6 min read<\/p>\n
Samsung Q2: memory remains scarcer than expected. For buyers, the shortage matters less than the consequences for prices, lead times and supply contracts in Europe.<\/strong><\/p>\n Das Wichtigste in K\u00fcrze<\/p>\n Related:<\/strong> Cheap AI from China: What procurement must check<\/a> \u00b7 When a German AI model actually pays off<\/a><\/p>\n year-over-year operating profit (Q2 2026)<\/p>\n Source: Samsung Electronics, Q2 2026 Results<\/p>\n<\/div>\n According to Samsung Electronics\u2019 Q2 2026 figures, group revenue reached 171.500 trillion won, with operating profit at 89.500 trillion won-both all-time highs. Quarter-on-quarter, revenue rose 28 percent and operating profit 56 percent. Year-on-year, the surge was dramatic because the prior-year quarter was still mired in a weak memory cycle.<\/p>\n What is HBM?<\/strong> High Bandwidth Memory is ultra-fast stacked memory for AI accelerators and servers. It sits close to the processor and delivers far higher data throughput than classic server DRAM. This exact blend of HBM and server DRAM is driving today\u2019s scarcity.<\/p>\n The growth engine is in the Device Solutions segment. Semiconductor contributed 127.500 trillion won in revenue and 89.200 trillion won in operating profit. Memory again hit record levels as Samsung prioritised scarce capacity for server products. DRAM and NAND both achieved all-time highs in bit shipments. At the same time, the company ramped HBM4 sales and, by its own account, shipped the first HBM4E samples to key customers.<\/p>\n Investments are chasing demand. Q2 CapEx totalled 16.800 trillion won, up 5.500 trillion won versus the prior quarter. Of that, 15.400 trillion won went to Device Solutions. Memory investments rose, among other things, for the Pyeongtaek fab and supporting infrastructure. If you\u2019re still banking on \u201cchip prices will fall again,\u201d read this as the opposite: the manufacturer is expanding capacity yet still signalling persistent scarcity.<\/p>\n Crucial context for SMEs: the absolute won figures are group-level. The real leverage for mid-market buyers lies in allocation. When server and HBM demand dominate the production lines, commodity modules and enterprise SSDs get caught in the same capacity decisions. Distributors then prioritise large customers and framework agreements. Mid-market reorders slip back, even if the volume seems small.<\/p>\n Samsung expects robust server demand in the second half of 2026, driven by sustained AI infrastructure capital expenditure and the broader adoption of agentic AI systems. Server DRAM, enterprise SSDs, and HBM are projected to grow faster than supply, while mobile and PC demand will only partially offset this imbalance. Ultimately, the market is set to remain undersupplied.<\/p>\n Long-term take-or-pay agreements further tighten the situation. Earnings communications and investor relations reports highlight finalized contracts with the five largest global data center customers, alongside advanced negotiations with additional major buyers. For mid-market buyers, this means predictable capacity is first allocated to hyperscalers and AI platforms-late orders translate into residual volumes and extended lead times.<\/p>\n Samsung also anticipates that supply constraints could intensify through 2027, with the sharp rise in token generation by AI workloads serving as a key mid-term driver. More tokens mean more inference operations, more servers, more HBM, and additional high-speed memory. This isn\u2019t confined to cloud environments: on-premises inference, edge boxes in manufacturing, and AI-capable industrial PCs demand the same memory mix.<\/p>\n Sources: Samsung Electronics Q2 2026 Results and earnings call materials; analysis by MyBusinessFuture.<\/p>\n<\/div>\n The second page of the memory rally is showing up in the equipment business. Samsung itself reports growing sales in mobile and network segments, yet faces margin pressure from expensive components. That same cost logic is hitting DACH companies: industrial PCs, edge-AI controllers, NAS and backup systems, GPU servers for pilot projects, and replacement storage for existing fleets are all getting pricier or arriving later.<\/p>\n Then there\u2019s energy and location. More local inference means more power-and often more cooling in the server room. Bringing workloads back in-house to cut token costs doesn\u2019t automatically save money. Memory, accelerators and energy costs all land in the same business case. Leave any of the three out, and every \u201cAI saves staff\u201d calculation is incomplete.