Samsung’s DDR5 Memory Just Went From $149 to $239 — blame AI Chips

Wednesday 12 August 2026 | Finance, The Long View

TL;DR — Samsung’s DDR5 memory contract prices jumped from $149 to $239 in a single quarter, driven by AI chipmakers buying up HBM and DDR5 capacity. Samsung and SK Hynix are reallocating fab lines toward higher-margin AI memory, tightening supply for everyone else.

Samsung raised the price of a 32GB DDR5 memory module from $149 to $239 in September, a 60% jump in a single price update. Nothing changed about how that memory works, and nothing changed about the laptop or desktop it was going to end up in. What changed is that the fab capacity that used to make it is now making something else: memory for AI accelerators. The AI buildout’s cost isn’t staying contained to hyperscaler capex lines. It’s showing up in the price of a memory stick.

Why HBM Is Eating the Memory Supply

High-bandwidth memory, HBM, is the stacked, high-speed memory that sits directly next to AI accelerator chips, feeding data in and out fast enough to keep GPUs busy instead of waiting. It’s a different product from the DRAM in a laptop, but it comes from the same fabs, competing for the same wafer starts. HBM production consumes roughly three times the wafer capacity per gigabyte that standard DRAM does. Every wafer a manufacturer commits to HBM is a wafer it isn’t using to make the memory that goes into phones, PCs, and servers that have nothing to do with AI.

Faced with a choice between AI-accelerator margins and consumer-electronics margins, manufacturers are choosing AI. Micron exited the consumer memory market entirely this year to focus on enterprise and AI customers. That’s not a company retreating from a shrinking business. It’s a company reallocating capacity toward the higher-margin one and leaving consumer buyers to compete for what’s left.

The Numbers Behind the Shortage

Samsung and SK Hynix raised HBM3E prices nearly 20% for 2026 orders. As of this month, HBM3E runs roughly $300 per 36GB stack, with next-generation HBM4 priced around $500 per 48GB stack. HBM capacity across SK Hynix, Samsung, and Micron is essentially sold out for 2026. That scarcity doesn’t stay contained to HBM. TrendForce projected conventional DRAM contract prices to rise 55% to 60% in the first quarter of 2026 alone, with server DRAM climbing more than 60%. The Samsung DDR5 module price jump isn’t an isolated data point. It’s what a wafer-capacity shortage looks like once it reaches retail.

Who Actually Pays This Tax

Every earnings report we’ve covered this season has framed AI infrastructure spending as something happening on hyperscaler balance sheets — Microsoft’s backlog, Meta’s capex guidance, Alphabet’s buyback trade-offs. Those are real costs, but they’re borne by shareholders who chose to hold that exposure. The memory shortage doesn’t stay inside that circle. It shows up in the bill of materials for every laptop, phone, game console, and server built by a company that has nothing to do with the AI race, and eventually in the price consumers pay for those products. Nobody buying a new laptop this year voted to subsidize GPU memory demand. They’re paying for it anyway, built into a component cost they’ll never see itemized.

Who’s Actually Winning the Squeeze

SK Hynix holds roughly 50% to 55% of HBM market share and is currently the only supplier reliably delivering both HBM3E and the next-generation HBM4, a technical lead that’s translating directly into pricing power. Samsung holds 35% to 40%, still catching up on HBM4 yields. Micron, at 5% to 10% of HBM share, made the more radical bet: walking away from consumer memory entirely rather than split capacity between a low-margin business under pricing pressure and a high-margin one with a multi-year backlog. That’s the same capital-allocation logic we’ve seen play out at the hyperscaler level all season, just one layer further down the supply chain, where the company making the choice doesn’t get an earnings call full of analysts asking about it.

We’ve tracked this same dynamic — a physical capacity constraint setting the real ceiling on how fast the AI buildout can convert into shipped products — in TSMC’s sold-out advanced packaging capacity. The memory squeeze is the same story at a different layer of the stack, and it’s the one layer of this cycle’s cost that’s already showing up in a receipt.

Sources: TrendForce DRAM and HBM pricing reports, 2026; Samsung and SK Hynix HBM3E pricing announcements, December 2025; Micron product line disclosures, 2026. Third Pole Markets holds no position in Samsung, SK Hynix, or Micron as of publication. This is not investment advice — see our About page for our full disclosure policy.

Memory pricing trends are tracked by TrendForce, with Samsung’s own disclosures at its investor relations page.

Tags: HBM | Memory Chips | Samsung | Semiconductors | SK Hynix

Author & Analysis

By Jack Coulter

Jack Coulter spent seven years on equity trading desks in Chicago and New York, four of them on the sell-side covering tech, then five more on the buy-side at a concentrated long-only fund. He left asset management in 2024, tired of writing research to fit a mandate instead of a conviction. Third Pole Markets is what came next: independent equity research, funded by his own positions, answerable to no client. Born and raised in Akron, Ohio, now based in New York, he holds long positions in the names he covers, disclosed in every piece, not buried in a footnote.

More on the Tape

Four Customers Are 61% of Nvidia’s Revenue. Nvidia Won’t Say Who.

Four Customers Are 61% of Nvidia’s Revenue. Nvidia Won’t Say Who.

TL;DR — Four customers account for 61% of Nvidia's revenue, and Nvidia won't name them in its filings. That concentration is the real risk sitting underneath Nvidia's growth numbers, not the chip roadmap. Nvidia's most recent quarterly SEC filing lists four customers,...