Hybrid bonding: Has the inflection point arrived?

In the past two years, HBM has almost become a key term that the entire industry cannot avoid.

On top of the booming AI accelerators, apart from the core GPU, the most eye - catching components are often the HBM chips closely attached to it.

Most of them are stacked through thermocompression bonding. Micron - scale metal bumps are arranged between DRAM chips, and the connection and formation are completed through heating and pressurization.

This process has been used for nearly a decade. It is mature and stable, and the yield rate is within a controllable range. However, the real bottleneck lies in the height: the standard organization JEDEC has set the upper limit of the packaging height at 775 microns. Within this space, the base chip and 16 layers of core chips need to be accommodated simultaneously. The existing thermocompression bonding process seems unable to meet the requirements.

At this time, hybrid bonding has come into the spotlight. What it does is to completely eliminate the bumps, allowing two copper surfaces to directly contact and complete the connection at the atomic level. This brings lower connection resistance, a more compact layer spacing, a shorter signal path, and better heat dissipation performance.

This approach has already been verified in image sensors and TSMC's 3D V - Cache. However, when the target becomes HBM, the difficulty has multiplied. The reason is not hard to understand. It is no longer the combination of two layers of chips, but a high - stack structure of 16 or even 20 layers of DRAM. Any mistake in one layer means the scrapping of the entire chip. The industry generally estimates that the yield rate of HBM hybrid bonding currently hovers around 10%, and to truly move towards commercialization, it needs to cross the 60% threshold at least.

Even so, Samsung and SK Hynix have both taken the first step in mass - production verification this year.

Two Companies, Two Strategies

SK Hynix is the current leader in the HBM market, with a global market share of over 70%. HBM revenue accounts for over 40% of its total DRAM sales. Its production capacity for 2026 has already been locked in advance. Its early adoption of hybrid bonding is more like a defensive move, holding the next - generation technology chips in hand in advance to avoid being preempted by Samsung or Micron in the HBM race.

According to South Korean media reports, SK Hynix placed an order for a hybrid bonding in - line equipment jointly developed by Applied Materials and Besi last month, with a price of about 20 billion won. This is the company's first purchase of a mass - production hybrid bonding system. This equipment integrates Applied Materials' CMP and plasma processing processes directly into Besi's bonding machine. It has been implemented on TSMC's production line for the mass production of AMD 3D V - Cache.

Interestingly, SK Hynix is also planning to introduce equipment from Hanwha Semiconductor during the same period and plans to start large - scale process verification from April to May this year. Betting on different suppliers makes it seem more at ease.

Samsung seems to be in a greater hurry.

According to reports, Samsung has chosen to directly promote hybrid bonding on HBM4 instead of following the MR - MUF route led by SK Hynix. There is a bit of helplessness behind this. After all, it has lagged behind in HBM packaging for nearly two years. If it continues to follow, it will only catch up with yesterday's technology and miss tomorrow's market.

In March this year, Samsung specifically released a video on hybrid bonding technology at NVIDIA GTC 2026, claiming that the thermal resistance can be reduced by more than 20% compared with traditional TC bonding. It also published a paper at ISSCC, showing the process of manufacturing 4F² DRAM using inter - wafer hybrid bonding technology, which is quite proactive in the usually low - key technology disclosure. Samsung CTO Song Jae - hyuk also emphasized the advancement of the technology in his keynote speech at the 2026 South Korean Semiconductor Exhibition. He said that when 12 - layer and 16 - layer HBM use hybrid bonding, the temperature of the base chip can be reduced by more than 11%.

However, Samsung's subsidiary Semes is a variable. Samsung is evaluating Semes' chip - to - wafer hybrid bonding machine. The logic of internalized production is clear: to reduce dependence on external suppliers and control the core manufacturing process. But the problem is that the overall quality of Semes' hybrid bonding equipment currently lags behind Besi. Before this path is successful, Samsung still has to bet on external equipment at the same time.

Backstab from the Standard Organization

While the two major manufacturers are making intensive moves, the standard organization has stabbed them in the back.

According to reports, JEDEC is discussing relaxing the HBM height standard from 775 microns to about 900 microns. The new standard is expected to take effect from the seventh - generation HBM4E. This means that it is feasible to stack more layers on thicker chips using existing TC bonding equipment. Hybrid bonding, which was necessary when reaching 20 layers, can now wait for a few more years.

