Foundry allocation and lead-time reference
Silicon Analysts tracks booking-to-delivery lead times and allocation status across semiconductor foundry capacity: leading-edge logic (TSMC N2, N3, N5, N7; Samsung 2nm and 3nm; Intel 18A; SMIC 7nm; Rapidus 2nm), mature and specialty nodes (UMC, GlobalFoundries), advanced packaging (CoWoS, SoIC, WMCM) and HBM supply from SK Hynix, Samsung and Micron. Lead times are for new capacity orders rather than existing long-term agreements. 7 of 13 records carry a quantified lead-time band; the remainder are tracked qualitatively because no defensible public number exists. A quarterly time series covers CoWoS-S, CoWoS-L, HBM3E and TSMC N3 from Q1 2022, capturing the 2023 shortage peak and the subsequent capacity relief.
Foundry Allocation & Lead Times
By our estimate, CoWoS (S + L) runs about 52 weeks booking-to-delivery; TSMC publishes no lead times. The longest estimated wait on a logic node is TSMC N2 / A16 at 78–156 wks. Lead times are our estimates for new orders, not existing LTAs — the number that decides whether a design can ship next year.
Baseline: April 2026 · signals current through 2026-10-26 · · Full series (34 observations, 4 series, 2022–2026) → · CSV · API
How current is this? The allocation table is a curated baseline last rebuilt in April 2026. A weekly pipeline tracks public capacity signals on top of it — most recently 2026-10-26 — and any record confirmed or moved since the baseline shows that signal inline below.
Changed this week
to
fab & CoWoS capacity — this page's own series, as compared in the change ledger.
No capacity readings moved past the significance threshold this week (2% or 1pp)
Lead-time history (quarterly, weeks)
CoWoS-L: 36 weeks (Q1 2026), Silicon Analysts estimate Mar 2026
2022–2024: 22 earlier prints — full ledger is Pro
- 1Q4 2023 — Peak Shortage: CoWoS sold out through 2024; HBM3E booked through 2025; 50+ week lead times confirmed by TSMC.
- 2Q2 2024 — Capacity Relief Begins: TSMC CoWoS capacity additions; Samsung HBM qualification; lead times begin gradual decline from peak.
- 3Q1 2026 — B300/Rubin Transition: HBM4 transition begins; CoWoS-L remains tight for next-gen chips while CoWoS-S eases.
Source: Silicon Analysts model, Industry estimates Q2 2024, DigiTimes Asia May 2024, TrendForce / SemiVision Jun 2024, Memory analyst consensus Jun 2024, TSMC investor day Dec 2024 and 3 more · Basis: booking-to-delivery · Anonymous view: the newest 3 observations per series (May 2024–Feb 2026); 22 earlier prints back to 2022 — full ledger is Pro.
CoWoS-S, CoWoS-L, HBM3E and TSMC N3, Q1 2022 through Q1 2026 — the series runs to the most recent quarter with a defensible public reading, not to today. Annotations mark the 2023 demand surge, the Q4 2023 shortage peak and the capacity relief that followed.
Pull this series into a spreadsheet: /api/v1/market-data/component-lead-times?format=csv · Excel & Google Sheets setup →
Current allocation by node
| Foundry | Node | Status | Lead time (our est.) |
|---|---|---|---|
| TSMC | N2 / A16 | Allocated (est.) | 78–156 wks |
| TSMC | N3E / N3P / N3B | Allocated (est.) | 104–156 wks |
| TSMC | N5 / N4P / N4NP | Constrained | 22–32 wks |
| TSMC | N7 / N6 | Available | 8–16 wks |
| TSMC | N16/12 / N28/40 | Available | 4–12 wks |
| Samsung | SF2 / SF2P (2nm) | Constrained | Not quantified |
| Samsung | 3GAE / 3GAP (3nm) | Available | Not quantified |
| Samsung | 4LPP (4nm) | Constrained | Not quantified |
| Intel | Intel 18A | Early Access | Not quantified |
| GlobalFoundries | 12nm-22nm | Available | 10–16 wks |
| SMIC | N+2 (7nm) | Severely Constrained | Not quantified |
| UMC | 14nm-28nm | Available | 8–14 wks |
| Rapidus | 2nm GAA | Early Access | Not quantified |
“Not quantified” means no defensible public lead-time number exists for that node — status and capacity are still tracked. Lead times cover new capacity orders; existing long-term agreements are priced and scheduled separately.
