Memory Prices Won't Stop Rising Until 2028

TL;DR: Memory prices aren't coming down until 2028 at the earliest. HBM (AI memory) is in a supply tsunami — 2027 capacity already locked in, prices doubling yearly. DRAM for your laptop/phone? Still structurally short. I told you to upgrade early in 2025. If you didn't listen, your wallet is about to hurt a lot more.

James here, CEO of Mercury Technology Solutions.

Hong Kong — July 2026

In early 2025, I told anyone who would listen: upgrade your laptop, your desktop, your phone. Do it now. Don't wait.

A year and a half later, memory and SSD prices have done exactly what I expected — they've gone vertical. And I'm here to tell you the same thing I told you then, except now with more data and worse news:

This isn't a blip. It's a structural supply crisis that runs through 2028.

Let me break down what's actually happening in the memory market, why your next hardware purchase just got more expensive, and why the AI boom is eating the entire semiconductor supply chain from the top down.

HBM: The Memory You Didn't Know You Needed (But AI Can't Live Without)

HBM — High Bandwidth Memory — is the stuff that goes into AI training chips. Not your laptop. Not yet, anyway.

Right now, there are exactly three suppliers: SK Hynix (55% market share), Samsung (~20%), and Micron (~20%). That's it. Three companies, entire global supply.

Here's the kicker: all three have their 2027 production capacity already locked in by long-term contracts. Every new wafer they bring online gets snapped up before it exists. Samsung is openly warning that 2027 HBM supply will be tighter than 2026.

The numbers are staggering:

| Year | HBM Market Size | Growth | |------|----------------|--------| | 2024 | $17B | — | | 2025 | $35B | +106% | | 2026E | $54B | +54% | | 2028-2030E | $100B+ | ~2x from 2026 |

Goldman Sachs — yes, the same Goldman Sachs — tried to call the top in July 2025. They predicted 2026 HBM demand would peak at $45B. Six months later, they quietly revised that number upward. Way upward.

The consensus now: HBM crosses $100B by 2028-2030. And that was before July 2026.

The China Variable Nobody Modeled

The forecasts above? All published before July 2026. Then China dropped two bombs on the market:

Kimi K3 and Qwen3.8 Max — open-weight models that demolished the cost-per-token floor.

I upgraded to K3 the week it launched. Quality jumped noticeably. But here's what caught my attention: a single 20,000-word report consumed 25% of my monthly token quota — about ¥50 ($7). That same report from a consulting firm? Starts at ¥20,000 ($2,800+). The economics are absurd.

So I upgraded to the ¥699/month plan. That's 2.5x the quota. Still not enough. I need 25x. And I don't want to wait 2 hours for a report — I want it in 10 minutes.

My personal compute requirement: ~120x what K3 currently delivers. And I'm not a power user. I'm a CEO who runs queries between meetings.

If every knowledge worker wants this experience — and they will — the compute demand is astronomical. Meanwhile, HBM manufacturers are running three shifts at full capacity and can only grow output 0.5-1x per year. Even at aggressive 2x annual growth, we're 7 years from satisfying this demand.

The shortage isn't ending. It's just getting started.

The HBM Technology Arms Race

HBM was a niche product in 2023 — 8% of memory revenue. Today it's 41%. Thirty percent of all wafer production now goes to HBM. That ratio keeps climbing.

Soon, traditional DRAM becomes the niche product. And here's where it gets interesting for consumers: legacy HBM — yesterday's AI memory — could trickle down to consumer devices. HBM3, even though it's "old," delivers 10x+ the bandwidth of current DRAM. Your next laptop might get a performance jump you didn't see coming.

The technology evolution is accelerating faster than spec sheets predict:

HBM4 Bandwidth (JEDEC spec: 8 Gbps)

• Samsung: 11.7 Gbps — 46% above spec

• Micron: 11 Gbps — 38% above spec

• SK Hynix: 10 Gbps — 25% above spec

Nvidia didn't expect this. The three memory vendors are outcompeting each other so aggressively that Nvidia now tiers its GPUs by HBM speed — Vera Rubin (top tier) exclusively uses Samsung HBM4. Mainstream products get SK Hynix and Micron.

