Samsung’s preliminary Q2 2026 guidance — ~89.4 trillion won in operating profit, up roughly 19x year-over-year — is not an earnings story. It’s a structural signal about who controls the AI buildout’s most critical input.
Preliminary/unaudited guidance. OP range: 89.3–89.5T won. Revenue range: 170–172T won. Full audited results due July 30, 2026. The implied ~52% consolidated operating margin is extraordinary — verify against the audited report. No divisional or ASP breakdown in the preliminary release; any DRAM/NAND ASP figures circulating are analyst estimates only.
Samsung Electronics released preliminary Q2 2026 earnings guidance on July 8, 2026, reporting operating profit of approximately 89.4 trillion won — a roughly 19x, or ~1,800%, jump versus the 4.68 trillion won it posted in Q2 2025. Revenue came in at approximately 171 trillion won. These are unaudited K-IFRS figures; the full audited breakdown, including divisional results and memory ASP data, will follow on July 30.
The context makes the trajectory clear. In Q1 2026, Samsung had already reported total operating profit of ~57.2 trillion won — a +185% sequential jump representing a ~42.8% consolidated margin — with its DS (Device Solutions/semiconductor) division alone contributing ~53.7 trillion won at roughly a 66% operating margin. That Q1 print was itself driven by AI memory demand, industry-wide pricingstrength, and the start of mass shipments of HBM4 and SOCAMM2 memory modules for NVIDIA’s Vera Rubin platform. Q2’s number extends that ramp. This is not a one-quarter anomaly.
The preliminary release does not include DRAM or NAND average selling price data — those figures come with the full report. Any ASP percentages cited in analyst notes or media coverage before July 30 are estimates, not confirmed numbers. What is confirmed: Samsung’s semiconductor division is printing margins that rival the best software businesses on earth, and the driver is a structural imbalance between AI infrastructure demand and memory supply capacity.
The Ramp — Key Milestones
Q2 2025 — Baseline
Samsung operating profit: ~4.68 trillion won. Memory cycle still recovering; HBM supply constrained. The number Q2 2026 is measured against.
Q1 2026 — Inflection
Total OP ~57.2T won (+185% QoQ). DS division: ~53.7T won at ~66% margin. HBM4 + SOCAMM2 mass shipments begin for NVIDIA Vera Rubin. AI pricing power confirmed at scale.
Q2 2026 — Preliminary (July 8, 2026)
OP ~89.4T won. ~19x YoY. Revenue ~171T won. Implied consolidated margin ~52%. Full audited results July 30.
Capital Markets Signal — Parallel
SK Hynix pursuing a ~$29B US listing — the same AI memory trade expressed in equity markets. The oligopoly is monetizing the supercycle at every layer.
The key insight: Samsung’s operating profit didn’t grow by 19x because it sold 19x more chips. It grew because the AI buildout has created a structural supply shortage in the one component — high-bandwidth memory — that no hyperscaler can build around. Pricing power, not volume, is the mechanism. And that pricing power belongs to a three-firm oligopoly: Samsung, SK Hynix, and Micron.
The Structural Read
The Map of AI framework maps the entire AI stack across nine layers — from raw silicon and memory at the foundation, through compute infrastructure, model training, and inference, up to the application layer where end-users interact with products. Most of the capital, most of the media coverage, and most of the valuation premium flows to the top layers: the model builders, the application developers, the agents. Samsung’s Q2 number is a reminder of where the actual margin is.
HBM — high-bandwidth memory — is not a commodity input into AI infrastructure. It is the binding constraint. A GPU without sufficient HBM cannot run large-scale inference at commercial throughput. NVIDIA’s Blackwell and now Vera Rubin architectures are physically designed around HBM stacks; the memory is co-packaged with the GPU die. You cannot separate the compute roadmap from the memory roadmap. They are the same story.
