator programs, leaving 65,000 wpm for all other customers. On the memory side, SK Hynix - which supplies approximately 50% of global HBM capacity - stated flatly that its entire 2026 HBM supply is allocated. HBM3E and HBM4 are both required for any competitive AI accelerator; there is no viable substitute in the performance tier these chips occupy. The bottleneck is not in a single layer but in three layers simultaneously, and the layers have different expansion timelines.
4. Andrej Karpathy's public statement frames the demand side: AI research throughput is now compute-bound, not researcher-bound.
On March 23, 2026, Karpathy - former Director of AI at Tesla and one of the five founding research scientists at OpenAI - stated publicly that "humans are now the bottleneck in AI research with easy-to-measure results." The investment implication is not about Karpathy specifically. It is a signal that AI model improvement is increasingly automated: AI systems writing code, running experiments, and evaluating results without waiting for human cycles. Automated AI research pipelines running continuously against GPU clusters represent a step change in sustained compute demand per unit of research output. Prior technology cycles produced demand with natural ceiling effects tied to human capacity - researcher bandwidth, enterprise procurement cycles, deployment windows. A compute demand driven by automated research loops does not have the same ceiling. This demand driver does not appear in consensus compute forecasts because no prior cycle had an equivalent.
Stay up to date on our best ideas
The Investable Bridge
Taiwan Semiconductor Manufacturing (TSM, NYSE, large-cap, approximately $900 billion market cap)
The direct pricing-power and visibility play. Each $1,000 increase in realized 2nm ASP above current consensus assumptions adds approximately $1.1 billion in annual revenue and approximately $600 million in net income at TSMC's approximately 55% gross margin on advanced capacity. The transactional-versus-consensus ASP gap of $3,000-5,000 per wafer represents $3.3-5.5 billion in annual revenue not in sell-side models at late-2026 wafer run rates. TSMC's fully booked lead times through 2028 are, in economic terms, contracted revenue visibility - not different in character from a long-term LNG offtake agreement, yet modeled with far less certainty by analysts. TSMC guided Q1 2026 revenue growth of 35-37% year-over-year; the pricing power component is not yet fully reflected in the 2027 outer-year estimates.
ASML Holding (ASML, NASDAQ, large-cap, approximately $280 billion market cap)
ASML is the sole supplier of EUV lithography machines required for 2nm production and the exclusive supplier of High-NA EUV machines required for TSMC's A16 node (the 2nm successor node, entering production in 2027-2028). TSMC's $52-56 billion 2026 capex is approximately 50-60% directed to ASML equipment by historical foundry capex mix. The High-NA ramp is the less-modeled element: each High-NA EUV machine carries an ASP of approximately $380 million, versus $180-200 million for standard EUV. Consensus models approximately 6-8 High-NA deliveries in 2026. Any revision toward 10 or more units would be above consensus and directly tied to the 2nm/A16 capacity buildout. ASML's order backlog stood at approximately 36 billion euros at year-end 2025 and covers the 2nm cycle as well as the parallel Samsung and Intel advanced node programs. The risk to ASML is customer concentration (TSMC is approximately 55% of revenue) and export license restriction on China sales, which has already been accounted for in current guidance.
Onto Innovation (ONTO, NYSE, mid-cap, approximately $4.4 billion market cap)
Onto Innovation makes process control, metrology, and inspection equipment for advanced logic and packaging nodes. At 2nm, defect density requirements tighten by approximately 30-40% versus 3nm, and the economic cost of a yielded defect at $30,000 per wafer creates strong incremental incentive for more inspection steps per wafer. Onto's Dragonfly G3 and Atlas platforms inspect advanced packaging interconnects and 3D IC structures - directly relevant to CoWoS and SoIC back-end packaging. In Q4 2025, Onto reported that orders for 2D packaging inspection for AI devices more than doubled sequentially. The company closed a $240 million multi-year volume purchase agreement with a leading HBM manufacturer for Dragonfly 2D inspection and 3D bump metrology through 2027, representing approximately 24% of full-year 2025 revenue booked in a single contract. ONTO trades at approximately 22-24x forward earnings versus a historical range of 28-32x during the 3nm-to-5nm node transition cycle in 2021-2022. The current discount relative to prior node transitions does not reflect the compound complexity increase at 2nm or the HBM inspection VPA visibility. Backlog at the most recent report date has nearly doubled year-over-year. Q2 2026 guidance was set above $300 million - the first quarter above that threshold.
