The AI memory shortage is creating an awkward problem for BE Semiconductor Industries, or Besi: customers urgently need more high-bandwidth memory, but that urgency can make them less willing to change how it is assembled. Investors who expected hybrid bonding to move rapidly into HBM are now reassessing when technical interest will become equipment orders for volume production.
The pressure concerns timing and valuation, not a finding that direct copper bonding has stopped working. Existing assembly methods are improving, a revised package-height allowance gives them more room, and manufacturers with scarce output have strong reasons to protect yields and delivery schedules. Strong demand for the finished memory does not automatically mean an immediate transition to Besi’s flagship process.
The downgrade challenges a revenue timetable
Besi shares fell 5.9% to €180.75 on October 7, 2026, after UBS moved its rating from Buy to Sell and its target from €370 to €159. Bank of America Global Research had separately shifted Buy to Neutral with a €212 target. Barclays also cut its target while retaining Equalweight. These are dated analyst actions, not current prices or an investment recommendation.
UBS estimated that Besi’s valuation implied about €936 million of hybrid-bonding revenue in 2030, against its own €424 million forecast. It also put HBM at roughly 10% of hybrid-bonding tool demand in 2028, versus consensus expectations it described as nearer 50%. Those are the bank’s estimates of revenues, demand mix and expectations—not confirmed customer commitments, disclosed orders or Besi guidance.
The gap explains the sensitivity to adoption news. Revenue shifted further into the future may be worth less in today’s valuation even if the eventual technology opportunity remains intact. UBS could describe Besi as a high-quality semiconductor-equipment company while disputing how much growth the shares already assumed.
Why a shortage can favor the older process
HBM stacks several DRAM dies vertically beside a processor or accelerator to supply bandwidth for AI workloads. Traditional assembly connects the dies through microscopic solder bumps. Thermocompression uses controlled heat and pressure; mass reflow melts many connections during a manufacturing step. Hybrid bonding instead aligns prepared copper pads and surrounding insulating surfaces and joins them directly without the solder layer.
Direct connections permit finer pitches, denser interconnects and thinner stacks, with potential electrical advantages. Those characteristics become more valuable as stacks gain layers and designers manage power and heat. But a factory also needs throughput, contamination control, manageable defects, integration economics and yield—the share of devices meeting specifications. Each extra interface in a stack creates another possible alignment or bonding failure involving already valuable dies.
In UBS’s assessment, HBM supply could stay tight through 2028. Under such conditions, a qualified line can sell nearly all its output, while a less mature process risks lost units or disrupted deliveries. Capital spending can therefore rise while funding more proven capacity rather than a change in the core assembly route. Memory producers benefit immediately from selling additional output; Besi benefits when they install its equipment. The two cycles need not move together.
JEDEC’s increase in permitted HBM package height from 720 to 775 microns adds another reason to wait. The extra space gives 16-layer HBM4 a path using microbumps rather than requiring immediate solder removal. The September roadmap analysis describes an extension of older methods, not a loss of hybrid bonding’s capability. It postpones the point where height becomes an unavoidable constraint.
Thermocompression equipment, mass-reflow processes, molding materials and stack designs also continue improving. Hybrid bonding competes with a moving alternative, not a frozen technology. Manufacturers can revisit it when stack height, connection pitch or thermal requirements outweigh transition risk. That does not mean they have abandoned evaluations or advanced-packaging facilities, or that every near-term HBM product needs direct bonding.
Orders show engagement, but evaluation is not a factory ramp
Besi’s first-quarter 2026 results said hybrid-bonding unit orders more than doubled from the fourth quarter of 2025, setting quarterly records in both units and value. Most of the increase came from one unnamed customer’s larger-than-expected capacity build. That is tangible demand, but concentration can make quarterly comparisons uneven rather than establish a smooth industry-wide ramp.
The company reported repeat HBM orders from one memory customer and delivery of two evaluation systems to a second. Overall hybrid-bonding adoption reached 20 customers. Research equipment, evaluation machines and production lines serve different purposes: the count does not establish 20 high-volume HBM users, the size of each installation or when tests turn into repeat production purchases.
Besi connected emerging logic, memory, co-packaged-optics and consumer applications with AI-related introductions expected during 2027–2030. Its guidance for second-quarter revenue growth of 30%–40% over the first quarter, based on backlog and customer feedback, covered the wider business. It was not an HBM-only or hybrid-bonding-only growth forecast and cannot resolve the adoption timing dispute.
