Hybrid bonding in advanced packaging for logic and HBM memory Report Interpretation

Overview

The report reassesses the timing of hybrid-bonding adoption in high-bandwidth memory after a shift toward lower HBM stack heights. Bernstein argues that the near-term delay is temporary: logic applications and eventual higher-stack HBM requirements should still drive substantial hybrid-bonder demand by 2028.

Core views

Bernstein argues that broad hybrid-bonding adoption in HBM is now unlikely within 2027. Industry attention has shifted toward 8-hi rather than 12-hi configurations for at least HBM4, and likely HBM4E, pushing 16-hi designs and the associated need for hybrid bonding further out. The report attributes this to memory supply constraints that require customers to spread memory across more XPUs, shipment challenges for some 12-hi products, and high memory costs. It therefore maintains its prior expectation that broad adoption is more likely with HBM5 in 2028. The delay does not change the report's long-term technology case. HBM packaging is constrained by the current JEDEC total-package-height limit of 775μm: adding layers forces thinner DRAM dies, increasing warpage and heat-related yield risks. Hybrid bonding removes microbumps and underfill, allowing core dies to be 24% thicker; it can reduce TSV/pad pitch to below 18μm versus 20–30μm for MR-MUF, raising interconnect density and bandwidth; and its copper-to-copper/oxide interface improves heat transfer, with approximately 35% lower thermal resistance at 20 stacks. Even if JEDEC relaxes the height limit to about 900μm, Bernstein sees this as extending conventional technology rather than eliminating the eventual need for hybrid bonding. SK hynix's roadmap suggests partial adoption from 16-hi and full adoption from 20-hi. Logic remains the principal near-term growth engine. Bernstein expects die-to-wafer hybrid bonding to expand in chiplet integration, XPU-to-memory designs and co-packaged optics because it provides denser, lower-power interconnects. It forecasts 2028 logic hybrid-bonding TAM of $887mn and 237 shipments, up from 103 in its prior estimate, reflecting stronger expected adoption at TSMC for SoIC packaging. The firm models the global logic hybrid-bonder installed base growing 4.5x from 137 units in 2025 to 582 in 2028. TSMC is expected to expand from about 65 units to about 440 units and increase SoIC capacity from 10kwpm at end-2025 to 80kwpm in 2028; Intel is modeled to rise from about 30 to about 65 units by 2028, with capacity additions resuming from 2027. HBM hybrid-bonding adoption is modeled as more gradual. Current TCB solutions remain effective, HBM4E is likely to remain mostly at 8/12-hi, and Samsung and SK hynix are pursuing alternative heat-dissipation approaches. Bernstein nonetheless expects Samsung, followed by SK hynix and Micron, to begin adoption in 2028 for 16-hi HBM5, while TCB remains mainstream through that point. Its HBM forecast assumes 15% blended hybrid-bonding penetration across 1,202kwpm of HBM capacity, implying $489mn of hybrid-bonding revenue and $1.8bn in total HBM bonding revenue including TCB. HBM accounts for 36% of the estimated $1.4bn total hybrid-bonder TAM in 2028. For Besi, Bernstein estimates roughly $1.1bn of 2028 hybrid-bonding revenue from about 281 units at a $3.75mn ASP, assuming 76% market share—80% in logic and 70% in HBM. The firm views the HBM timing pushout as non-permanent because Besi still benefits from logic hybrid bonding and conventional packaging. For DISCO, it argues that strong overall DRAM wafer consumption should sustain dicing and grinding demand regardless of whether memory uses HBM or conventional packaging; NAND wafer-to-wafer bonding, BSPDN and potential silicon-photonics upside are additional drivers. On memory stocks, Bernstein expects supply to remain short in 2027 and margins exiting 2028 to remain very healthy even if memory normalizes in 2028, supporting its Outperform ratings on Samsung, SK hynix and Micron; it retains Underperform on KIOXIA because of the long-term threat from China.

Analysis framework

Bernstein starts with the revised HBM stack-height roadmap and the technological constraints behind hybrid bonding, then separates demand into logic and HBM applications. It forecasts installed capacity, equipment shipments, market size, vendor market share, unit pricing and company revenue before linking the results to ratings and price targets.

Methodology notes

• Industry AnalysisSupply-demand framework

  Memory supply constraints, HBM stack configurations and capacity additions

  The report links current memory tightness and lower stack-height choices to delayed HBM hybrid-bonding adoption, then models adoption as supply conditions and stacking needs evolve.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Advanced-packaging technology adoption flowing into equipment demand

  Bernstein traces demand from chiplet and HBM packaging roadmaps through foundry and memory capacity needs to hybrid-bonder, dicing and grinding equipment demand.

• Valuation methodsP/E and PEG Valuation

  Forward EPS multiple valuation

  Price targets are derived using stated P/E multiples applied to forward EPS estimates, generally over forward Q5–Q8 periods.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Besi

  Hybrid-bonding equipment beneficiary despite delayed HBM adoption

  Strengths

  Logic hybrid bonding and conventional packaging remain demand drivers; Bernstein models c. $1.1bn of 2028 HB revenue.

