📖 Reading time: approx. 6 minutes · 1,107 words · 6,990 characters
Samsung has recently had an uncomfortable problem in the AI hardware boom: it is big everywhere, but not leading everywhere where the highest margins are currently being made. In HBM, SK hynix dominated NVIDIA’s supply chain for a long time, and in the foundry business Samsung continues to face pressure from TSMC. That makes the statement by Jun Young-hyun, co-CEO of Samsung Electronics and head of the chip division, all the more interesting: he said that on 8 June 2026 in Seoul he spoke with NVIDIA CEO Jensen Huang about cooperation on next-generation foundry generations. Reuters reports that Samsung and NVIDIA are already working together on autonomous driving chips and Groq AI accelerators. In addition, both sides reportedly discussed long-term cooperation on HBM4E and HBM5. That is not a minor footnote. It is precisely these memory and manufacturing topics that determine whether Samsung will move closer to NVIDIA again in the AI hardware market – or continue to miss the most lucrative bottlenecks.
Symbolbild
The Reuters report is relatively brief, but substantively dense. Jun Young-hyun accordingly said that he had spoken with Huang about cooperation on next-generation foundry chips. The discussion also concerns future semiconductor products and long-term collaboration on HBM4E and HBM5. This is an official statement by a senior Samsung executive to the press, not merely industry hearsay. What is not confirmed, however, is that NVIDIA will have a major next-generation GPU main chip manufactured at Samsung. That would be a clearly different statement. Based on the available information, the topic so far concerns discussions, existing cooperation, and future possibilities. This distinction is particularly important in the foundry business, because even small wordings are quickly overinterpreted as “NVIDIA is switching from TSMC to Samsung.” There is no evidence for that in the sources available here. What is concretely confirmed, however, is cooperation around Groq. Reuters points out that Huang presented a new NVIDIA AI inference processor based on Groq technology in March 2026 and that Samsung is to manufacture the Groq LP30, with shipment planned for the second half of 2026. For Samsung, this is a visible foundry foothold in an NVIDIA-adjacent platform, even if it is not equivalent to a classic NVIDIA GPU die. Samsung prominently displayed its HBM roadmap at GTC 2026. In its own Semiconductor newsroom, Samsung describes the presentation of HBM4E and HBM5 architecture as part of a broader “Total AI Solution” that includes memory, foundry, and packaging. Samsung cites a speed of up to 13 Gbit/s per pin for HBM4 and refers to hybrid copper bonding packaging with improved thermal behavior. For HBM5, Samsung describes a technical direction: the core die is to be based on a sixth-generation 10-nm-class DRAM process, while the base die is to be manufactured with a 2-nm foundry process. This is where the strategy becomes visible. Samsung does not want to supply memory chips alone, but to sell the combination of DRAM, logic base die, packaging, and foundry as an integrated package. That is a decisive difference. HBM is long no longer a simple stacked memory. The more closely GPU, memory, interposer, base die, and packaging interact, the more tightly manufacturing and memory development become linked.
For Samsung, this is an opportunity to leverage its breadth: memory, foundry, and advanced packaging come from one conglomerate. For customers such as NVIDIA, that can be attractive, provided performance, yield, power consumption, and supply capability are right. The obvious reading is: Samsung wants to compete more strongly with SK hynix at NVIDIA again. That is true, but not complete. In HBM, the relevant players are indeed SK hynix, Samsung, and Micron. But on the foundry side, Samsung is primarily up against TSMC – and increasingly against Intel Foundry as a potential challenger in the high-end and packaging business. The talks with NVIDIA are therefore doubly important. An HBM order strengthens Samsung’s memory position. A foundry or packaging collaboration, by contrast, could build confidence in Samsung’s advanced logic manufacturing. It is precisely this confidence that is difficult to gain and easy to lose in the foundry business. Customers have to align their designs early with process libraries, PDKs, yield expectations, and packaging options. A single statement from Seoul does not replace production approval, but it shows that Samsung at least is once again directly at the table. Reuters also places the talks in the broader South Korea visit by Huang. While SK hynix was highlighted as NVIDIA’s largest memory partner, Samsung is clearly trying to differentiate itself through the breadth of its portfolio. Not just HBM, not just foundry, not just packaging – but all of it together. That is ambitious. But it is also necessary if Samsung does not want to remain merely a second-choice supplier. It remains unclear which specific NVIDIA products could result from the talks. It is also open whether HBM4E or HBM5 will actually come from Samsung in significant volume in future NVIDIA platforms. There are also no hard details on the foundry side: no node, no production timeframe, no volume, no specific chip class. Anyone deriving a switch from TSMC to Samsung from this is reading more into it than the sources support. What can be stated reliably, however, is that Samsung is once again positioning itself aggressively as a complete AI provider. HBM4E, HBM5, Groq LP30, foundry talks, and packaging are not being communicated simultaneously by coincidence. This is an attempt to narrow the gap to SK hynix in HBM and to TSMC in foundry, at least in narrative terms – and ideally later also through orders.
Conclusion:
Samsung needs NVIDIA not only as a customer, but as a seal of quality. If NVIDIA trusts a memory or foundry partner, the market pays attention. The talks that have now become public are therefore no guarantee of large orders, but they are an important signal. Samsung is trying to translate its old strength – breadth from memory to manufacturing – into the AI cycle. Whether that succeeds will not be decided on a stage, but in yield data, delivery schedules, and thermal limits. Less glamorous, but exactly there this market is won.
Noch keine sichtbaren Antworten im Forum gefunden. Der Thread ist bereits angelegt und kann direkt im Forum geöffnet werden.