This table as CSV: /api/v1/allocation?format=csv
Advanced packaging
CoWoS (S + L)
TSMC + OSATs
TSMC said its advanced packaging capacity is very tight and that it has to work with its OSAT partners (TSMC 1Q26 call, 2026-04-16), that its back-end capacity is still in shortage mode (2Q26 call, 2026-07-16), and that CoWoS is still the number one choice for its customers (3Q26 call, 2026-10-15). Advanced packaging was slightly over 10% of TSMC revenue in 2025, and TSMC expects low-teens percent in 2026 (4Q25 call, 2026-01-15). TSMC's 4Q25-3Q26 calls give no CoWoS capacity, lead time or customer split. Reported, not a TSMC statement: the NVIDIA, Broadcom, AMD, Amazon/AWS and Marvell shares of 2026 CoWoS wafer demand in this record come from a 36Kr report whose NVIDIA figure cites a Morgan Stanley report (36Kr, 2025-12-04); the Amazon/AWS and Marvell shares are the wafers it says each reserved (50,000 and 55,000) over its 1 million-wafer 2026 total, and it says MediaTek reserved 20,000 wafers for Google's TPU project, which this record counts in the Others share. The capacity, lead time, allocation tier and the residual Others share in this record are Silicon Analysts estimates from trade reports, not TSMC figures.
Customer Allocation
2026 target: 130K WPM
SoIC (3D hybrid bonding)
TSMC
TSMC's 4Q25-3Q26 earnings calls give no SoIC capacity, growth rate or customer figures; asked about SoIC on its 1Q26 call, TSMC said it works with its customers to meet their demand (TSMC 1Q26 call, 2026-04-16). The capacity values, status and customer shares in this record are Silicon Analysts estimates from press reports, not TSMC figures.
Customer Allocation
2026 target: 20K WPM
WMCM (Wafer-Level MCM)
TSMC
TSMC's 4Q25-3Q26 earnings calls do not mention WMCM, and TSMC does not comment on specific customers (TSMC 3Q26 call, 2026-10-15). The capacity values, status and the Apple share in this record are Silicon Analysts estimates from press reports, not TSMC figures; its 2026 capacity target of 60,000 wafers/month is a Silicon Analysts estimate.
Customer Allocation
2026 target: 60K WPM
CoWoS is the binding constraint on AI accelerator supply. SoIC and WMCM are tracked by status and capacity ramp only — neither has a public booking-to-delivery figure we can stand behind.
HBM supply
SK Hynix
50% of HBM revenue, Q2 2026 (Counterpoint Research)
“Discussion is underway for 2027 HBM supply volume and pricing with our key customers” — Kim Woo-hyun, CFO (via interpreter), SK hynix Q2 2026 call (Investing.com transcript), 2026-07-29.
HBM3E: Higher sales of 12-high HBM3E and server DDR5 drove record Q3 2025 results (SK hynix 3Q25 results, 2025-10-29).
HBM4: Development completed and mass production entered in 2025 (3Q25 results, 2025-10-29); mass shipments began in Q2 2026 (2Q26 results, 2026-07-29).
FY2025 revenue KRW 97.1T (+47% YoY), with a 58% operating margin in Q4 2025 (SK hynix FY2025 results, 2026-01-28). SK hynix says it has finalized long-term memory supply agreements, not specific to HBM, with around 10 customers, including key strategic partners (2Q26 results, 2026-07-29).
HBM3E $/stack (our est.)
$330
HBM4 $/stack (our est.)
$520
NVIDIA HBM4 allocation: ~55% (reported by Hankyung, 2026-02-09, citing industry sources, as the mid-50% range of the provisional HBM4 volumes NVIDIA allocated in December 2025; a press report, not a supplier or NVIDIA figure). Hankyung article
Key customers (public reports)
NVIDIA · Google · AMD
Samsung
33% of HBM revenue, Q2 2026 (Counterpoint Research)
“HBM supply agreements that we have already signed with our customers for 2027” — Samsung Q2 2026 call (Investing.com transcript), 2026-07-30. Samsung did not say what share of its 2027 HBM supply these agreements cover.
HBM3E: Being sold to all related customers (Samsung 3Q25 results, 2025-10-30).
