Autoresearch: does CoWoS 4× show up in BESI hybrid-bonding backlog?

Generated by /autoresearch on 2026-08-17. Synthesized across 3 rounds from 8 successful web fetches (1 timeout), anchored by Grokipedia Besi. See Provenance. Treat as raw material — review before promoting into a project or thread. Context: vault/projects/stock-market

Priors capture skipped — unattended weekday shift; user instructed skip.

X pass enabled. 0 high-signal posts surfaced (0 via direct fetch, 0 via search snippets). Four site:x.com / site:twitter.com discovery queries returned no permalinks. X MCP needsAuth in this environment. See Provenance.

Summary

The June 2026 question asked whether TSMC's CoWoS 4× expansion (wiki: ~35k → ~130k wpm; Layer 3 street 140k → 220k) translates into a BESI hybrid-bonding backlog multi-year growth cycle, versus TSMC making tools in-house or buying Applied / others. This pass finds the order cycle is real and still compounding after Q1, but the identity is wrong.

BESI's Q2-26 print (July 23) beat the Q1 +30–40% sequential revenue guide: revenue €249.9M (+35.2% QoQ, +68.7% YoY), orders €292.9M (+8.6% QoQ, +128.8% YoY), LTM orders a record €987.6M. Hybrid-bonding customers went 15 → 21. Q3 is guided +10–15% sequential. That is not a one-quarter surge that then digested (Besi IR Q2-26; Roic AI Q2 transcript).

What it is not: a clean "CoWoS wafer × 4 = BESI hybrid-bonder backlog." CoWoS-like (flip-chip / TCB / oS) is already "a significant part of the orders." Hybrid bonding is the SoIC / AP7 / CPO / future-HBM layer. TSMC is buying BESI systems for AP6/AP7, not making bonders in-house. Applied is a 9% partner, not a substitute. The August CoW overflow to ASE is talking to Korean dicing/bonding suppliers (Hanmi et al.), not naming BESI. HBM hybrid bonding is still evaluation at the memory three — the 2027–2030 second cycle remains unverified for volume.

The wiki's €400M backlog was an analyst calculation on the Q1 call, not a BESI disclosure. BESI does not publish a backlog number.

Findings

Theme 1 — Q2-26: the order cycle compounded. It did not digest

Besi reported Q2 and H1 2026 on 2026-07-23 (Besi IR; full table in the Yahoo GlobeNewswire reprint):

Q1-26 official (Besi IR Q1): revenue €184.9M (+28.3% YoY), orders €269.7M (+104.5% YoY), "primarily due to a significant increase in bookings for hybrid bonding systems." Q2 guide was +30–40% sequential revenue / 64–66% GM. Q2 beat the revenue guide (+35.2%) and printed 65.7% GM.

Q3-26 outlook (Besi IR Q2): revenue +10–15% vs €249.9M; GM 63–65% on a less favorable mix; opex flat to +5% vs €55.3M.

CEO Richard Blickman on the Q2 call (Roic AI): LTM orders a record €987.6M; AI applications ~60% of H1-26 system orders vs ~50% in H1-25; "capacity shortages, particularly in memory and in CoWoS and CoWoS-like"; "customers indicate that we are in a multiyear AI CapEx cycle"; strategy is "increasing our penetration of CoWoS, CoPoS, and Photonics." Reuters via CNA quotes the same multi-year / Q3-momentum language.

Book-to-bill (computed, not disclosed): Q2 292.9 / 249.9 ≈ 1.17; H1 562.6 / 434.7 ≈ 1.29. Orders still exceed shipments. That is the operational content of "backlog grew" without a company number.

Theme 2 — The €400M backlog was never official

The question page and the BESI entity treat "~€400M backlog" as a Q1-26 fact. BESI does not disclose backlog. On the Q1 call, Morgan Stanley's Nigel van Putten said: "I know you don't really disclose the backlog, but I calculate around EUR 400 million by March end" (Roic AI Q1 transcript; Investing.com transcript reprint). Blickman did not confirm the figure. Q2 IR and the Q2 call again say only "based on our backlog and feedback from customers" for the Q3 guide — no euro amount, no unit count (Besi IR Q2; Roic AI Q2). Asked for hybrid-bonder units shipped or booked in Q2, Blickman: "No, we haven't disclosed units. We did disclose the number of customers" (same transcript).

Correction for the wiki: treat €400M as an analyst estimate as of 2026-03-31, not a BESI primary. Visibility evidence is orders + sequential revenue guides, not a backlog line.

