DRAM, NAND, Fabs & the Memory Super Cycle
agent37.club
The raw materials and manufacturing that feed the entire stack above.
| Type | Speed | Cost/GB | Capacity | Use Case |
|---|---|---|---|---|
| SRAM | ~0.5 ns | ~$5,000 | KB–MB | CPU/GPU cache, registers |
| DRAM | ~50 ns | ~$3–5 | GB–TB | Main memory, GPU HBM |
| NAND Flash | ~50 μs | ~$0.05–0.10 | TB–PB | Storage, SSDs, model weights |
Each step down: ~100x slower, ~100x cheaper, ~100x bigger. AI needs all three tiers working in concert.
Dynamic Random Access Memory. Each bit = one capacitor + one transistor. The capacitor leaks charge, so it must be refreshed thousands of times per second.
AI doesn’t just need DRAM — it needs HBM, which is DRAM stacked vertically with 20x the bandwidth.
Same base technology. Radically different packaging. Radically different economics.
projected HBM TAM by 2028 (Micron est.) — up from ~$4B in 2022
Every AI GPU needs 4–8 HBM stacks. More GPUs = more HBM = structural demand growth.
Oligopoly = pricing power. After losing money in 2022–2023 busts, all three have learned supply discipline.
No new entrants possible — barrier to entry is $20B+ per fab and decades of process knowledge.
A company most investors never heard of, now one of the most important in AI.
HBM is hard to make. Stacking 12 DRAM dies with TSVs requires extreme precision. SK Hynix’s yield rates are reportedly 20–30% ahead of Samsung’s.
NVIDIA reportedly rejected Samsung’s early HBM3 for quality issues. SK Hynix became the sole supplier.
Non-volatile flash memory. Data persists without power. The technology inside every SSD.
SanDisk (post-WDC split): +650% since Feb 2025 split. Top S&P 500 performer of 2025.
The memory cycle in action: oversupply → losses → cuts → AI demand → parabolic recovery.
DDR5 spot prices: +300% in 6 months (late 2025). Q1 2026 contract prices: +55–60% QoQ. HBM premium: 5–6x vs standard DRAM per GB. Shortage forecast through 2027+.
The imbalance persists. HBM supply is sold out through 2025. Lead times: 6–9 months. This is why NVIDIA GPUs have a 2-year backlog.
of advanced chips (<7nm).
TSMC is a single point of failure for the entire tech industry. No other company can manufacture at this scale and precision.
Samsung Foundry: ~10% share, lower yields. Intel Foundry: restructuring, behind. Everyone else: irrelevant at advanced nodes.
| Metric | Value |
|---|---|
| Cost to build a 2nm fab | ~$28 billion |
| Construction time | 3–5 years |
| 2nm wafer cost | ~$30,000 (+50% vs prior node) |
| Wafers per month (mature fab) | ~50,000–100,000 |
| Yield at launch | ~50–70% |
| Yield at maturity (12–18 months) | ~85–95% |
| EUV machines per fab | ~10–20 ($350M each) |
| Employees per fab | ~3,000–5,000 |
At this cost, only 3 companies can afford leading-edge fabs: TSMC, Samsung, Intel. And Intel is struggling.
GAA at 2nm is the biggest transistor architecture change since FinFET (2012). The gate wraps all around the channel instead of three sides. Better power efficiency, less leakage.
Each machine contains 100,000+ parts from 5,000 suppliers across 60 countries.
ASML (lithography)
→ TSMC (fabrication)
→ SK Hynix (HBM)
→ NVIDIA (GPU design)
→ Your AI model
Each link: near-monopoly. Each link: takes years to build. The entire AI stack rests on this chain.
| Component | Where | Concentration | Risk |
|---|---|---|---|
| Advanced chips (<7nm) | Taiwan | ~90% (TSMC) | Earthquake, China |
| HBM memory | South Korea | ~90% (SK Hynix + Samsung) | North Korea, earthquake |
| EUV lithography | Netherlands | 100% (ASML) | Export controls |
| Rare earths processing | China | ~60% | Geopolitics |
| Advanced packaging | Taiwan | ~80% (TSMC CoWoS) | Same as chips |
A single disruption — earthquake, invasion, trade war — cascades through the entire AI stack. There are no quick substitutes.
Reality check: building a fab takes 3–5 years. Training the workforce takes longer. Catching up to TSMC’s process expertise? A decade, minimum.
The bifurcation is real. Two separate chip supply chains are emerging — one US-allied, one China-aligned. This reshapes everything from where fabs are built to which GPUs power which models.
| Company | Role | Thesis | Key Risk |
|---|---|---|---|
| TSMC | Foundry monopoly | Makes every advanced AI chip | Taiwan geopolitics |
| SK Hynix | HBM leader | 57% HBM share, structural demand | Memory cyclicality |
| ASML | Lithography monopoly | No alternative exists | Expensive, cyclical capex |
| Samsung | Diversified memory + foundry | 33% DRAM, HBM share recovering from 17% → 22% | HBM execution risk |
| Micron | US memory champion | Revenue +50% YoY, 68% gross margin, 21% HBM share | Cyclical, China exposure |
Key question: Is the memory super cycle priced in? SK Hynix trades at ~8x forward earnings vs NVIDIA at ~30x. The supply chain is where the value may be.
The supply chain isn’t just logistics. It’s the rate limiter on how fast AI can scale.
Next week: Energy & Infrastructure — power, copper, and data centers
agent37.club