The memory supercycle is here.

Vera Rubin $NVDA will consume 60% of global HBM capacity and prices are shooting into the stratosphere.

The AI revolution doesn’t rest on GPUs alone — it rests on memory. And there is dramatically too little of it. Goldman Sachs talks about the worst DRAM deficit in 15 years, Micron $MU is sold out for 2026, and SK Hynix $HY9H.F and Samsung $SSNLF are fighting for tens of percentage points of market share. Who will benefit the most and where is the trap?

I’ve always said the best opportunities in the tech sector aren’t where the biggest spotlights shine. Nvidia owns the headlines, but the real litmus test of the AI era sits one floor down — in memory chips. And right now one of the largest cycles of the past 15 years is unfolding in that segment.

Let’s look at what’s actually happening, who has the advantage, and where the real spot is for an investor thinking beyond the next quarter.

Vera Rubin: a monster that changes the rules of the game

In March 2026 at GTC Nvidia confirmed that its new platform Vera Rubin (R200) carries 288 GB HBM4 memory on a single GPU — that’s the same as Blackwell B300, but with 62.5% higher bandwidth (13 TB/s vs 8 TB/s).

That alone wouldn’t be so dramatic. The shock comes with Rubin Ultra (R300) in H2 2027:

| Parameter | Blackwell B300 | Rubin R200 (H2 2026) | Rubin Ultra (H2 2027) |

|---|---|---|---|

| HBM capacity / GPU | 288 GB HBM3e | 288 GB HBM4 | 1,024 GB HBM4e |

| HBM stacks | 8 (12-Hi) | 8 (12-Hi) | 16 (16-Hi) |

| Bandwidth | 8 TB/s | 13 TB/s | ~32 TB/s |

| FP4 performance | – | 50 PFLOPS | 100 PFLOPS |

3.5× more memory per GPU within 12 months. That is an unprecedented step in the semiconductor industry.

And now the question few ask: where will that memory come from?

Let’s run the numbers. And a problem emerges.

Global HBM bit demand grows according to UBS and Tech Investments like this:

- 2025: ~12.6 billion Gb

- 2026: 19.0 billion Gb (+51% YoY)

- 2027: 26.3 billion Gb (+38% YoY)

If I use a conservative estimate of 5 million Rubin R200 units in 2026 (NVIDIA targets 20M Blackwell+Rubin GPUs by EOY 2026, with Rubin ~22–29% of that), Rubin alone would consume 11.5 billion Gb of HBM — i.e. approximately 60% of global HBM capacity.

Yes, you read that right. One single product line from one GPU vendor will consume more than half of global HBM production.

And if we add Rubin Ultra with 1 TB of memory by 2027, we reach numbers where the Rubin family in a theoretically pessimistic scenario would exceed 100% of today’s capacity. That’s why we see fights over 5‑year contracts, OpenAI negotiating 900,000 wafers/month commitments for Stargate (40% of global DRAM production!), and hyperscalers lining up at SK Hynix.

Three players who must handle it — and each has a different story

SK Hynix: the king that’s slightly faltering

In Q1 2025 SK Hynix for the first time in history overtook Samsung in the total DRAM market — 36% vs 34%. Reason? Dominance in HBM, where it held 59% bit share in 2025.

But here’s the problem for SK Hynix stock: Samsung has finally caught up on development. Bloomberg in January 2026 confirmed Samsung passed Nvidia qualification for HBM4, and in February it started mass production. Consequence:

- 2026: SK Hynix drops to 50–53% bit share

- 2027: further to 49%

Capex is nonetheless phenomenal: $13.3 billion just on equipment in 2026. The M15X fab in Cheongju ramped 4 months earlier than planned, and by the end of 2026 SK Hynix will have 620,000 wafers/month DRAM capacity, of which 30% will go to HBM. In 2027 that will be 40% to HBM — a huge shift.

The megaproject Yongin cluster Y1 will enter production in H2 2027 with 150,000 wafers/month capacity (and can reach 300k+ overall).

Samsung: a comeback with a question mark

Samsung was out in the HBM3e era. Yields were poor, and Nvidia qualification took years. But on February 12, 2026 they announced the world’s first commercial mass production of HBM4 with 11.7 Gb/s pin speed and 3.3 TB/s bandwidth — beating the JEDEC standard by a solid margin.

The plan is dramatic: increase HBM capacity from 170,000 wafers/month to 250,000 wafers/month by end‑2026 (+47%). Fab P5 in Pyeongtaek resumed construction, and mass production is expected in 2027.

Risk: Yield for the 1c node HBM4 is, according to [TweakTown](https://www.tweaktown.com/news/108316/samsung-1c-dram-for-hbm4-yields-rumored-to-hit-around-50-percent-to-battle-sk-hynix-and-micron/index.html), only around 50%, while SK Hynix is already at ~80%. If Samsung doesn’t close that gap by H2 2026, the 2027 thesis is at risk.

Micron: the outsider with the best product

In March 2026 Micron announced high‑volume production of HBM4 36GB 12‑Hi designed directly for Nvidia Vera Rubin — 2.8 TB/s bandwidth, with 20% better power efficiency than HBM3e.