Theme 3 — Hybrid bonding is compounding. It is not the same thing as CoWoS 4×

Hybrid-bonding customers: 15 at YE2025 → 21 at end-Q2-26. Use cases: logic, memory, co-packaged optics, consumer. Q2 HB orders: two repeat customers + one new hyperscaler. Blickman split those as CPO, logic, and (probably) wearables — not HBM volume (Roic AI Q2).

On CoWoS specifically, JPMorgan's Sandeep Deshpande asked whether Besi had "broken into the CoWoS market" (CoWoS-L and original CoWoS at contract manufacturers). Blickman: "That's already a significant part of the orders. We always call that CoWoS-like. And that already started, you can say, in Q4, Q1, Q2 and should continue to expand" (same). Asked what "increasing penetration of CoWoS and CoPoS" means — flip-chip or TCB? — Blickman: "Both, and also hybrid." He also named "the oS, on-substrates from CoWoS, same like chip-on-wafer" as part of the 2.5D roadmap (same).

Read: CoWoS wafer expansion shows up in BESI as CoWoS-like assembly tools (flip-chip / TCB / oS) plus a hybrid-bonding overlay for SoIC / CPO / panel. It does not mean every incremental CoWoS wpm is a hybrid bonder. Grokipedia Besi is the vocabulary primer: 43% flip-chip/die-attach share (2023); hybrid Gen 2 (100 nm) / Gen 3 (50 nm); 9800 TC Next for HBM stacks up to 16 layers; 8800 Chameo for 2.5D CoW. Treat the encyclopedia as a primer, not a 2026 order table.

TC Next (10–20 µm fluxless TCB) is expanding in qualifications; "systems ordered, single systems, not yet major volume. There's one who has several" (Roic AI Q2). A large 2025 memory TCB order has not repeated — Blickman said that is generation-timing, not a hybrid-bonding hesitation (same).

Theme 4 — TSMC is buying BESI tools for AP6/AP7. It is not making bonders in-house

Blickman, Q2: "we have started now to fill the AP7 with the first systems, which is publicly informed to the world." On shipment cadence: "published total capacity should be roughly double that of AP6. In AP6, there are about 60 bonders… first round orders received in Q1. There may be another round in the second half… lead time of about 6 months" (Roic AI Q2).

On the Q1 call he said AP7 first-round orders were pulled forward from Q2 to Q1, the program enlarged toward an "anticipated 100 bonders. That number we are told may be significantly higher, plus orders placed for co-packaged optics." Hybrid-bonder factory capacity was raised 180 → 250 systems/year; 100 nm lead time ~6 months (Roic AI Q1).

TrendForce (2026-01-21, citing Commercial Times) identifies AP7 as TSMC Chiayi: Plant 2 equipment install H2 2025 / production 2026; Plant 1 move-in 2026 / MP 2027; initial WMCM for Apple iPhone 18, then SoIC and 2.5D; CoPoS mass production targeted end-2028 on 310×310 mm (equipment validation from Q3 2026). TSMC advanced-packaging capex CAGR 24% 2025–27; packaging + photomask 10–15% of 2026 capex (TrendForce). Blickman independently said Taiwan selected 310×310 and Besi already has panel orders (Q2 call).

BNP's Martin Jungfleisch asked whether the "big Taiwanese customer" might accelerate CoWoS in favor of SoIC and slow hybrid-bonding orders. Blickman: "we don't see that currently. We only see continued… pressure on us to expand the capacity" (Roic AI Q2).

Internalization risk, as framed in the question (TSMC builds hybrid-bonding capability in-house): the fetched record is the opposite for tools. TSMC is filling AP7 with BESI systems and pressing for more. TSMC does internalize CoWoS process capacity (Layer 3: 10 packaging facilities, street 140k → 220k wpm) — that is wafer capacity, not a BESI-replacement tool line. No fetched source says TSMC designs or builds hybrid bonders.

Onshoring (Q1 call): TSMC / Amkor / Micron US programs — "in the next 2 years, so '26, '27, preparation, building fabs and then as of '28, volume production" (Roic AI Q1). That matches Layer 3's Amkor Arizona 2028 clock. It is not 2026 BESI revenue.

Theme 5 — Applied is a partner. ASE overflow talks name Korean TCB/dicing, not BESI HB

Applied Materials bought 9% of Besi on 2026-04-14 (market purchases; no board seat; no further-share plans). The firms have collaborated since 2020 on "the industry's first fully integrated equipment solution for die-based hybrid bonding" — Applied front-end process + Besi die placement (Applied IR). Blickman on Q2: "Besi with its hybrid bonding technology, together with Applied Materials in automated line concepts is at the forefront" for memory (Roic AI Q2). Applied is an enabler / shareholder, not the substitute the question feared.

