The race to deploy Artificial Intelligence on a large scale is no longer driven solely by GPU power. By 2026, the industry’s focus is shifting toward a less flashy but critical component: memory, especially HBM (High Bandwidth Memory) and advanced DRAM that power data center accelerators. This change in priorities is increasingly reflected in the financial forecasts circulating in the market.
Two forecast charts attributed to Morgan Stanley Research show upward revisions for Samsung Electronics and SK Hynix for 2026-2028, with a message that directly interests a tech-oriented audience: if AI continues to advance as it has so far, memory is not just another component, but a limiting factor that could redefine margins, investments, and industrial strategy.
Samsung: When profit is explained by semiconductors (and, within, by memory)
For Samsung Electronics, revised estimates for FY2027 (FY27E) increase the sales of the company to 647.680 billion won from 599.480 billion won in the previous scenario, an 8% growth. But the most notable jump is in operating profit: from 241.454 billion won to 317.379 billion won, a 31% revision.
This shift is not a minor adjustment: it means the estimated operating margin for FY27E rises from 37.0% to 49.0%, a gain of 12 percentage points. In a conglomerate with multiple divisions, such a move usually signals one thing: the “core” business is driving the rest.
Breaking down by areas, the data shows most of this momentum coming from semiconductors and especially the memory line. For FY27E, the “Memory” segment increases its revised sales to 388.347 billion won from 309.708 billion won, a 25% rise, and the revised operating profit reaches 279.302 billion won from 216.777 billion won, a 29% increase. At the same time, the semiconductor division (DS) also makes significant upward revisions: 437.900 billion won in sales (up from 357.471 billion won, +22%) and 279.284 billion won in operating profit (up from 217.040 billion won, +29%).
The technological contrast becomes clear when examining the rest of the group. For FY27E, “Mobile/Network” shows revised sales down to 129.753 billion won (–18% compared to previous projections) and a dramatic decrease in revised operating profit to 9.885 billion won, a -52% drop from earlier estimates (20.672 billion won). This snapshot reflects a mature market: mobile devices remain mass-market, but they no longer drive profitability as centralized data centers take over the spotlight.
SK Hynix: “Software-like” margins for a hardware business… if HBM maintains its scarcity
In SK Hynix, the picture is more straightforward because the company is much more focused on memory. For FY27E, sales are revised from 258.009 billion won to 304.878 billion won (+18%) and operating profit from 181.336 billion won to 225.354 billion won (+24%). The estimated operating margin rises from 70% to 74% (an increase of 4 points).
The interesting part, from a tech perspective, is in the “DRAM (& HBM)” breakdown. It illustrates the story that the sector has been telling for months: memory associated with AI is changing the rules. For FY27E, revised DRAM/HBM sales climb to 237.721 billion won (+21% over previous projections), and revised operating profit reaches 187.568 billion won (+27%). NAND also improves but less significantly: revised sales are 65.843 billion won (+11%), with revised operating profit at 37.786 billion won (+14%).
In other words: if the HBM supply chain remains tight, SK Hynix appears positioned to sustain margins traditionally associated more with services than with industrial manufacturing. This is precisely why the debate around the sector has become so polarized: some investors see it as a structural supercycle; others perceive it as a cyclical peak that historically corrects itself once new capacity arrives.
Why this matters to infrastructure teams and the AI ecosystem
For a tech-focused audience, the key question isn’t just who will earn the most in 2027, but what this suggests about AI infrastructure design:
- HBM and advanced DRAM are becoming elements of planning. Reserving GPUs alone is no longer enough; memory and packaging must be secured within the same time window.
- The focus shifts to architecture: optimizing models, batching, quantization, and inference strategies are no longer only about computational efficiency; they’re also about memory efficiency.
- Manufacturers capable of industrializing new generations and delivering stable volumes can set the tone for the market, just as GPUs did during the early years of the boom.
The data tables reinforce this narrative: for Samsung, the driver of improvements isn’t mobile devices, but the semiconductor and memory segments. For SK Hynix, most of the improvement is driven by DRAM/HBM. This isn’t just an accounting detail; it’s a map of technological power in AI.
The “big risk” behind the figures: supercycle or episode?
The same forecasts that excite the market also contain an implicit caution: these are projections, dependent on assumptions about demand, prices, capacity, and yields. If memory scarcity eases earlier than expected due to new capacity, margins could normalize. And if AI spending becomes more selective — for efficiency or return reasons — the bottleneck may shift elsewhere in the stack.
Nevertheless, the strategic shift is already real: for the first time in years, memory is again a key component driving schedules, budgets, and priorities. As these revisions show, it’s also starting to shape the financial narrative of major industry players.
Frequently Asked Questions
What is HBM memory, and why does it significantly impact AI performance?
HBM is high-bandwidth memory designed to feed accelerators with much higher speeds than conventional memories, reducing bottlenecks when training or running large models.
Why are forecasts improving so much for 2027 and not just for 2026?
Because the market assumes AI demand isn’t just a one-time spike: it’s evolving into sustained deployments in data centers, which drive longer-term contracts, capacity planning, and margin expectations.
What risk could threaten this high-margin memory scenario?
A relaxation of scarcity due to new capacity, industrial performance improvements, or shifts in demand. Historically, the supply-demand balance in memory can change quickly.
Are these numbers actual revenues or estimates?
They are revised estimates (Previous vs. Revised) for fiscal years (FY26E, FY27E, FY28E) in billions of won, according to forecasts attributed to Morgan Stanley Research.