AMD Chief Executive Lisa Su held discussions with Samsung Electronics and SK Hynix focused on artificial intelligence, according to a Chosun Ilbo report dated to the current reporting cycle. The meetings place the chief executive of the second-largest merchant AI accelerator vendor directly in front of the two memory suppliers that, together with Micron, control the global high-bandwidth memory (HBM) market.

Why does HBM allocation matter now?

HBM stacks supply the working memory on which AMD's Instinct MI300X, MI325X and forthcoming MI350 accelerators depend. AMD shipped the MI300X in volume through 2024 and disclosed in its Q4 2024 earnings call that Instinct revenue exceeded $5 billion for the full year. Each MI300X package carries 192 GB of HBM3E, and AMD's stated roadmap moves the next generation to HBM4.

Memory accounts for a disproportionate share of accelerator bill of materials. Trendforce and other supply-chain trackers have estimated HBM at roughly 35-50% of an AI GPU's silicon cost depending on stack count. Constraining SM sales to any single supplier in this segment translates directly into delivery schedules for hyperscaler customers such as Microsoft, Meta and Oracle.

What does AMD gain from dual sourcing?

SK Hynix entered mass production of 12-layer HBM3E in 2024 and has held the dominant share of NVIDIA's HBM allocations. Samsung qualified its 12-layer HBM3E later in the same cycle and has invested in HBM4 development at its Pyeongtaek facilities. Engaging both Korean suppliers in parallel gives AMD leverage on pricing, allocation priority and qualification timelines for next-generation stacks.

• SK Hynix: 12-layer HBM3E in mass production; HBM4 sampling reported in 2024
• Samsung: 12-layer HBM3E qualified in late 2024; HBM4 ramp targeted for 2026
• Micron: third incumbent, not named in the Chosun Ilbo report

Where does this leave the competitive picture?

NVIDIA's HBM allocation is widely reported to run ahead of AMD's in absolute volume, reflecting both earlier qualification cycles and larger Blackwell shipment forecasts. Su's outreach to Samsung — historically the runner-up in HBM market share — suggests AMD is seeking to widen the supplier base rather than concentrate orders with SK Hynix alone.

Lisa Su has previously stated on AMD earnings calls that supply diversification across memory partners is a stated priority. The Chosun II report does not disclose whether the talks produced specific volume commitments, qualification milestones or pricing terms.

What remains undisclosed?

The Chosun Ilbo headline identifies the parties and the topic. It does not disclose:

• The meeting location or date
• Whether other AMD executives attended
• Whether the talks covered HBM4 specifically or earlier HBM3E allocations
• Whether Samsung's foundry or system LSI divisions were represented
• Any pricing, volume or timeline commitments emerging from the discussions

How should customers read the meeting?

For hyperscalars evaluating AMD's 2025 and 2026 accelerator roadmaps, the signal is structural rather than transactional. Engaging Samsung and SK Hynix in parallel indicates that AMD intends to qualify multiple memory suppliers for its HBM4 generation, a step that historically took NVIDIA roughly two years from initial sampling to volume ramp.

Samsung's HBM4 timeline places it roughly 12-18 months behind SK Hynix in customer qualification, based on disclosed company guidance. AMD's willingness to engage Samsung now is consistent with the pattern of building parallel qualification paths ahead of MI400-series and MI350 launch windows.

The report leaves the supply picture in qualitative terms. No new product announcement, contract figure or production milestone accompanied the disclosure.