HBM4: Mass production and first commercial shipments announced 2026-02-12, on 1c DRAM with a 4nm logic base die (Samsung newsroom).
Share of HBM revenue rose from 15% in Q2 2025 to 33% in Q2 2026 (Counterpoint Research). Samsung said on 2026-02-12 that it expects its HBM sales to more than triple in 2026 versus 2025.
HBM3E $/stack (our est.)
$310
HBM4 $/stack (our est.)
$500
NVIDIA HBM4 allocation: ~25% (reported by Hankyung, 2026-02-09, citing industry sources, as the mid-20% range of the provisional HBM4 volumes NVIDIA allocated in December 2025; a press report, not a supplier or NVIDIA figure). Hankyung article
Key customers (public reports)
Google · Broadcom · NVIDIA · AMD
Micron
18% of HBM revenue, Q2 2026 (Counterpoint Research)
“We have completed agreements for the vast majority of our calendar 2027 HBM bit supply” — Sanjay Mehrotra, CEO, Micron FQ4 2026 call (Investing.com transcript), 2026-09-30.
HBM3E: 12-high HBM3E made up the majority of its HBM shipments in fiscal 2026 (Micron FY2026 Form 10-K, filed 2026-10-09).
HBM4: Volume production of HBM4 36GB 12-high began in fiscal 2026 (year ended 2026-09-03); 48GB 16-high samples delivered to multiple key customers (FY2026 Form 10-K).
Fiscal Q2 2026 revenue $23.86B and a 74.9% non-GAAP gross margin (Micron results, 2026-03-18). Its HBM4 base logic die and DRAM core dies are designed and manufactured in-house (FY2026 Form 10-K).
HBM3E $/stack (our est.)
$320
HBM4 $/stack (our est.)
$510
NVIDIA HBM4 allocation: ~20% (reported by Hankyung, 2026-02-09, citing industry sources, as about 20% of the provisional HBM4 volumes NVIDIA allocated in December 2025; a press report, not a supplier or NVIDIA figure). Hankyung article
Key customers (public reports)
NVIDIA · AMD
HBM is contracted through long-term agreements rather than booked at a quoted lead time, so each card shows its supplier’s own latest statement on contracted supply, quoted with its source and date, instead of a week count.
Recent allocation changes
- 2026-07-29
SK hynix begins HBM4 mass shipmentsSK Hynix HBM4
SK hynix said it began mass shipments of HBM4 in the second quarter and will ramp production in the second half (2Q26 results, 2026-07-29). Hankyung reported that NVIDIA's provisional HBM4 allocation, made in December 2025, gives SK hynix a share in the mid-50% range (Hankyung, 2026-02-09; a press report, not an SK hynix or NVIDIA figure).
Source: SK hynix 2Q26 results, 2026-07-29; Hankyung, 2026-02-09
- 2026-04-02
Samsung Taylor fab reported to begin EUV tool testingSamsung Samsung 2nm
Reported, not a Samsung statement: the Taylor, Texas fab began testing EUV lithography equipment in March 2026 (The Korea Herald, 2026-04-02, citing industry sources). Samsung said its foundry business would focus in 2026 on beginning operations at its Taylor fab in a timely manner (Samsung 3Q25 results, 2025-10-30). Samsung's DART filings disclose a KRW 22.76 trillion (about $16.5B) foundry contract with Tesla running to 2033-12-31 (filed 2025-07-28; Tesla named in the 2025-07-31 correction); the filings name no fab.
Source: The Korea Herald, 2026-04-02; Samsung 3Q25 results; Samsung DART filings
- 2026-04-01
Intel buys back Apollo's Fab 34 joint-venture stake for $14.2BIntel Intel 4 / Intel 3
Intel agreed on 2026-04-01 to repurchase the 49% equity interest that Apollo-managed funds held in the joint venture related to its Fab 34 in Ireland, for $14.2B financed with cash on hand and a $6.5B bridge loan; the repurchase closed on 2026-04-08, and Intel owns 100% of the joint venture (Intel Form 8-K, 2026-04-08). Intel's FY2025 Form 10-K lists its Ireland production on Intel 4 and Intel 3.
