Samsung flags Q2 profit rebound on AI memory as TSMC June sales rise, spotlighting HBM and packaging bottlenecks
Kai Tanaka ·
Samsung Electronics said in preliminary second quarter guidance this week that operating profit rebounded on the back of a memory upcycle tied to AI servers. Taiwan Semiconductor Manufacturing Co reported June revenue growth on Friday in Taipei compared with a year earlier. Together, the updates indicate the AI infrastructure build remains on track into the second half of 2026.
The near term constraint still sits in high bandwidth memory and advanced packaging, the two inputs that gate accelerator output. HBM3E supply and capacity for 2.5D CoWoS style packaging remain the pacing items for Nvidia and AMD shipments to hyperscalers.
Semiconductors set the weekend tone
In Seoul, Samsung’s device solutions unit benefited from firmer DRAM pricing and a richer mix that includes HBM, according to its Korean filing. CFO Park Hark kyu has told investors the company will prioritize HBM ramps and advanced DRAM nodes through this year.
SK Hynix, a leading HBM supplier to Nvidia platforms, is expanding capacity at Icheon and Cheongju while migrating to finer DRAM process nodes known in the industry as 1b and beyond. The company has focused on higher stack counts and advanced thermal designs to meet accelerator requirements.
In Taipei, TSMC’s monthly sales underscore continued ramp of its N3 logic nodes and advanced packaging used in AI accelerators and networking chips. CEO C. C. Wei has highlighted persistent order backlogs for CoWoS and incremental capacity additions that stretch into 2026 to support customers including Nvidia, AMD, and Broadcom.
Japanese equipment and materials vendors remain key to these ramps. Tokyo Electron supplies etch and deposition tools used in DRAM and leading edge logic, while JSR and Shin Etsu provide resists and materials that support advanced nodes and interposer technologies.
Policy remains a swing factor. South Korea continues targeted tax credits and infrastructure support for the K Semicon cluster anchored by Samsung’s Pyeongtaek complex and SK Hynix sites, and Taiwan has backed packaging expansions to secure export earnings and domestic employment tied to AI manufacturing.
For global technology teams, the message is practical. Capacity for HBM, advanced substrates, and packaging slots is the gating resource for model rollouts and data center upgrades, more than GPU core counts alone. Procurement plans that lock in memory and packaging early will set deployment timelines and costs through year end.