Commercial logic products nevertheless provide evidence beyond laboratory tests. AMD has used direct bonding with vertically stacked cache, TSMC has deployed SoIC, and Intel uses direct bonding within Foveros. Those examples show that architecture and economics can justify production. They do not automatically establish the throughput or cumulative yield economics of repeated-die HBM stacks.
Besi has a bridge technology and an integrated approach
Besi’s TC Next thermocompression platform targets next-generation AI devices. In the first quarter it reported two new orders and adoption across six customers. This gives it exposure to a bridge process customers can use before hybrid bonding becomes necessary. It weakens the idea that Besi has only one route into AI packaging, without erasing concerns about the growth and product mix previously attached to direct bonding.
With Applied Materials, Besi is also developing Kinex, an integrated die-to-wafer hybrid-bonding system combining surface preparation, bonding and process control. Coordinating those steps can reduce contamination risk and movement between separate operations. It addresses manufacturing barriers, but equipment readiness and a customer’s need to change products remain different variables.
Counterpoint’s analysis of a memory adoption path involving SK hynix, Applied Materials and Besi emphasizes continuing co-development rather than abandonment. Better integration can make future production more viable; it cannot force a manufacturer to switch before the complete economics justify it. Supplier revenue needs do not determine a memory factory’s optimal transition date.
A later market gives rivals more qualification time
ASMPT has developed hybrid-bonding systems, including memory-oriented tools. The broader field includes Hanwha Semitech, EV Group, Applied Materials and other suppliers and partnerships, with different capabilities and roles. A delayed opportunity gives rivals more time to qualify equipment and build customer relationships. Besi can participate in a successful technology while capturing less revenue or share than early expectations implied.
Bank of America also pointed to possible ASML involvement. Its extent and form remain unclear; this is not a confirmed die-bonding entry or product launch. ASML is known for lithography, though its process-control capabilities and industry relationships would make any strategic move consequential to watch. Speculation should not be converted into an established competitive win.
Qualifications can begin long before large orders. Validating several vendors gives customers bargaining power and reduces dependence, so Besi’s early expertise in precise die placement does not guarantee eventual share. Conversely, repeat orders for its platform with Applied Materials would support the view that an integrated process is becoming embedded. Both outcomes require evidence rather than assumptions about an inevitable winner.
Beyond HBM, the issue is still scale
Dense vertical links can serve logic, chiplets, image sensors, consumer processors and co-packaged optics. The latter places optical communication components near computing silicon to reduce distance and power in moving data between processors and networks. It is a potential additional market, but gradual development should not be described as an immediate replacement for delayed memory revenue.
HBM attracted especially large expectations because every stack requires multiple bonding steps. Several smaller applications can support growth without equaling that volume. The bullish case rests on physical limits to solder bumps and eventual scaling needs; the bearish case rests on how long existing methods can keep serving customers and what valuation assumes. Neither a downgrade nor a delayed generation is a technical verdict.
The next scheduled reporting point is Besi’s third-quarter results on October 22, 2026. Useful distinctions include repeat hybrid-bonding purchases versus evaluation deliveries, named production applications and memory capacity commitments. A future shift in adoption language from HBM4E toward HBM5 would imply a longer wait, but should not be reported as a shift already confirmed.
SK hynix, Samsung and Micron control the relevant memory production decisions. Confirmed use in a shipping stack, qualification milestones and line-specific capital spending would be stronger signals than demonstration samples. Continued mass reflow or thermocompression in HBM4E could support the delay thesis; one supplier’s high-volume direct-bonding commitment could put competitive pressure on others. Orders without a clear application, products without equipment details or technical wins with limited yields remain incomplete evidence.
For accelerator buyers and infrastructure teams, the issue reaches beyond one equipment share: packaging constraints affect availability, system costs and deployment speed. Yet an industry tailwind does not specify a supplier’s revenue date. Besi’s predicament illustrates how scarcity can slow a technically attractive transition precisely because dependable existing output has become so valuable. The decisive change is when direct bonding becomes economically necessary, rather than simply promising.
Sources
Analyst downgrade reporting · HBM height and bonding roadmap · Besi Q1 2026 results · Counterpoint memory co-development analysis · Bloomberg report