Source: Intel Form 8-K, 2026-04-08; Intel FY2025 10-K
- 2026-03-31
Taiwan approves 3nm for TSMC's second Japan fabTSMC N3
Taiwan's Ministry of Economic Affairs approved on 2026-03-31 TSMC's plan to use 3nm at its second fab in Kumamoto, Japan, a facility approved in June 2024 for 6- to 12nm chips (reported by the Taipei Times with CNA, 2026-04-01). TSMC said on its 1Q26 call (2026-04-16) that it now plans 3-nanometer technology for its second Japan fab, with volume production scheduled in 2028.
Source: Taipei Times (CNA), 2026-04-01; TSMC 1Q26 call, 2026-04-16
- 2026-03-18
Micron reports record $23.86B quarterly revenueMicron HBM
Revenue of $23.86B versus $8.05B a year earlier, with a 74.9% non-GAAP gross margin (74.4% GAAP); Micron said it set new records across revenue, gross margin, EPS and free cash flow (fiscal Q2 2026 results, 2026-03-18).
Source: Micron fiscal Q2 2026 results (Form 8-K)
- 2026-02-27
Rapidus raises $1.7B from the IPA and 32 companiesRapidus Rapidus 2nm
Rapidus completed a 267.6 billion JPY (about $1.7B) funding round: 100 billion JPY from the Information-Technology Promotion Agency (IPA), under METI, and 167.6 billion JPY from 32 companies including Canon, Fujitsu, NTT, SoftBank and Sony Group. Rapidus said the funding lets it progress to mass production of 2nm logic semiconductors by 2027 (Rapidus, 2026-02-27). In a joint release with Preferred Networks (PFN) and SAKURA internet, Rapidus said that under a basic agreement among the three it manufactures the latest model of AI semiconductors in the MN-Core series to be designed by PFN (Rapidus, 2025-01-08).
Source: Rapidus, 2026-02-27; Rapidus, PFN and SAKURA internet, 2025-01-08
- 2026-02-12
Samsung ships commercial HBM4, which it calls an industry firstSamsung HBM4
Samsung said it had begun mass production of HBM4 and shipped commercial products to customers, a first in the industry by its account, built on its 1c DRAM process with a 4nm logic base die (Samsung Newsroom, 2026-02-12).
Source: Samsung Newsroom, 2026-02-12
- 2026-01-28
ASML reports EUR 13.2B in Q4 net bookingsASML Equipment
Q4 2025 net bookings were EUR 13.2B (EUR 7.4B of it EUV systems), 56% from Memory and 44% from Logic, and ASML ended 2025 with a backlog of about EUR 38.8B (ASML Q4 2025 investor call, 2026-01-28).
Source: ASML Q4 2025 investor call transcript, 2026-01-28
Latest entry 2026-07-29 — this log tracks structural allocation moves, not weekly signal readings.
Methodology
Lead times are Silicon Analysts estimates of booking-to-delivery for new capacity orders, compiled from trade press, procurement reporting and supply-chain research; foundries do not publish them, and earnings calls supply only context (capacity, tightness). Existing long-term agreements are scheduled and priced separately and are not represented here.
The allocation table is a curated baseline. A weekly agent searches public sources for capacity signals, requires corroboration across independent sources before publishing, and stacks each weekly reading rather than overwriting the previous one — so a record shows how an estimate evolved, not just its latest value. Readings that are single-source or that move sharply are held for human review. A reading whose only source is a social-media post is shown at low confidence and marked “SNS · unconfirmed”, and is never a newsletter headline or a “changed this week” move. An older reading cited only to one of our own pages is shown the same way, marked “Unconfirmed source”. A customer-allocation share that no primary or trusted source confirms is shown at low confidence, marked “Unconfirmed estimate · secondary/social sources”, until a trusted source confirms it. A share no public source states is our own estimate, shown at low confidence and marked “Our estimate · no public source”.
Where a node carries no lead-time band, that is a deliberate absence rather than a gap in collection: no public figure exists that survives corroboration. We would rather show “not quantified” than infer a number.
For full data provenance and terms of use, see Terms & Data Provenance.
Filter allocation by foundry, node or customer →
Allocation Explorer — the same records with per-node detail, customer concentration and capacity ramp targets.
What the constrained supply costs →
HBM Pricing — contract and blended $/GB by generation, dated and sourced.
One short sourced email a week, plus an alert only when this series moves past 2% or is restated. No ads, one-click unsubscribe.
Weekly brief + a watch alert only when fab & CoWoS capacity moves past 2% or is restated — at most once a week. No ads. One-click unsubscribe.