Current production: 12-layer stacks. All three are sampling 16-layer. But the path to 16 layers is splitting into two competing technologies:

SK Hynix (conservative): Stick with MR-MUF — the "solder hundreds of tiny balls and fill gaps with glue" approach. They think there's still headroom.

Samsung (aggressive): Go all-in on Hybrid Bonding — wafers pressed together with <2nm gaps, essentially becoming a single crystal. No solder, no glue, just atomic-level fusion.

SK Hynix has Hybrid Bonding working in the lab too. They just think it's too expensive for now. Samsung is betting the company that Hybrid Bonding will be ready for HBM5. Whoever wins this bet will likely be #1 in market share by 2028.

DRAM: The Forgotten Shortage

Now let's talk about the memory in your actual devices — phones, laptops, desktops. The shortage here is just as real, just less headline-grabbing.

The optimists (Silicon Motion, TrendForce) say DRAM supply eases by late 2027. The pessimists (Goldman, Morgan Stanley) say late 2028 at the earliest.

Here's a data point that should terrify anyone shopping for RAM:

| Product | Jan 2025 Price | July 2026 Price | Increase | |---------|---------------|-----------------|----------| | DDR5 16Gb | $4.60 | $49.00 | +965% | | DDR4 16Gb | $3.20 | $80.00 | +2,400% |

DDR4 — a product that was supposed to be obsolete — is now the hottest commodity in memory. Why? Because DDR5 is so scarce that data centers are buying DDR4 just to keep systems booting. When a product with minimal remaining capacity faces panic buying, prices go parabolic.

Even Nanya Technology — a legacy Taiwanese DRAM maker that was an afterthought two years ago — is printing money.

The demand driver isn't consumers. PC and phone sales are down 10-15%. People are holding onto devices longer. The demand is coming from AI servers — every GPU cluster needs standard DRAM to run the host OS and manage data flow. Consumer demand collapsed, but enterprise AI demand more than absorbed the slack.

My personal data point: I bought a 24TB hard drive on JD.com in January for ¥6,700. Today that same drive costs ¥11,000+. And that's a mechanical hard drive. Not even flash storage.

The ChangXin Miracle That Isn't

"But James, ChangXin Storage is ramping DDR5 production. Won't that flood the market?"

Short answer: no.

Long answer: ChangXin just became the world's 4th largest DRAM fab in Q1 2026, passing Nanya. They've been expanding aggressively since 2024 — the fastest DRAM capacity growth globally.

SemiAnalysis ran the numbers: even with ChangXin's expansion, plus Samsung, SK Hynix, Micron, and Nanya all adding capacity, 2027 still shows a ~15% supply shortfall. ChangXin's 2027 output is already pre-sold.

ChangXin will soften DDR5 price increases. It won't crash them. The supply-demand math doesn't allow it.

The 2028 Inflection Point

My view hasn't changed from a year ago: memory prices stop rising in 2028 at the earliest.

But the dynamics are shifting. Two years ago, memory makers were scarred by the last downturn and expanded cautiously. Today, they're convinced this is a multi-year supercycle. New fabs are under construction everywhere. Most come online in 2028.

That creates a potential collision: massive new supply hitting the market right when... well, that's the question.

Scenario A: AI demand keeps accelerating. The 2028 supply wave merely slows price growth. Memory stays expensive. Your phone and laptop upgrade cycles stretch to 4-5 years.

Scenario B: AI demand plateaus or slows. The 2028 supply glut triggers a bloodbath. Prices crash. Margins evaporate. Weak players die. Strong players consolidate. Rinse and repeat — it's the semiconductor cycle, just compressed.

I'm not betting on Scenario B. The demand signals I'm seeing — my own usage patterns, the K3 upgrade path, the enterprise adoption curves — suggest we're still in the early innings of AI compute demand.

What You Should Do

If you need hardware, buy it now. Not next quarter. Now.

If you're running a data center or AI workload, lock in long-term memory contracts if you can. The spot market is going to get uglier.

If you're an investor, the memory supply chain — particularly HBM packaging and testing equipment — is a leveraged bet on AI demand that most people aren't tracking.

And if you're a regular consumer? Your next laptop purchase just got 30-50% more expensive. Sorry. I warned you in 2025.

Mercury Technology Solutions: Accelerate Digitality.