Here’s the catch: Micron has the better product but lacks wafer capacity. That’s why Nvidia gives it only Rubin CPX (a mid‑tier inference accelerator) and not the flagship Vera Rubin. The big leagues will be split between Samsung and SK Hynix.

Still:

- HBM capacity 2026 sold out

- 5‑year contracts into 2027+

- CEO Sanjay Mehrotra: “We can satisfy only 50–66% of the demand of our key customers”

- Capex rising to ~$14 billion

- Singapore HBM packaging meaningful by CY 2027, Idaho fab FY 2027

In 2026 Micron is expected by analysts to deliver +100% revenue growth YoY, and its HBM market share sits between 19–22%.

What’s happening with regular DRAM and NAND?

This is the part most analysts overlook. HBM has a 3:1 trade‑ratio versus DDR5 — producing one HBM4 wafer consumes the same capacity as 3 DDR5 wafers. For future HBM5 (2028+) the ratio will be even worse.

Consequence? Conventional DRAM is bleeding:

- Goldman Sachs raised its forecast for DRAM price increases in 2026 from 150% to 250–280% YoY

- NAND from 100% to 200–250% YoY

- IDC: DRAM supply growth in 2026 only +16% YoY (historical average 25–30%)

- DDR4 goes to the grave — wafer starts at Samsung and SK Hynix fall in H2 2026 to low single digits

The NAND story is similar:

- Western Digital has HDD capacity sold out for all of 2026

- Solidigm ships 122 TB QLC SSDs, and 245 TB drives are coming in 2026

- Phison CEO warns NAND shortage could last up to 10 years

- Kioxia is ramping 332‑layer NAND in Kitakami in 2026, capex +41% to $4.5 billion

Main question: underbuilt or overbuilt?

Goldman Sachs says it bluntly: memory chips undersupplied at least until H1 2027, the worst DRAM deficit in 15 years (-4.9% in 2026, -2.5% in 2027).

SK Group Chairman Chey Tae‑won goes further — wafer shortage potentially into 2030.

Timeline — how it looks today:

```

2025 ████████░░ Tight, prices rising

2026 ██████████ DEFICIT — everything sold out, prices soaring

2027 ████████░░ Still deficit, new fabs ramping

2028 ██████░░░░ Possible balance — BUT Rubin Ultra will tighten demand

```

Key risk: If Yongin Y1, Samsung P5 and Micron Idaho/Singapore all ramp at once by 2028 and AI capex at hyperscalers slows, we could see a classic memory cycle similar to 2022–2023, but from a much higher base. This cyclicality is notorious in the memory sector and investors should not ignore it.

Practical view: what this means for portfolios

This is no longer academic. Let’s go to the tickers:

SK Hynix (000660.KS) — best positioned HBM player, but losing share. SK Securities’ 3M KRW target assumes Samsung fails. If Samsung succeeds, expect a downside re‑rating.

Micron (MU) — my favorite for investors with higher risk tolerance. Best product on the market, fully sold out, +100% revenue growth expected in 2026. Wall Street consensus per [Investing.com](https://www.investing.com/analysis/micron-analysts-see-more-upside-as-hbm-demand-surges-200673271) still only sees ~20% growth — materially undervalued.

Samsung (005930.KS) — a volatility play. If HBM4 ramp in H2 2026 works, we could see a strong re‑rating. If not, the stock stays in sideways consolidation.

Nvidia (NVDA) — paradoxically the memory bottleneck reduces its near‑term GPU shipments (Rubin share in 2026 fell from expected 29% to 22%). Long‑term thesis remains intact.

Kioxia (285A.T) — fresh IPO, BiCS9 332‑layer NAND is technologically ahead. For those who believe in a NAND shortage.

Equipment plays — ASML, Applied Materials, Lam Research, ASM International. The memory capex supercycle = their golden era. Less volatile exposure than direct memory names.

Solidigm/SK Hynix via parent — for NAND/eSSD exposure in AI datacenters.

Conclusion — key takeaways

1. Vera Rubin is a structural driver of the HBM supercycle at least into 2027. I don’t see any “peak AI capex” here yet.

2. The market is dramatically undersupplied — all three main HBM suppliers are sold out for 2026, prices are rising, supply gap is the worst in 15 years.

3. Market shares are being redistributed — SK Hynix falls from 59% to ~50%, Samsung returns with a yield question, Micron holds ~20% but is growing fastest fundamentally.

4. Watch 2028 — everyone is building at once. If AI capex slows, a classic memory cycle will hit from a much higher base. That is the biggest risk for long positions with a 3+ year horizon.

5. Equipment players may be less sexy but offer a less volatile path to the same thesis.

For me this is one of the clearest macro setups in the tech sector. The question isn’t “if the memory supercycle continues,” but “how long and where is the peak.” We’ll have to wait for the answer exactly as long as Samsung struggles with HBM4 yields.

Disclaimer: This article is an investment opinion, not investment advice. Do your own analysis and consider your risk preferences before making any decision. The author may have personal exposure in some of the positions mentioned.