Samsung Electronics' foundry business is showing signs of recovery after a prolonged downturn, as AI chip projects and HBM4-related demand begin to raise utilization at its advanced-process lines.
Ddaily reported, citing industry sources, that Samsung Foundry has recently secured a series of 4nm projects and is preparing to expand supply in the second half of 2026. Major AI chip programs are expected to enter mass production during that period, while base dies for sixth-generation high bandwidth memory, or HBM4, are also expected to add volume.
Samsung Foundry could return to profit as early as the second quarter of 2026, after an estimated operating loss of about KRW1 trillion in the first quarter of 2026, according to Korean media reports and analyst estimates. Utilization has recently climbed to the 80% range, supported by the production of Exynos 2600 and HBM4 base dies.
The improving outlook comes as Samsung's memory business is also benefiting from stronger demand. Together, the memory turnaround and a potential foundry recovery could support Samsung's semiconductor earnings in the second half of 2026.
4nm demand builds around AI chips
Samsung's 4nm process has become a key node in the company's effort to regain momentum in advanced foundry. The company began mass production of its first 4nm process, SF4E, in late 2021, followed by second-generation SF4 and third-generation SF4P. Its fourth-generation high-performance computing process, SF4X, launched in 2025 and has gained a firmer footing in 2026.
One of the most visible SF4X customers is Groq, a US AI semiconductor company that develops language processing units, or LPUs, designed to reduce token-generation latency during large language model inference.
Samsung's SF4X process is expected to be used for the mass production of Groq's next-generation LPU chips for rack-scale inference accelerators in the second half of 2026. The project has become an important reference case for Samsung's 4nm recovery, particularly after Nvidia CEO Jensen Huang highlighted Samsung's role in producing the Groq-related chip at GTC 2026 in March.
Samsung also displayed a 4nm wafer for the chip at the event. Reuters separately reported that Nvidia's new AI inference processors, based on technology from Groq and made using Samsung's 4nm process, are expected to ship in the second half of 2026.
Other companies — including IBM and Ambarella in the US, Baidu in China, and South Korea's Rebellions and HyperAccel — are also expected to use Samsung's 4nm process, including SF4X.
HBM4 turns memory demand into foundry volume
HBM4 is emerging as another driver for Samsung Foundry because its base die is produced using a 4nm logic process.
In earlier generations of HBM, the base die mainly acted as an interface layer beneath stacked DRAM core dies, connecting the memory stack to external chips and supporting data movement and power delivery. Performance competition focused mostly on DRAM core-die density and bandwidth.
That role is changing as more data moves between AI accelerators and HBM. Signal control, power management, and input-output efficiency have become more important, pushing HBM4 base dies further into advanced logic manufacturing.
That shift creates a natural link between Samsung's memory recovery and its foundry utilization. As HBM4 shipments to major customers such as Nvidia and AMD expand, Samsung's 4nm base-die output is expected to rise, helping lift utilization at advanced lines, including those at its Pyeongtaek campus.
The outlook is also being reflected in earnings expectations. Kiwoom Securities analyst Park Yu-ak said Samsung Foundry, including System LSI, could post a small operating profit in the second quarter of 2026, supported by HBM4 base-die and Exynos 2600 production. Park revised his expected timing for a return to profit forward from the third quarter to the second quarter.
Kang Seok-chae, executive vice president of Samsung's foundry business, said during the company's first-quarter 2026 earnings call on April 30 that advanced-process line utilization had reached its highest level. Demand for advanced products, including HBM4 base dies, is expected to improve second-quarter results, he said.
2nm becomes the next test
Samsung's next-generation 2nm process is also drawing attention as the company tries to turn early recovery signs into a more durable foundry rebound.
The company has moved into mass production of Exynos 2600 using its first-generation 2nm process, SF2, and is preparing Exynos 2700 on its second-generation 2nm process, SF2P, for the second half of 2026. Samsung is also aiming to make progress with SF2P+, its third-generation 2nm process planned for mass production in 2027, and SF2T, a Tesla-dedicated process.
The longer-term question is whether Samsung can convert improving yields and utilization into more external orders. Business Post reported that Samsung is seen as having a growing chance of winning 2nm business from major US technology customers, including AMD, Meta, and Broadcom, though those orders have not been confirmed.
AMD is viewed as a potential customer for next-generation CPUs or consumer GPUs using Samsung's 2nm process. The company currently relies heavily on TSMC for CPU production, but capacity constraints at the Taiwanese foundry could push AMD to diversify its manufacturing base.
AMD CEO Lisa Su visited Samsung's Pyeongtaek campus in March and met Samsung Electronics Chairman Jay Y. Lee to discuss HBM4 supply and foundry cooperation. The visit has added to market speculation about deeper cooperation between the two companies.
Shinhan Securities analyst Huh Sung-kyu said AMD's foundry strategy appears to be entering a dual-sourcing phase to protect supply and margins. PC and server CPUs, along with consumer GPUs, are likely to be priority targets for diversification because of margin pressure and consumer price resistance, he said.
Samsung's HBM4 supply could also become part of that equation. With much of SK Hynix's HBM capacity tied to Nvidia, AMD may need to rely more heavily on Samsung for advanced memory, creating a potential opening for broader foundry cooperation.
Meta and Broadcom are also viewed as possible ASIC customers. Meta has announced plans to release its in-house AI chip, MTIA, every six months, while Broadcom designs Google's tensor processing units. Hana Securities analyst Kim Rok-ho said Big Tech ASIC shipments are expected to begin in earnest this year and surge in 2027, giving Samsung room to grow around advanced processes if TSMC capacity remains tight.
Samsung is focusing on stabilizing 2nm yield and improving power, performance, and area, or PPA, as it seeks larger AI customers. Its first-generation 2nm yield has reportedly risen above 50% through Exynos 2600 production, while Exynos 2700 is expected to be produced in more than double the volume of Exynos 2600, helped by second-generation 2nm technology and yield gains.
The company is also preparing its Taylor fab in the US for future 2nm production. Tesla's AI5 and AI6 chips are scheduled to enter mass production there from 2027.
Han Jin-man, president of Samsung's foundry business, said at Samsung's shareholder meeting in March that the Tesla projects would use the company's most advanced 2nm process, making it important to raise utilization for the node. He said Samsung would build experience with major customers and use that experience to improve yield.
The transition to HBM4 is turning the memory supercycle into a more direct opportunity for Samsung's foundry business. The longer-term challenge is whether Samsung can stabilize production for major 2nm customers and secure next-generation mobile application processor orders through Exynos 2700.
Article edited by Jerry Chen