It is worth noting that both NVIDIA and Amazon AWS are reported to plan to adopt TSMC's SoIC advanced packaging solution. SoIC vertically stacks system semiconductors, which naturally increases the overall packaging thickness. This leaves more breathing space for the entire GPU ecosystem in terms of HBM height specifications.

In addition, the pad pitch of HBM4 happens to be around the 10 - micron critical point. At this pitch, the cost - competitiveness of hybrid bonding is difficult to surpass that of the micro - bump solution. Even without standard adjustment for this generation of products, there may not be enough commercial reasons to force the implementation of hybrid bonding.

The change in the standard has put Besi in the Netherlands at the forefront of the storm.

The Equipment Leader: Soaring and Plummeting

Besi is the most prominent beneficiary of this hybrid bonding boom and also the most exposed target.

According to the prediction of the Financial Times, Besi's hybrid bonding business revenue is expected to soar from 36 million euros in 2023 to 476 million euros in 2026, accounting for about one - third of the company's total revenue at that time. This growth prediction is an important pillar of Besi's high valuation. So when the news from JEDEC came out, Besi's stock price fell by more than 19% during intraday trading on that day. Before that, this stock had just risen by 58% in the past year and more than 200% in five years.

But in less than two weeks, Besi rose by more than 10% due to another event. According to Reuters, Besi has received acquisition inquiries from multiple external parties, and Lam Research is one of them. Applied Materials, which already holds about 9% of the shares, is also considered one of the potential buyers. For the same company in the same period, it fell once because hybrid bonding might be delayed and rose once because hybrid bonding is so important that someone wants to buy it, which is quite ironic.

It is worth noting that Michael Roeg, an analyst at Degroof Petercam, pointed out that JEDEC has previously relaxed the HBM stacking height specifications. This discussion is more likely to change the adoption curve of hybrid bonding rather than completely end it. UBS believes that the performance and heat - dissipation advantages will still make hybrid bonding an unavoidable option when the number of layers is higher.

From Besi's own perspective, it is not completely dependent on hybrid bonding equipment. It also has a place in the TC bonding and high - precision chip - mounting equipment markets, but its growth rate is bound to slow down for now.

Seeking Their Own Ways Out?

The ambiguity of HBM standards has made every equipment manufacturer start to rethink its development route.

Hanmi Semiconductor is currently the undisputed leader in the TC bonding market. TechInsights data shows that its market share in the global HBM TC bonding machine market is as high as 71.2%. SK Hynix is almost its exclusive major customer, and Micron has also started to place orders later.

The standard change of JEDEC is not bad news for Hanmi. It means that the life cycle of its existing business has been extended. At the Semicon Korea 2026 exhibition in February this year, Hanmi Semiconductor also displayed a new wide - width TC bonding machine specially designed for HBM5 and HBM6. It pointed out that due to the delay in the commercialization of hybrid bonding, this equipment is expected to fill the market gap.

But Hanmi is also investing more than 50 billion won in its Incheon No. 7 factory to specifically build a hybrid bonding production line. It plans to start production in the first half of next year. The hybrid bonding machine for HBM is expected to be launched in 2027, and the version for SoC will be postponed to 2028. Its strategy seems clear: TC bonding is responsible for the present, and hybrid bonding will be fully promoted when the track really heats up.

The situation of another South Korean local equipment manufacturer, Hanwha Semitech, is much more delicate. This company has made great efforts to cooperate with the Dutch precision manufacturer Prodrive to develop the second - generation hybrid bonding machine SHB2 Nano, with an accuracy of ±100 nanometers, which meets the accuracy index of Besi's flagship product. The equipment has been assembled in the Netherlands and shipped back to South Korea, and is planned to be delivered to SK Hynix for testing.

The problem it faces is that the cost structure of cooperative development is higher than that of independent R & D, and the market demand has become more uncertain due to the standard discussion. Some industry insiders point out that the R & D speed of SHB2 Nano is faster than Prodrive's normal rhythm, and the long - term stability of its fineness still needs to be verified in mass production. More straightforwardly, Hanwha Semiconductor has placed a large number of chips on a track with an unclear schedule.