<\/p>\n Procurement and IT shouldn\u2019t wait for the next price dip. Vendor messaging is pointing to persistent shortages and multi-year visibility for large customers. If you\u2019re planning bigger roll-outs in 2027, start looking for alternatives today: qualified SKUs, second-source suppliers, cloud-bursting for spikes, and clear priorities on which workloads truly need on-prem storage.<\/p>\n More memory and more accelerators on-site mean measurably higher power draw. Bringing inference back from the cloud only cuts token costs if electricity price, cooling and utilization are included. In energy-intensive operations this collides with existing grid fees and whether the server room has any spare capacity at all.<\/p>\n For senior management it\u2019s a control problem. IT requests hardware for AI pilots. Facilities counts kilowatt-hours. Procurement sees rising storage prices. Without a shared spreadsheet each department acts alone-and the project shows up in the next forecast as a surprise. Samsung\u2019s messaging on sustained AI-infrastructure demand is therefore also a signal to energy managers: capacity planning belongs in the same calendar as memory procurement.<\/p>\n 1. Catalog critical SKUs.<\/strong> List server DRAM modules, HBM-dependent systems, enterprise SSDs and edge devices with AI components. Flag anything that must be replaced or expanded by Q1\/Q2 2027. Without this list, negotiations are blindfolded.<\/p>\n 2. Lock in supply.<\/strong> Ask distributors and OEMs for firm delivery windows and single-source dependencies. Check whether framework agreements or pre-reservations are possible. In parallel, identify which workloads can shift to cloud or rented GPU capacity if hardware slips.<\/p>\n 3. Recalculate the total budget.<\/strong> Line up hardware, electricity, cooling and ongoing token costs side by side. Agentic-AI projects rarely fail at the demo stage; they stumble on the second wave-more memory, more instances, more energy. Samsung\u2019s numbers are an early indicator that this second wave will cost more than the 2024\/2025 budget assumed.<\/p>\n Samsung is setting the tempo for the memory market. The beat is this: AI infrastructure and agentic systems tie up capacity, prices and schedules. For executive teams and IT the choice is clear. Either lock in critical hardware and supply paths now, or accept that projects in 2027 will hit memory limits long before the software does. Waiting in this cycle only pushes the risk into the next forecast.<\/p>\n No. Hyperscalers are the first to lock in large volumes of HBM. Mid-sized companies feel the impact through higher prices and longer lead times when upgrading servers, deploying edge AI, industrial PCs, or expanding storage.<\/p>\n<\/details>\n Agentic systems generate more interaction steps and therefore more inference tokens. This drives demand for server DRAM, fast SSDs, and HBM-not just training clusters in the cloud.<\/p>\n<\/details>\n Only if the business case is rock-solid, covering memory, power, and operations. A hybrid approach-secure on-premises hardware plus cloud burst capacity-often proves more resilient than a large in-house build without guaranteed supply.<\/p>\n<\/details>\n Revenue, operating income, segment contributions, and CapEx are drawn from Samsung Electronics\u2019 Q2-2026 disclosures. Outlook statements on scarcity and agentic AI reflect manufacturer forecasts and should be read as such.<\/p>\n<\/details>\n\n
What Samsung reported for Q2<\/h2>\n
Agentic AI and LTAs: Why Structural Scarcity Persists<\/h2>\n
\n\n
\n \nSignal<\/th>\n Q2 \/ Outlook<\/th>\n Mid-market Interpretation<\/th>\n<\/tr>\n<\/thead>\n \n Revenue \/ Operating Profit<\/td>\n 171.5 \/ 89.500 billion Won<\/td>\n Memory cycle has rebounded into a boom phase-budgeting for falling hardware prices is risky<\/td>\n<\/tr>\n \n Product Mix<\/td>\n Servers, HBM4, HBM4E samples<\/td>\n Standard DRAM and SSDs remain tied to AI allocation<\/td>\n<\/tr>\n \n CapEx<\/td>\n 16.800 billion Won, DS 15.4 billion<\/td>\n Expansion yes, easing no-capital spending follows scarcity<\/td>\n<\/tr>\n \n LTAs \/ Customers<\/td>\n Top-5 data centers secured, more in pipeline<\/td>\n Volume and timelines prioritized for major buyers<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n Where mid-sized companies are already feeling the bill<\/h2>\n
Energy and location: the invisible cost block<\/h2>\n
Three checks for procurement, IT and executive teams<\/h2>\n
The next decision<\/h2>\n
Frequently Asked Questions<\/h2>\n
Does memory scarcity only affect hyperscalers?<\/strong><\/summary>\n
What does \u201cagentic AI\u201d mean for storage needs?<\/strong><\/summary>\n
Should mid-sized companies now invest in their own GPU servers?<\/strong><\/summary>\n
Which figures can be relied on?<\/strong><\/summary>\n
Editor\u2019s Reading Picks<\/h2>\n