SK hynix wins TSMC Partner of the Year again after validating HBM5 on CoWoS packaging

SK hynix said on September 28, 2026 that it was named TSMC's Partner of the Year in the 3DFabric-HBM Integration category for the second consecutive year, an honor that recognizes its joint validation of HBM5, the eighth generation of high bandwidth memory, on TSMC's CoWoS 2.5-dimensional advanced packaging technology. The recognition was tied to the TSMC Open Innovation Platform (OIP) Conference held in Santa Clara, California, on September 23, where the company also displayed a 12-high 36GB HBM4 device and a 96GB SOCAMM2 module mounted inside Nvidia's next-generation Vera Rubin platform.

The award lands in a part of the semiconductor industry where memory and logic are increasingly designed together rather than in isolation. With every new HBM generation, the number of stacking layers rises and data processing speeds increase, which means the advanced packaging that connects the memory to the GPU on a single substrate has to advance in step. SK hynix and TSMC have been jointly validating packaging compatibility and performance since the early stages of HBM5 development, and SK hynix has been participating in the related TSMC forums since 2024.

HBM5 sits three generations ahead of HBM3E, the fifth-generation HBM widely used in AI servers today. SK hynix began shipping HBM4, the sixth generation, in the second quarter of 2026, and it has already supplied samples of HBM4E, the seventh generation, to major customers. HBM5 remains in the validation stage, where the objective is to verify its performance together with TSMC's advanced packaging.

CoWoS, short for chip on wafer on substrate, is TSMC's 2.5-dimensional packaging technology that integrates HBM and the GPU on a single substrate. The exhibition at the Santa Clara conference ran under the theme 'Memory Supporting Large-Scale Intelligence' and covered a lineup stretching from HBM to server DRAM and enterprise SSDs.

Key Facts

ChosunBiz reported on September 28 that SK hynix was named partner of the year for the second straight year after being recognized for collaborative results in validating HBM5 with TSMC, the world's largest foundry. Edaily reported on September 28 that the award came in the 3D Fabric-HBM Integration category for validating HBM5 based on CoWoS, following the company's win last year as well.

The product display was unusually specific. Alongside Nvidia's Vera Rubin, SK hynix showed the HBM4 36-gigabyte 12-high part and the 96GB SOCAMM2. HBM4 is mounted around the Rubin GPU, while SOCAMM2, a low-power DRAM-based AI server memory module, serves as main system memory for the Vera CPU. The booth also featured HBM4E 48GB 12-high, HBM4 48GB 16-high, HBM3E 36GB 12-high and 192GB SOCAMM2, plus server DRAM products including 256GB 3DS RDIMM, 128GB MRDIMM and 64GB RDIMM built on the 10nm-class sixth-generation (1c) process.

Park Dae-sik, a senior engineer in charge of PCIe design at SK hynix, projected that the PCIe-based eSSD market will grow about 3.8 times, from 509 exabytes in 2026 to 1,933 exabytes in 2030.

On the supplier side, Seoul Economic Daily reported on September 29 that SK hynix will prepay up to 50 percent of its suppliers' initial research and development costs. President and CEO Kwak Noh-jung and other top executives met representatives from about 100 suppliers at the company's 2026 Partners Day, an event that marked the 25th anniversary of the launch of its Win-Win Growth Council. The program is tailored to four stages of a supplier's growth: technology development, testing and verification, mass production, and sales. At the development stage, a newly created R&D Challenge Reward program will cover up to 50 percent of initial development costs, and even if development fails or does not result in a purchase, SK hynix will settle the costs based on technical results and contribution.

ChosunBiz reported on September 29 that the company decided to double its delivery payment support fund to 100 billion won, about 68 million US dollars, from 50 billion won. It is also accelerating plans to build Trinity Fab, a process demonstration facility of roughly 3,300 square meters (1,000 pyeong) at the Yongin semiconductor cluster, targeting a launch next year, and it will provide low-interest funding through its win-win growth fund while extending consulting to second- and third-tier suppliers. The Win-Win Growth Council was launched in 2001 with 79 companies, and this year awards went to 11 outstanding partners across four categories: technological innovation, quality improvement, co-prosperity, and ESG.