Applied Materials (AMAT) has a different logic. It bought about 9% of Besi's shares early and became its largest single shareholder. At the same time, they jointly developed the Kynex D2W hybrid bonding in - line system. This solution, which integrates CMP, plasma processing, and bonding processes, has been in mass - production operation on TSMC's production line. AMAT's idea is to extend its front - end process capabilities accumulated over the years to the back - end, embed front - end strengths such as CMP into the packaging process, and build a complete platform that is difficult for competitors to replicate independently. At the same time, with the help of Besi's equipment network and customer relationships, it can directly enter the supply chain vision of SK Hynix and Samsung. In fact, in SK Hynix's first large - scale hybrid bonding purchase, the jointly developed equipment was the winning bidder.

Another equipment leader, K & S (Kulicke & Soffa), has a more balanced layout. This veteran manufacturer, which has been doing wire bonding for more than 70 years, announced in March 2026 the expansion of its full - line product portfolio of interconnect solutions for storage. It is expected that its TCB business will increase by about 70% quarter - on - quarter in the 2026 fiscal year, and it has accelerated the R & D project of hybrid bonding, which has attracted the interest of early - stage customers. It can be said that K & S tends to focus on TCB first and at the same time promote the R & D of hybrid bonding in parallel. It will make targeted investments when the market is truly mature.

ASML, Uninvited

In the game of hybrid bonding, the most unexpected player is actually a company that has never made packaging equipment.

According to South Korean media The Elec, ASML has started to design the overall architecture of hybrid bonding equipment and has cooperated with long - term partners Prodrive Technologies and VDL - ETG. They are respectively responsible for the linear motor drive of the EUV magnetic levitation system and the manufacturing of the mechanical structure. These two companies are also the core partners of Hanwha Semiconductor's SHB2 Nano. If ASML really enters the market, to some extent, it will cut off the construction path of latecomers from the component supply chain level.

ASML's CTO Marco Peters publicly stated that after seeing the technology roadmaps of storage manufacturers such as SK Hynix, he confirmed that the demand for stacking process equipment is real, and he is personally examining the product portfolio direction of the company in the semiconductor packaging field. But ASML's CEO also admitted that the probability of hybrid - bonding - related business having a significant impact on the company's revenue before 2030 is low. What ASML needs to do is to complete technology reserve and ecological positioning before the track matures.

After careful consideration, in fact, ASML's cross - border positioning is not low in technical feasibility. Currently, ASML's High - NA EUV system has an overlay accuracy of 0.7 nanometers. The alignment control required for hybrid bonding is within the order of magnitude that it is already familiar with. Once it really enters the market, its technical starting point will not be the initial stage of other packaging equipment manufacturers, but will directly bring decades of experience in nanometer - level precision control. Some industry insiders even evaluate that once ASML's hybrid bonding machine is launched, it will immediately change the existing competitive landscape.

Implementation Is the Real Problem

All these layouts, bets, entries, and wait - and - sees are ultimately waiting for the same answer: when can hybrid bonding really run on the HBM mass - production line?

From the existing roadmap, the answer generally points to HBM5, that is, the generation with 20 - layer stacking. The three major storage manufacturers have basically confirmed that they will introduce hybrid bonding on a large scale in this generation. According to the general industry expectation, the time window for HBM5 falls around 2028 - 2030. Before that, TC bonding will continue to evolve through processes such as flux - less bonding. The standard game of JEDEC around the packaging height is essentially to gain more time and space for this transition period.

However, the biggest uncertainty in this seemingly clear roadmap comes from NVIDIA. Samsung and SK Hynix have started to send HBM3E samples based on hybrid bonding to NVIDIA for testing. Once NVIDIA puts forward leap - forward requirements for bandwidth, power consumption, or latency on the next - generation Rubin or even more subsequent platforms, the rhythm of the entire industry may be forced to be rewritten in advance.

Yield rate, standards, and NVIDIA's demand - any deviation in these variables may further postpone the mass - production time of hybrid bonding.

For equipment manufacturers, this is not a choice of whether to do it or not. The direction of hybrid bonding is beyond doubt, and the only difference lies in the rhythm. Some are waiting and seeing, some are accelerating, and some are positioning in advance. As for the standard adjustment of JEDEC, it is just a short - term supply in this long - distance race.

This article is from the WeChat official account “Semiconductor Industry Observation” (ID: icbank), author: Shao Yiqi, published by 36Kr with authorization.