Nvidia Now Buys HBM From Three Suppliers. That Changes Everything Downstream.

Korea's Memory Duopoly · Part 2 of 3

Nvidia Now Buys HBM From Three Suppliers. That Changes Everything Downstream.

For two product generations, SK Hynix was effectively the sole credible supplier of high-bandwidth memory at scale. That era ended quietly this year. Working backwards from the supply chain to the income statements reveals two very different companies.

24 August 2026 · Part 2 of 3 — The businesses

Part 1 recap — SK Hynix and Samsung announced roughly $107bn in combined shareholder returns across three days in August, reversing a violent July selloff that was driven by a leverage unwind rather than deteriorating fundamentals. On the day the KOSPI recovered, the KOSDAQ fell 4.63%.

Key points

• Nvidia has qualified all three memory makers for HBM4. Supply concentration is loosening.
• SK Hynix retains the lead but its share is forecast to fall from 59% to 50% in 2026.
• SK Hynix earns a 74% operating margin versus Samsung's 48% — despite Samsung being 2.3× larger by revenue.
• 97.4% of Samsung's first-half operating profit came from semiconductors. Phones, TVs and appliances contributed almost nothing.

The qualification that reset the board

The competitive question in AI memory has never really been about specifications on a datasheet. It is about who can pass a customer's qualification process and then ship reliably at volume. For HBM3 and HBM3E, that customer was overwhelmingly Nvidia, and the answer was overwhelmingly SK Hynix.

HBM4 changed the answer. Nvidia's next-generation AI platform ships with HBM4 from Samsung, SK Hynix and Micron. Jensen Huang has confirmed all three passed quality validation and entered production.

The sequence is worth noting because both Korean firms can claim a first. SK Hynix established HBM4 mass-production capability in September 2025. Samsung was first to ship 12-high HBM4 in volume, in February 2026, and followed with 12-high HBM4E samples in May.

Figure 1. SK Hynix stays first, but the gap narrows. Counterpoint's series shows the same trajectory more sharply: 69% in Q1 2025 falling to 57% by Q4.

For SK Hynix this is genuinely double-edged. The addressable market is expanding — HBM is forecast at roughly $58bn in 2026 within a $563bn memory market — but exclusivity is eroding. And a customer with three qualified vendors negotiates differently than a customer with one.

Now look at what this produces on the income statement

Figure 2. Converted at roughly ₩1,400/USD. Samsung's half-year revenue combines ₩133.9tn in Q1 and ₩171.5tn in Q2.

The margin gap is the whole story, and the usual explanation — that Samsung is diversified across phones, displays and appliances while SK Hynix is a pure-play — is correct but drastically understates what has happened.

The 97.4% problem

Samsung's first-half operating profit was ₩146.7tn. Its semiconductor division alone produced ₩142.9tn of that.

97.4% Share of Samsung Electronics' H1 2026 operating profit generated by the semiconductor division. In Q2 alone the figure was 99%.

Figure 3. Foundry and System LSI remain loss-making, though management expects a turn as 2nm second-generation production ramps in the second half.

Read that again in portfolio terms. By profit, Samsung Electronics is already a memory company. Galaxy phones still sell in enormous volume; they simply do not move consolidated earnings anymore.

This matters for anyone holding Samsung as the "safer, diversified" way to own the memory cycle. Owning many business lines is not the same as having diversified earnings. On current numbers, Samsung's profit is nearly as exposed to DRAM pricing as SK Hynix's is.

Which explains the buyback structures

Part 1 left a question open: why did SK Hynix choose outright cancellation while Samsung opted for a larger but mixed programme? The balance sheets answer it.

	SK Hynix	Samsung
Structure	₩40tn buyback, full cancellation	Up to ₩110tn, dividends included
Business mix	Effectively pure memory	Memory plus foundry and devices
Competing calls on cash	Memory capacity expansion	Capacity plus a loss-making foundry
Result	Concentrated, irreversible	Larger, but deliberately flexible

SK Hynix held roughly ₩69.4tn in net cash at end-June, up more than ₩34tn in one quarter, against a simple set of uses. Samsung spent over ₩55tn on R&D and capex in the half and still needs foundry to reach breakeven. The simpler business could afford the more aggressive gesture.

Shared risks

Cyclicality

Memory has always been cyclical, and peak-out concerns are already circulating. Samsung's counter is that shortage conditions persist into 2028, and it is raising long-term agreements to 60–70% of capacity to dampen volatility. Whether this cycle genuinely breaks the pattern is the central open question for both stocks.

Customer leverage

Three qualified HBM vendors means pricing power shifts toward the buyer. This is the direct cost of the supply-chain broadening described at the top.

One-off costs

Record profits carry record profit-sharing. SK Hynix negotiated to pay 60% of its profit-sharing bonus in shares rather than cash — which conserves cash, but the expense is still recognised.

Coming next

Part 3 — The concentration problem. Why the KOSPI and KOSDAQ decoupled, what index concentration does to foreign investors benchmarking Korea, and which indicators actually describe market breadth.

Disclosure — This article summarises corporate disclosures, research-house estimates and press reporting for informational purposes. It is not investment advice and does not recommend the purchase or sale of any security. Market-share forecasts are third-party estimates and may prove inaccurate. Figures are current as of 23 August 2026. Won-to-dollar conversions are approximate, at roughly ₩1,400/USD. Readers should conduct their own research and consider consulting a licensed financial adviser.