The short version

When Lisa Su landed in Seoul on October 7, 2026, the stated agenda was memory. AMD's next AI accelerator, the Instinct MI455X, carries 432GB of HBM4 (AMD's published specification) — 50% more memory than its predecessor — and a single Helios rack system reportedly needs 31TB of it. Nobody ships that without a memory supplier. But according to Korean press reports, the October meetings may also have touched on foundry cooperation with Samsung, though AMD has not confirmed any plan. No contract was announced. The terms were not disclosed. And that is precisely what makes this story worth a buyer's attention — not as foundry gossip, but as a live demonstration of the oldest negotiation tactic in semiconductors: bundling a scarce resource with a contested one. To be clear about what follows: the bundling dynamic described here is pieced together from media reporting and industry analysis — neither Samsung nor AMD has confirmed that HBM4 supply is conditioned on foundry business.

Advanced AI accelerator chip package on a circuit board

What actually happened (and what didn't)

Separating verified fact from reported discussion matters here, because the gap between the two is where bad procurement decisions get made. **Company-confirmed:** In March 2026, the two companies signed an MOU to align on primary HBM4 supply for MI455X, collaborate on high-performance DDR5 for 6th Gen EPYC, and discuss a foundry partnership in which Samsung would manufacture next-generation AMD products. Lisa Su confirmed MI455X and Helios shipments have started, both using HBM4, with support from Samsung and SK hynix. She told reporters AMD 'continues to explore opportunities' with Samsung across memory and foundry (as rendered in Money Today's English translation), without confirming any foundry plan. **Media-reported:** Lisa Su visited Seoul on October 7 and, according to industry officials, met Samsung DS division chief Jun Young-hyun. Korean reporting (TrendForce and others) says the October meetings revisited the foundry question. AMD says it receives HBM4 from both Samsung and SK hynix; Samsung is the supplier AMD named as primary for MI455X in the March MOU — and Korean analysis reads that primary-supplier position as leverage for Samsung's foundry bid, an interpretation rather than a documented exchange condition, since no source shows memory supply actually conditioned on foundry business. A senior industry official was quoted in Korean coverage describing AMD's view of entrusting foundry work to Samsung as 'too risky' — an unnamed official's characterization, not an AMD statement. Estimates of Samsung's 2nm yield range from about 60% (Seoul Economy, August) to above 70% (KB Securities, September); Samsung has not published a figure, and major customers reportedly want about 70%. **Analyst/industry judgment:** Qualifying a new foundry process takes years (TSMC's C.C. Wei put it at about five years from technology choice to ramp), so nothing is likely to change on AMD's production lines in the near term, whatever the talks produce. The critical sentence for buyers: *no contract has been announced.* Everything about the foundry side is discussion and possibility — which is exactly the stage where leverage tactics are most visible.

Why HBM4 works as a bargaining chip

Bundling only works when three conditions hold — and if a supplier wanted to run the play here, this case shows all three. First, the bundled good must be genuinely scarce. HBM4 is. TrendForce noted in late September that HBM and conventional DRAM share limited advanced-process capacity, with tight supply expected into next year. And the demand math is unforgiving: each MI455X needs 432GB, up 50% from the prior generation — so the same number of accelerators sold consumes half again as much memory. Second, the buyer must need it on a fixed timeline. AMD has started shipping MI455X and Helios systems. Lisa Su herself described memory supply as "very tight" and said AMD is asking partners to ramp production "as quickly as possible." A buyer with product already shipping cannot credibly threaten to walk away from memory. Third, the contested good must be something the supplier desperately wants to sell. Samsung Foundry does. Samsung's foundry and System LSI units were estimated in Korean coverage to have lost roughly 2 trillion won in Q3 (an estimate published before Samsung reported divisional results); TrendForce put TSMC at 72.5% of foundry revenue in Q2 against Samsung's 5.9%. Samsung's 4nm lines are busy, but the industry sees 2nm customer wins as the real test. One flagship customer would change the narrative — and AMD's logic business is exactly the prize Samsung has chased. Scarce memory plus a shipping product plus a hungry foundry: that is the complete anatomy of a bundling opportunity — whether Samsung is actually running that play is precisely what has not been confirmed.

Macro shot of stacked memory chips on a substrate

The buyer's lesson: how to spot a bundling play

You will probably never negotiate HBM4. But the pattern — a supplier tying allocation of something you must have to something they want you to take — shows up at every tier of the supply chain: a distributor tying hot parts to slow movers, a manufacturer requiring forecast commitments for allocation priority, a foundry offering capacity only with a multi-year take-or-pay. Use this three-question test whenever a supplier's offer feels packaged: | Question | What to look for | Red flag | |---|---|---| | 1. Is the "must-have" item genuinely scarce, or artificially scarce? | Check independent supply data: lead times, allocation notices, multi-source availability | Scarcity exists only in this supplier's quotes, not in the broader market | | 2. Does the tied item solve *your* problem or *theirs*? | The tied item should reduce your risk, cost, or lead time on its own merits | The tied item is a new risk (unqualified process, unproven node) you're asked to absorb | | 3. What happens if you unbundle? | Ask for the must-have item standalone and compare total cost | The standalone price is punitive or availability "cannot be guaranteed" without the bundle | In the Samsung-AMD case, question 2 is the live one: *if* memory security were being traded for a foundry commitment, the bid would ask AMD to absorb qualification risk (new process, yield questions, a multi-year qualification cycle) in exchange. AMD's reported posture — keeping the door open without committing — is the correct posture either way: acknowledge the bundle, evaluate the tied item on its own merits, and never let scarcity on one line item lower your standards on another.