  Weaknesses

  Near-term HBM hybrid-bonding adoption is delayed.

  Comparison

  Bernstein's 2028 shipment estimate is between Besi's mid and high cases.

  Risks

  Delayed adoption, a prolonged mainstream assembly-market recovery, hybrid-bonding share loss, or weak uptake of Besi's TCB offering.

• DISCO

  DRAM wafer-processing equipment beneficiary

  Strengths

  Strong DRAM wafer consumption supports dicing and grinding demand; NAND W2W bonding, BSPDN and silicon photonics are additional drivers.

  Risks

  AI GPU/HBM overcapacity, competition and unfavorable exchange-rate movements.

• Samsung Electronics

  Potential early HBM hybrid-bonding adopter

  Strengths

  Bernstein expects Samsung to lead HBM hybrid-bonding adoption, potentially starting with 16-hi HBM5 in 2028.

  Weaknesses

  HBM4E is expected largely to remain conventional TCB at 8/12-hi.

  Comparison

  Expected to lead adoption ahead of SK hynix and Micron.

  Risks

  Earlier end to favorable memory pricing, weaker demand or higher supply, valuation sentiment, and China's memory progress.

• SK hynix

  HBM supplier and potential 2028 hybrid-bonding adopter

  Strengths

  Its packaging roadmap highlights hybrid bonding's capacity, performance and thermal benefits.

  Weaknesses

  TCB is expected to remain mainstream through 2028.

  Comparison

  Expected to follow Samsung in HBM hybrid-bonding adoption.

  Risks

  Earlier end to favorable memory pricing, weaker demand or higher supply, valuation sentiment, and China's memory progress.

• Micron

  Memory supplier and potential 2028 hybrid-bonding adopter

  Strengths

  Bernstein expects adoption for HBM5 in 2028 and retains an Outperform rating.

  Weaknesses

  TCB remains the mainstream technology through 2028.

  Comparison

  Expected to follow Samsung in HBM hybrid-bonding adoption.

  Risks

  Earlier end to favorable memory pricing, weaker demand or higher supply, valuation sentiment, and China's memory progress.

• KIOXIA

  NAND memory supplier

  Strengths

  Potential upside from stronger NAND demand, Japanese capex subsidies and cost reduction.

  Weaknesses

  Bernstein retains Underperform because of the long-term threat from China.

  Risks

  China's progress in memory, particularly NAND; weaker demand or higher supply ending favorable pricing.

Key data

• 2028 hybrid-bonder TAMc. $1.4bnCombined logic and HBM hybrid-bonding equipment market forecast.
• 2028 logic hybrid-bonding TAM$887mnForecast to grow eightfold from 2025.
• 2028 HBM hybrid-bonding TAM$489mnBased on 15% blended penetration across 1,202kwpm of HBM capacity.
• 2028 hybrid-bonder shipments367 units24% above Besi's mid case and 8% below its high case.
• Logic installed base582 units in 2028 versus 137 in 2025Bernstein's global logic hybrid-bonder installed-base forecast.
• Besi 2028 hybrid-bonding revenuec. $1.1bnAssumes 76% overall market share, about 281 units and $3.75mn ASP per unit.
• HBM thermal resistance~35% lower at 20 stacksReported hybrid-bonding benefit versus the current solder/underfill approach.

Impact & implications

The report sees a timing reset rather than a structural break in hybrid-bonding demand. Logic adoption supplies most of the 2027 demand base, while higher HBM layer counts and I/O requirements support a later migration; this underpins Bernstein's constructive view on Besi, DISCO and selected memory suppliers.

Risks

• For Besi, hybrid-bonding adoption could be lower than expected or delayed, the mainstream assembly recovery could remain weak, or the company could lose share or see limited TCB uptake.
• For DISCO, AI-related GPU and HBM overcapacity could impair earnings and valuation; competition and exchange-rate movements are additional risks.
• For memory suppliers, weaker demand or higher supply could end favorable pricing earlier than expected, while investor sentiment and China's memory progress—especially in NAND—are risks.
• KIOXIA also faces risk from weaker NAND demand, although the report cites stronger NAND demand, Japanese capex subsidies and cost reduction as upside factors.

What to watch

• Whether HBM4 and HBM4E remain concentrated at 8-hi/12-hi designs or move toward 16-hi stacks.
• The timing of JEDEC's possible relaxation of HBM package-height specifications from 775μm to about 900μm.
• TSMC's SoIC and hybrid-bonding capacity expansion, including adoption by AMD, Apple, Google TPU, Groq and Nvidia.
• The timing of Samsung, SK hynix and Micron hybrid-bonding adoption for HBM5 in 2028.
• Memory supply tightness in 2027, AI demand in 2028 and the resulting path for memory margins and pricing.