Layer 3's August CoW-outsourcing wave (TrendForce/ETNews, 2026-08-05): TSMC expanding CoW to OSATs "including ASE"; ASE / Amkor / SPIL already licensed for WoS; OSAT equipment POs focused on dicing and bonding; "PO discussions are underway with South Korean materials, components, and equipment suppliers" (TrendForce). Secondary recaps name Hanmi, Avaco, Wonik for laser/bonding (search-level; not re-fetched as primary). BESI is not named in the fetched CoW-overflow equipment list.

Implication for the question: 2026 TSMC captive packaging (AP6/AP7, SoIC/hybrid) is the BESI HB vein. 2026 ASE overflow residual is a different tool basket (Korean TCB/dicing first in the Korean press). Do not equate "ASE takes 2026 CoW wafers" with "BESI books those lines."

Competition the question named: ASMPT is the usual die-to-wafer HB peer in industry commentary (not first-party this pass). Kulicke & Soffa / Hanmi are TCB/HBM. EV Group / SUSS are wafer-to-wafer. TOWA is molding-adjacent in the same commentary set — no fetched 2026 TOWA-vs-BESI CoWoS order. Grokipedia Besi names K&S and ASMPT as primary assembly rivals.

Theme 6 — HBM hybrid bonding is still the unverified second cycle

The question's 2027–2030 HBM4/HBM4E hybrid-bonding wave is not 2026 backlog. Q2: evaluations "at one of the major — one of the 3 memory companies"; "another one is starting"; "some are voicing that it may delay. Others… that it is imminent." All three have tools. Hurdles: cost and yield. Taiwan logic took "over 3 years" to reach yield/cost; US logic is "finally… becoming a mainstream application." H2-26 / Q3–Q4 is when it becomes clearer whether HBM4E or even HBM3 goes hybrid (Roic AI Q2).

JEDEC loosening HBM4E thickness: Blickman said height relief lets the industry ship more TCB devices into a shortage, but "does not change" the performance/heat case for hybrid. "The year '26, all 3 are preparing to adopt at some point" (same).

Q1 color (secondary recap of the April call, useful only as a pointer back to the transcript): first memory customer targeting mid-2026; second end-Q3/Q4; all three evaluating "for one specific end customer" with hybrid-bonded stacks wanted by end-2026 for 2027 end-markets (druckfin Q1 recap — qualify as secondary). Blickman on Q1 (Bits&Chips): hybrid-bonding adoption "increasing as we approach the timing for new AI-related product introductions anticipated in the 2027-2030 period" (Bits&Chips).

Do not graduate the HBM4 hybrid-bonding leg. Gate: Q3-26 print 2026-10-22 (Besi calendar on the Q1 release) and any first-party memory-maker hybrid-HBM production statement.

Theme 7 — Cycle risk is live. Blickman will not call a super-cycle

Deutsche Bank's Robert Sanders asked about 2.5D overcapacity (CoWoS look-alikes vs China). Blickman: typical up-cycle 6–8 quarters then overcapacity; "as soon as you hear new fabs… and then also the whole subcontractor universe expanding… you're bound to… see at some point an overcapacity." "There are many always other prophets who are saying, no, this time, it's different… We've also heard that before." OSATs making ~18% GM going after ~40% CoWoS-look-alike GM is "the name of the game." Besi adjusts the supply chain "3 times a week" (Roic AI Q2).

This is a first-party caution against reading H1-26 order doubling as a locked multi-year compounder. The multi-year customer comment (agentic AI / data-center CPUs) and the 6–8 quarter equipment-cycle comment sit next to each other. Leave both on the page.

No August 2026 Besi update after July 23. Next official gate is Q3 on 2026-10-22.

Contradictions and open questions

• CoWoS 4× ≠ hybrid-bonding units. The question's chain step 2 ("each incremental CoWoS tool line requires BESI hybrid bonders") is too tight. CoWoS-like is already a significant order slice and is flip-chip/TCB/oS as well as hybrid. Hybrid bonding is SoIC / AP7 / CPO / future HBM.
• €400M backlog is not a BESI number. Analyst calc at Q1. Do not treat it as confirmed or as a Q2 run-rate.
• TSMC internalization. Process capacity is being internalized (Layer 3). Tools are being bought from BESI (AP6 ~60, AP7 filling toward ~100+). Those are different claims.
• ASE 2026 overflow ≠ BESI HB. Layer 3's residual is ASE. The fetched CoW equipment talks name Korean suppliers. Besi's "CoWoS-like" OSAT orders are real (Q4–Q2) but unquantified vs TSMC-captive AP7.
• HBM hybrid bonding volume: still evaluation. Needham-style 2027–2030 second cycle is Blickman's own horizon language, not a booked backlog.
• Units undisclosed. Customer count (21) is not a wafer or tool-unit bridge to 140k→220k wpm.
• Already priced in. Q1 doubling was public in April; Q2 confirmed continuation. This pass does not re-underwrite BESIY / BESI.AS multiples. Twelve Data still cannot price BESI on the free plan (wiki ledger).
• No August primary after July 23. Do not invent a Q3 preview.