Amkor Technology (AMKR, NASDAQ, mid-cap, approximately $5.5 billion market cap)
Amkor is the world's second-largest OSAT (outsourced semiconductor assembly and test) provider. The direct-CoWoS argument for Amkor is weak: Amkor does not provide a CoWoS substitute for AI accelerators. The relevant mechanism is different. As TSMC prioritizes its internal packaging capacity for high-value CoWoS AI workloads, the conventional OSAT volume that TSMC previously handled for non-AI customers - consumer electronics, automotive, RF, industrial - routes to third-party OSATs. Amkor is the largest beneficiary of this displacement. Amkor's GIGA Center in Vietnam and its Arizona facility, opened in 2023 adjacent to TSMC's Fab 21 in Phoenix, are positioned specifically for this category of volume. The Arizona co-location was a deliberate strategic move: semiconductor packaging adjacent to an advanced logic fab reduces logistics cost and enables co-optimization of assembly and test processes for the same customer base that TSMC Fab 21 serves. AMKR trades at approximately 9-10x forward earnings, which is a multi-year trough valuation. The Arizona utilization trajectory over the next two quarters is the specific metric to watch.
Risks and Failure Modes
The ASP gap thesis for TSMC depends on whether Apple and Nvidia's framework agreements allow transactional pricing above contracted price escalators. If large customers have multi-year price caps embedded in their supply agreements, TSMC's realized ASP improvement on existing volume commitments could be lower than spot-transactional pricing implies. The pricing power would then accrue only on incremental new orders, not on the full volume base.
The CoWoS expansion thesis is the most exposed to execution risk. TSMC's plan to reach 130,000 wpm from 35,000 wpm over approximately 24 months is a 271% increase in packaging capacity - faster than any prior TSMC packaging expansion. Equipment lead times for advanced packaging tools, qualified technician availability in new AP facilities (AP7 in Chiayi is the primary expansion vehicle), and yield ramp timelines are all variables where execution could slip. If TSMC reaches 100,000 wpm CoWoS by mid-2026 rather than 130,000, the constraint extends but loses some pricing leverage as partial relief arrives.
For ONTO, customer concentration is a real risk. A delay or reduction in TSMC capex spending would be directly transmitted to Onto's order rate with a 6-12 month lag. The HBM VPA provides revenue floor for 2026-2027, but growth above that floor is dependent on foundry and advanced packaging customer capex continuing at current rates.
The Karpathy-signal demand driver is the least quantifiable. If AI model scaling encounters empirical diminishing returns before the 2nm constraint resolves - specifically, if leading AI labs conclude that current architectures cannot efficiently use additional compute after a certain training scale threshold - the demand growth assumption weakens. This is a live debate in the AI research community.
What to Monitor Next
1. TSMC Q1 2026 earnings (expected April 2026) and N2 ASP commentary. TSMC does not report per-wafer ASP directly, but quarterly revenue divided by implied advanced node wafer shipments provides an inferred ASP range. A reading above NT$930,000 per 2nm wafer (approximately $28,500 USD) would confirm the pricing gap is larger than consensus models. Any forward commentary on wafer pricing power at the 2nm node is a direct thesis calibration event.
2. ASML Q1 2026 earnings - High-NA EUV order intake (expected mid-April 2026). The number of High-NA units in new orders and order backlog revision is the leading indicator for TSMC A16 production volume in 2027-2028. Consensus models 6-8 High-NA deliveries in 2026. Any guidance or order commentary suggesting 10 or more units is above consensus and extends the TSMC pricing power runway into the A16 generation.
3. Onto Innovation Q1 2026 results (expected April/May 2026) and advanced packaging order mix. The specific metric is the proportion of advanced packaging orders in the mix versus conventional logic. If AI device packaging orders maintain the greater-than-double sequential growth rate reported in Q4 2025, and if Arizona OSAT peers including Amkor report rising utilization in the same quarter, it confirms both the inspection demand thesis and the OSAT overflow thesis operating simultaneously.
This is for informational purposes only and does not constitute investment advice. Past observations are illustrative. Data may be incomplete or subject to revision.