Samsung’s HBM4 shipments for Vera Rubin, which began ramping in Q1 2026, represent a new generation of that dependency. SOCAMM2, the memory module form factor designed for AI server blades, extends the same logic to inference-at-scale deployments. Every new NVIDIA platform generation tightens the coupling between compute and memory — and every tightening strengthens the pricing position of the firms that can manufacture at yield and volume. Right now, that is effectively three companies globally.
Map of AI — Foundation Layer
“The most valuable layer in any technology stack is not the most visible one — it is the one where supply is structurally constrained and substitution is technically impossible. In the current AI buildout, that layer is high-bandwidth memory. The hyperscalers have the capital. The model builders have the talent. The memory oligopoly has the leverage.”
Analysts tracking cloud-provider capital expenditure patterns have flagged that memory now accounts for an estimated 50%+ of AI server BOM costs this year, with projections suggesting that figure could approach 70%+ in 2027 as inference workloads scale. Those are external estimates — the confirmed data point is Samsung’s margin. A ~52% consolidated operating margin, if it holds in the audited report, would be extraordinary for a diversified hardware manufacturer at this scale. It reflects a market that is not clearing at competitive prices. It reflects a chokepoint.
Three Implications
IMPLICATION 1 — THE OLIGOPOLY PREMIUM IS REAL AND DURABLE (FOR NOW)
Samsung, SK Hynix, and Micron control a supply chain that took decades and hundreds of billions in capex to build. New entrants cannot close that gap in a single AI cycle. The ~52% implied margin Samsung is printing is the market’s way of pricing that moat. SK Hynix’s pursuit of a ~$29B US listing is capital-markets confirmation of the same thesis: investors want direct exposure to the HBM chokepoint, not just downstream AI application plays. The oligopoly is monetizing at every layer simultaneously.
IMPLICATION 2 — NVIDIA’S ROADMAP AND SAMSUNG’S P&L ARE THE SAME DOCUMENT
The Vera Rubin platform’s performance specs are only achievable with HBM4 at the yields Samsung and SK Hynix are now shipping. This is not a vendor relationship — it is co-dependency at the architecture level. When NVIDIA announces the next platform generation, it is simultaneously announcing a step-change in HBM demand. Investors modeling NVIDIA’s revenue without modeling Samsung’s and SK Hynix’s supply capacity are reading half the spreadsheet. The memory boom and the compute roadmap are one trade.
IMPLICATION 3 — THE CYCLE RISK IS THE ONLY REAL RISK, AND IT IS REAL
Memory is one of the most cyclical industries in the history of technology. The current supercycle is real — but supercycles end. The single biggest risk is not a technology substitution or a new competitor; it is a pause or reallocation in AI infrastructure investment by the hyperscalers. If Microsoft, Google, Amazon, or Meta recalibrate their capex timelines — whether due to ROI pressure, regulatory intervention, or a demand plateau at the application layer — the memory market can flip from undersupply to oversupply faster than Samsung can adjust its fab utilization. The 19x YoY gain is a record. Records are also peaks until proven otherwise. Verify the ~52% margin figure against the July 30 audited report before treating it as the new normal.
Business Engineer Framework
The Map of AI — Where Samsung Sits in the Nine-Layer Stack
The Map of AI maps 200+ companies across nine layers of the AI value chain — from raw memory and silicon at the foundation through inference infrastructure, model platforms, and application distribution at the top. Samsung’s Q2 result is a case study in why the foundation layers — invisible to most users, inaccessible to most competitors — capture disproportionate margin during buildout phases. Understanding where each company sits in that stack is the prerequisite for any serious AI investment or competitive analysis.
Samsung’s ~89.4 trillion won operating profit — 19 times what it earned in the same quarter a year ago — is the clearest single number yet that the AI buildout’s economic center of gravity sits not with the model builders or the application developers, but with the three firms that manufacture the memory without which none of the inference infrastructure runs. That is a structural fact about the current phase of AI deployment, and it will remain true until either supply catches up with demand or hyperscaler capex pauses — whichever comes first. The supercycle is real. So is the cycle.
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