Provenance

Rounds run: 3 of 3 (full; Round 3 was a short "any August update?" pass — none found)

Sub-questions by round:

Round 1 (broad survey):

1. What did BESI report for Q2 2026 (and any Q3 comments) on orders, backlog, hybrid vs TC mix?
2. Have TSMC, ASE, or Amkor been named as BESI hybrid-bonder customers for the CoWoS expansion?
3. Who else supplies hybrid bonding / TCB (Applied, TOWA, ASMPT, K&S), and is TSMC making tools in-house?
4. Does CoWoS actually use hybrid bonding, or is it still TCB / microbump?

Round 2 (drill-down):

1. Is AP7 TSMC's packaging fab, and what is public on BESI tool count? — targeted the Taiwan/AP7 unnamed-customer gap
2. Is the €400M backlog official? — targeted the wiki's load-bearing backlog figure
3. Does ASE/Amkor CoW overflow generate BESI orders vs Hanmi/TOWA? — targeted Layer 3 overflow vs BESI identity

Round 3 (resolve remaining uncertainty):

1. Any August 2026 Besi comment after the July 23 print? — targeted a stale-Q2 risk. None found; next gate 2026-10-22.

Anchor source (Grokipedia, fetched before round 1):

• Besi — 25,017 chars extracted (capped) via _lib/grokipedia.py — primer on Blickman, die-attach share, HB Gen 2/3, TC Next, Applied 9% (April 2025), 2025 Investor Day €1.5–1.9B 2029 target. Stale vs 2026 prints; not used as an order table.

X sources (--include-x; x-fetch skill + _lib/x.py):

• 0 items surfaced (0 direct fetch, 0 search snippets)
• Queries: BESI Q2 2026 earnings hybrid bonding backlog site:x.com; BESI hybrid bonding TSMC ASE Amkor tool orders 2026 site:x.com; BESI Q2 2026 results hybrid bonding backlog site:x.com; "BESI" OR "BE Semiconductor" hybrid bonding 2026 site:twitter.com; BESI Blickman hybrid bonding site:x.com 2026
• X MCP needsAuth in this environment. Same failure mode as the 2026-08-17 CSP CapEx, private-credit, and Amkor layers.

URLs fetched (8 successful, 1 timeout):

Round 1:

• Besi Q2-26 / H1-26 results (2026-07-23) — official / IR — revenue, orders, GM, Q3 guide
• Applied Materials 9% stake (2026-04-14) — official / IR — partner, not acquirer; integrated D2W HB line
• CNA / Reuters Besi orders (2026-07-23) — news (Reuters) — 21 customers; Blickman multi-year quote

Round 2:

• TrendForce TSMC AP7 Chiayi (2026-01-21) — industry press — AP7 = TSMC Chiayi; WMCM/SoIC/CoPoS timing
• TrendForce CoW outsourcing (2026-08-05) — industry press — ASE-first CoW; Korean equipment POs
• [Failed/timeout: https://www.globenewswire.com/news-release/2026/04/23/3279584/0/en/BE-Semiconductor-Industries-N-V-Announces-Q1-26-Results.html] — timeout; Q1 numbers taken from Besi IR Q1 instead

Round 3:

• Besi Q1-26 results (2026-04-23) — official / IR — Q1 orders/revenue; Q2 guide that Q2 then beat
• Bits&Chips Q1 recap — trade press — Blickman 2027–2030 adoption quote

Also used (search-extracted full text, not a separate WebFetch):

• Roic AI BESI.AS Q2 FY2026 transcript — earnings-transcript reprint — Blickman on AP6/AP7, CoWoS-like, HBM eval, Applied, units undisclosed
• Roic AI BESI.AS Q1 FY2026 transcript — earnings-transcript reprint — AP7 100+ bonders; 180→250/yr; van Putten €400M calc
• Yahoo / GlobeNewswire Q2 reprint — wire reprint — full Q2 table + Blickman prepared remarks
• Investing.com Q1 transcript reprint — transcript reprint — corroborates the €400M-as-analyst-calc line

Tools used: WebSearch, WebFetch, grokipedia-fetch (skill + _lib/grokipedia.py), x-fetch (skill; discovery only, zero items). Generated: 2026-08-17 14:05 UTC