Analysis

The most striking detail in the SK hynix disclosure is not the trophy but the stage of the silicon behind it. HBM5 is the eighth generation of high bandwidth memory, three generations beyond HBM3E, which is still the mainstream part in AI servers, and the company is validating it with TSMC's CoWoS before the product is anywhere near mass production. What this really means is that the competitive frontier in AI memory has shifted from stacking DRAM dies higher to proving, jointly with a foundry, that a memory stack and a GPU can be packaged together and still hit their performance targets. That work cannot be done by a memory maker alone, and it cannot be done by a foundry alone.

Seen in that light, a second consecutive Partner of the Year award in the 3DFabric-HBM Integration category is less a marketing prize than evidence of a working engineering relationship. ChosunBiz reported on September 28 that HBM5 is still in the validation stage to verify performance with TSMC's advanced packaging, which is exactly the kind of quiet, schedule-defining work that determines whether a next-generation AI accelerator ships on time. Every generation raises stacking layers and data rates, so the packaging roadmap and the memory roadmap have to be interlocked years ahead of revenue.

There is also a portfolio argument in the booth display. SK hynix did not show a single hero part. It showed 12-high 36GB HBM4, 16-high 48GB HBM4, 12-high 48GB HBM4E, 12-high 36GB HBM3E and two SOCAMM2 capacities, 96GB and 192GB, plus three classes of server DRAM. That breadth suggests the company is positioning itself as a full memory supplier for AI systems rather than a specialist in one die stack, and it is using the TSMC platform, with Nvidia's Vera Rubin as the showcase, to make that case to the customers who actually design the racks.

The Partners Day announcements point in the same direction from the other end of the supply chain. A supplier base that is asked to develop materials, parts and equipment for future HBM generations carries real risk, and the R&D Challenge Reward program absorbs up to half of the initial development cost even when a project fails. The bigger picture here is that the memory maker is buying development capacity and loyalty from roughly 100 suppliers at a moment when the packaging and process demands of HBM5 and beyond are tightening.

Why It Matters

For Nvidia, the practical consequence is supply. The Vera Rubin platform depends on HBM4 around the Rubin GPU and on SOCAMM2 as main memory for the Vera CPU, and both were shown running with SK hynix parts. Any slip in memory validation or packaging capacity flows straight into accelerator shipments, which is why the CoWoS line and the HBM roadmap are watched as one system rather than two product categories.

For TSMC, the award is a signal about how its advanced packaging franchise is being used. CoWoS was created to integrate HBM and logic on a single substrate, and the 3DFabric-HBM Integration category exists precisely because that integration is now a named engineering discipline. A partner that validates eighth-generation memory on the packaging technology gives the foundry a concrete example to offer other customers weighing the same architecture.

For the wider memory market, the display of sixth, seventh and eighth generation parts in one booth is a reminder of how fast the stack is turning over. HBM3E is still the workhorse, HBM4 only began shipping in the second quarter of 2026, HBM4E exists as samples, and HBM5 is in validation. Enterprise storage is moving at a similar pace: the PCIe eSSD market is projected to grow about 3.8 times, from 509 exabytes in 2026 to 1,933 exabytes in 2030, which is roughly the volume of data that AI training and inference clusters will need to move, store and re-read.

Next Up

The next visible milestones are already scheduled. HBM5 must move from validation on CoWoS toward qualification with customers, HBM4E samples must convert into orders, and Trinity Fab, the roughly 3,300 square meter process demonstration facility in the Yongin cluster, is targeted to launch next year so that suppliers can verify yields and reliability on SK hynix production lines.

On the corporate calendar, the delivery payment support fund doubles to 100 billion won, the R&D Challenge Reward covers up to 50 percent of initial development costs, and the win-win growth fund is to provide low-interest financing as consulting reaches second- and third-tier suppliers. Kwak Noh-jung closed the partners meeting with a line that frames the next phase as a shared one: 'Building on the trust we have accumulated so far, let us pioneer the future together and open a new 25 years of even greater leaps forward.'