The Sins of the Past

HARVIC JOHN — 7 MIN READ — JAN 1, 2026

THE MEMORY WARS.

If you tried to buy a GPU or upgrade your RAM in the past six months, you already know something is wrong. Prices have doubled, tripled, quadrupled. Shelves that should be full aren't. And the culprit isn't a factory fire, a pandemic, or a shipping crisis. It's artificial intelligence — and its appetite shows no sign of slowing.

What is unfolding right now is now being called a DRAM supercycle a stturtural, demand-driven shortage unlike the cyclical downturns the industry has weatherd before. The AI buidout that began accelerating in 2023 has now fully meastasized into the consumer electronic supply chain, and the damage is severe.

By late 2025, contract prices for DDR5 chips had surged nearly 300% in just three months. A DDR5 16GB kit that cost around $90 in early 2025 now fetches $240 or more in retail. NVIDIA has reportedly cut RTX 50-series GPU production by 30–40% compared to the first half of last year. And the PC market — which was supposed to boom in 2026 due to the Windows 10 end-of-life cycle — is instead projected to contract by up to 9%.

Where Did All the Memory Go?

The answer is both simple and maddening: the hyperscalers took it. Companies like Microsoft, Google, Amazon, and Meta are deploying AI infrastructure at a scale that was, until recently, almost impossible to comprehend. Each AI GPU cluster requires not just raw compute — it requires enormous quantities of High-Bandwidth Memory (HBM), a stacked, ultra-fast memory architecture that enables the data throughput AI models demand.

The three companies that control over 95% of global DRAM production — Micron, SK Hynix, and Samsung — are not charities. When AI chip manufacturers came knocking with orders for HBM at $60 to $100 per module, compared to the $5 to $10 a conventional DDR5 module fetches, the math was obvious. Production lines were retooled. Consumer memory was deprioritized. And the ripple effects are now everywhere.

It gets worse. The production process for HBM is extraordinarily complex — Micron stacks 12 to 16 layers of memory dies into a single cube-like package. This isn't a matter of turning a dial; retooling a fab for HBM production takes months and billions of dollars. And in the meantime, consumer DDR5 and GDDR7 supplies have dwindled to critical levels.

The Sins of the Past

This crisis did not appear from nowhere. The memory industry planted these seeds years ago during the 2022–2023 downturn, when post-pandemic demand collapsed and DRAM prices cratered. Samsung, facing catastrophic losses, cut production by roughly 50%. The entire industry slashed capital expenditure and built essentially no new fab capacity through most of 2024 and into early 2025.
It was a rational response to a cyclical bust. It was also a catastrophic miscalculation about what was coming.

2022-23 • Indusrty Cuts Production
Samsung cuts output ~50%. CapEx across all three DRAM makers falls to multi-year lows. No new fabs are ordered.
Mid-2025 • AI Demand Surges
Hyperscaler GPU deployments accelerate. HBM demand spikes. Manufacturers begin reallocating consumer DRAM capacity.
Oct-Dec 2025 • Prices Explode
DDR5 contract prices rise ~300% in three months. Japanese retailers introduce purchase limits. NVIDIA quietly plans GPU production cuts.
Q1 2026 • Crisis Goes Mainstream
DRAM prices rise another 90% quarter-over-quarter. Dell, ASUS, Acer announce price hikes. PC and smartphone market forecasts sharply revised downward.
2027-28 • New Fabs Come Online
Micron's Idaho and New York facilities begin volume production. Analysts expect partial price relief — but baseline pricing stays elevated.

When AI demand surged in mid-2025, the industry had no capacity buffer. The fabs that should have been ordered in 2023 won't be online until 2027 or 2028 at the earliest. That is an immovable timeline — you cannot rush semiconductor fab construction.

What This Means for Consumers

The damage is cascading through every tier of the market.

Gaming GPUs

Video memory — GDDR7 and GDDR6X — is produced by the same facilities being retooled for HBM. Reports suggest VRAM now accounts for over 80% of the bill of materials for some high-end consumer GPUs. NVIDIA and AMD both signaled price hikes beginning in early 2026. NVIDIA's reported 30–40% cut to RTX 50-series production is the most direct consequence: fewer cards will be made, and the ones that exist will cost more.
The 16GB tier — already borderline for demanding modern titles — is thinning out. The path to an affordable 16GB card is narrowing precisely when software demands make it most necessary.

PCs and Laptops

HP has reported that memory now accounts for 35% of PC build materials — up from a fraction of that figure just a year ago. Dell is planning hardware price hikes in the hundreds of dollars. Framework is raising the cost of RAM upgrades. IDC now projects the PC market will contract between 5% and 9% in 2026 — a reversal of what should have been a boom year driven by AI PC demand and the Windows 10 end-of-life replacement cycle.

THE CRUEL IRONY
AI PCs — which require 16GB of RAM minimum for Microsoft's Copilot+ certification — are arriving just as RAM becomes prohibitively expensive to include. The industry is being forced to ship less memory into devices that need more of it, at exactly the wrong moment.

Smartphones

Global smartphone shipments are projected to fall from roughly 1.26 billion units in 2025 to about 1.12 billion in 2026 — a 12.9% decline that would be the steepest drop since the global financial crisis of 2008. Xiaomi's CFO has publicly warned that memory pressures will drive smartphone prices higher in 2026. Some Android OEMs are quietly downgrading base configurations — offering 4GB of RAM instead of 8GB — to manage costs and secure supply.

When Does It End?

The uncomfortable answer: not soon. Micron is building two major fabs in Boise, Idaho, expected to begin producing memory in 2027 and 2028. A third facility in Clay, New York is scheduled to come online around 2030. SK Hynix and Samsung are similarly expanding, but fab construction timelines are measured in years, not months.

Even when new capacity arrives, it may not deliver the price relief consumers hope for. AI demand is also growing — at roughly 20–30% annually, by most estimates. Supply expansion may simply keep pace with consumption rather than outpace it. IDC's analysts are cautious: "the end of an era of cheap, abundant memory" may be a structural shift, not a temporary shock.

Some alternatives are emerging. XPUs — custom accelerators from Google (TPUs), Amazon (Trainium), and Intel (Gaudi) — are gaining traction as organizations seek GPU alternatives. About 31% of enterprises are now evaluating TPUs and 26% are evaluating Trainium, according to a recent survey. These chips diversify the compute supply chain, but they don't solve the underlying memory constraint: they, too, need HBM.

For now, the consumer sits at the bottom of the allocation stack. AI companies are first in line. Server makers are second. Consumer GPU and PC makers get what's left. And what's left, increasingly, isn't much.

If you need RAM today, the best advice from industry insiders is blunt: buy it now, before the next wave of price increases arrives. The shortage is expected to deteriorate further before it improves. The fabs are coming — but the machine doesn't care about your upgrade plans.

NOTABLE QUOTES
"Because our demand is so high, every HBM supplier is gearing up, and they're all doing great."
— Jensen Huang, NVIDIA CEO, CES 2026
"Consumers shouldn't wait for their PC upgrades — the situation is expected to deteriorate further."
— Wallace Santos, MAINGEAR CEO

Sources: Unknown