Second-source due diligence: yield and ecosystem, not price

This is where our October 7 guide on the Rapidus 2nm bet connects directly. The diligence framework is the same whether the second source is a Japanese startup or a Korean giant: **the real thresholds are yield, PPA (power/performance/area), and ecosystem — never the quoted wafer price.** Apply it to the Samsung foundry question: | Dimension | What to verify | Status per reporting | |---|---|---| | Yield at volume | Defect density and yield ramp on the specific node, not the marketing node name | No verified public data — the ~70% figure is a reported industry expectation, not a universal standard; yield must be verified product by product, which cannot be done from press reports | | PPA vs incumbent | Benchmarked power/performance/area against the current supplier's process | Unverified in public reporting — no data to evaluate | | PDK and IP maturity | PDK availability, third-party IP portfolio, EDA reference flows | Samsung has a shipping 4nm base but 2nm customer proof points are missing | | Supply-chain resilience value | Does the second source actually diversify your risk (geography, capacity, timeline)? | Partial — Samsung adds geographic and capacity diversity, but a multi-year qualification cycle means no near-term relief | | Contractual protection | Yield guarantees, capacity reservation terms, exit clauses | Nothing announced — nothing to evaluate | A second source whose yield cannot be verified is not a second source you can plan on; it is a science project with a purchase order. The reported 'too risky' characterization is exactly how a disciplined buyer treats an unproven node. Price never appears in this table, because a cheaper wafer you cannot ship on is the most expensive wafer of all.

Semiconductor cleanroom corridor with process equipment

What this means for your DRAM supply

There is a second-order effect buyers should not miss, and it connects to our DDR5 timing guide. Note carefully what does and does not follow. What is supported: every HBM4 stack Samsung allocates — to AMD, to NVIDIA, to Broadcom — consumes wafer capacity that could have been conventional DRAM, tightening the shared pool. What does not follow: Samsung winning more logic foundry business has no bearing on DRAM supply — foundry logic orders do not consume DRAM wafers. These are two separate causal chains; do not let the bundling narrative merge them. KB Securities expects Samsung's HBM selling prices to rise more than 100% in 2027, and TrendForce forecasts blended HBM average selling prices up 121% year over year — a mix-level forecast, not a per-device price hike — which tells you which direction the capacity is flowing. For procurement, the implication is the same one we laid out in the DDR5 guide: plan 2027 as if standard DRAM stays tight, qualify alternates now, and treat any 2028 relief as upside. Supplier power games at the top of the stack always compress the buyers at the bottom of it.

Warehouse shelves with boxed memory modules

Frequently asked questions

**Will AMD actually use Samsung Foundry for its chips?** Unknown. As of October 2026, only discussions have been reported — no contract, no timeline, no capacity allocation. AMD's public line is "continues to explore partnership opportunities," which is diplomatic for "not yet." A flagship accelerator handoff would come much later, if ever. **What is SF2P, exactly?** SF2P is Samsung's second-generation 2nm process, which Seoul Economy reports is expected to enter mass production for mobile products in the second half of 2026. The naming matters less than the diligence: whatever the suffix, the gates are yield at volume, PPA vs TSMC, and PDK/IP maturity, none of which have public proof points yet. **What is Samsung's 2nm yield right now?** No verified public figure exists. Estimates range from about 60% (Seoul Economy, August) to above 70% (KB Securities, September), and Samsung has not published a figure. Do not treat any specific yield number circulating online as confirmed — the only confirmed fact is that the yield question is the reported sticking point. **Why doesn't AMD just stay with TSMC?** It largely does — TSMC builds AMD's current CPUs, GPUs, and the MI455X itself. The Samsung talks are about *diversification*: spreading supply-chain risk, gaining negotiating leverage, and securing capacity for 2027 expansion. Second-sourcing is a hedge, not a breakup. **Is HBM4 really that scarce?** The evidence points to genuine tightness, not theater: TrendForce's forecast of tight supply into next year on shared advanced-process capacity, Lisa Su's own "very tight" characterization, AMD's dual-sourcing of HBM4 across Samsung and SK hynix, and per-unit memory content rising 50% generation over generation. Multiple independent signals beat any single claim. **How long would a Samsung foundry switch take?** Qualifying a new foundry process takes years — TSMC's C.C. Wei put it at about five years from technology choice to ramp: technology handoff, test-chip tape-outs, yield ramp, and qualification. But the start date of any Samsung-AMD foundry engagement is not public — the March 2026 MOU only covered discussing the possibility — so no firm product year can be derived from this week's talks. What can be said: a multi-year qualification cycle means no near-term switch is plausible, whatever the talks produce. **Should smaller buyers care about this story?** Yes, for the pattern rather than the players. The bundling tactic and the second-source diligence framework apply at every scale: when any supplier ties something scarce to something unproven, run the three-question test and the yield-gate checklist before you sign. And the HBM capacity squeeze affects standard DRAM pricing for everyone — that part of the story lands directly on your BOM. ---

Sources

Use the latest guidance for the exact part and manufacturer when deciding acceptance criteria.