Two headlines that ran the same week in 2026 look unrelated until you trace the supply chain between them: "Samsung and SK Hynix post record profits, warn shortages will persist into 2027," and "Nvidia to cut RTX 50-series GPU production by up to 40%." They're the same story told from two ends of one supply chain — and understanding the connection explains why a graphics card, or a laptop, got more expensive this year for reasons that have nothing to do with GPUs themselves.
The shortage, from the source
Samsung's memory division delivered 53.7 trillion won ($36.1 billion) in operating profit in Q1 2026 alone — roughly 94% of the company's entire quarterly profit — driven by AI memory demand. SK Hynix posted record quarterly revenue of 52.6 trillion won ($35.5 billion) and operating profit of 37.6 trillion won ($27.8 billion) on booming HBM sales. Both companies used their earnings releases to warn that the good news for shareholders is bad news for buyers: Samsung's memory chief Kim Jaejune stated on April 30, 2026 that "significant shortages" across memory products are expected to continue through at least 2027, explicitly because "the expansion of supply in 2026 and 2027 is expected to be limited due to constrained cleanroom space within the industry" even as AI-linked demand stays strong. Goldman Sachs independently sized the damage: a 2026 DRAM supply-demand gap of 4.9%, the most severe in 15 years, with DRAM spot prices up roughly 52% since January 2026.
Who holds the HBM market, and why the ranking keeps moving
| Company | HBM share (Q2 2025 snapshot) | 2026 trajectory |
|---|---|---|
| SK Hynix | ~62% | Defending lead; first to demo HBM4 at 16-Hi, 2TB/sec bandwidth |
| Micron | ~21% | Overtook Samsung for #2 per one industry count; guiding capex up 40%+ |
| Samsung | ~17% | Expected to strengthen as HBM3E clears qualification and HBM4 ramps; ~$74B total 2026 capex announced |
The industry's real 2026 battleground is the HBM3E-to-HBM4 transition. JEDEC's HBM4 standard supports 12-high and 16-high stack configurations at 24Gb or 32Gb die densities; a 16-high stack using 32Gb dies delivers up to 64GB per cube, roughly double HBM3's capacity, at up to 2TB/sec of bandwidth. SK Hynix has already shown working 12-Hi and 16-Hi HBM4 stacks. Getting there costs real, disclosed capital: Samsung's 2026 spending plan exceeds 110 trillion won (~$74 billion) across its semiconductor business, with its Taylor, Texas fab — the largest foreign investment in the state's history — targeting 2026 risk production. Micron and SK Hynix are each guiding capex up more than 40% year over year. None of that capital converts into shipped supply immediately, which is exactly why Samsung's own executives are pointing to cleanroom space, not money, as the binding constraint.
Where it shows up on a store shelf
This is the connection most data-center-focused HBM coverage skips: the same memory makers building HBM for AI accelerators also make the GDDR7 that goes into consumer graphics cards and the DRAM that goes into laptops — and when AI customers pay more and buy more predictably, consumer allocations lose the argument. Nvidia is cutting RTX 50-series GPU production by up to 40% in early 2026, with mid-range cards like the RTX 5070 and RTX 5060 Ti hit hardest, because Samsung and SK Hynix are prioritizing GDDR7 and HBM allocation toward AI data centers over gaming GPUs. The consumer-facing numbers vary by source and GPU tier but point the same direction: one tracking estimate put global graphics-card price increases at roughly 19% since late 2025 alongside a 20% cut to partner supply, while other reporting projects prices up 30% or more with continued scarcity through at least Q3 2026. Either way, the mechanism is the same one driving the data-center HBM story — memory suppliers allocating a genuinely constrained resource toward whichever buyer pays the AI premium.
The honest takeaway
Every figure above is dated and sourced to the companies or analysts making the claim, and they agree on the shape of the story even where they disagree on the exact number: real, disclosed shortage guidance from the two largest HBM makers, an independently estimated multi-decade-worst supply gap, tens of billions in capex that won't resolve the bottleneck before 2027 at the earliest, and a documented, causal path from that shortage to higher prices on ordinary consumer hardware. The "AI compute boom" and "why is RAM suddenly expensive" stories are, mechanically, the same story.
Sources: DataCenterDynamics on Samsung and SK Hynix's shortage warnings, Tom's Hardware on the 2027+ shortage outlook, Network World on Samsung's 2026 shortage/price warning, CNBC on SK Hynix's Q1 2026 earnings, Astute Group on 2026 HBM market share, TweakTown on SK Hynix's HBM4 demo, Tom's Hardware on Samsung's Taylor, TX capex, WebProNews on Nvidia's RTX 50-series production cut, TechTimes on RTX 50 pricing and supply impact.
