By Edgen Research | 2026-05-15
Tickers: $NVDA, $AMD, $MU, $SNDK, $AVGO, $MRVL, $POET, $LITE, $COHR, $TSM
SanDisk is up 3,362% over the past 12 months (Yahoo Finance). Micron is up around 340% in the same window (Yahoo Finance). Both are AI hardware. Neither is NVIDIA.
If you call yourself "long AI" but own only NVDA, you're not long AI — you're long one basket, and the other four ran without you. SNDK and MU are the proof of how expensive that miss can be.
Here's the map. Five baskets. Where I'd put new money today (Basket 4) and where I wouldn't (Basket 1). Bottom line first.
The call: add to Basket 4
If you can only add to one basket today, I'd add to Basket 4 — Optical Interconnect. Not NVDA. Not HBM. Not foundry.
One sentence: AI's bottleneck is shifting from "memory can't feed the GPU fast enough" (the 2024 problem) to "the wires can't connect 100,000 GPUs at once" (the 2026-2028 problem). The market hasn't fully priced this yet. AI-themed funds barely own any of the optical names.
I wouldn't add to Basket 1 (NVDA, AMD) at today's prices. Everyone owns NVDA. Buying another share doesn't tell you anything new.
Basket 1 — GPU: everyone owns it, probably overweight
Don't add at these prices. NVDA is the most-watched stock in tech. Holding one more share gives you no edge. AMD (AMD) is the credible second name, but CUDA software lock-in keeps real share shift slow — bulls have been promising the AMD pivot for years and it keeps not happening at scale.
If you already own NVDA (NVDA) in size, you're positioned correctly. Adding more is harder to justify than rotating into a basket the market hasn't figured out yet.
Basket 2 — HBM Memory: bottleneck already paid out
Hold, don't add. HBM is real — SK Hynix dominates, Micron is the only US name, Samsung is the third player (Astute Group). But this story is done — MU is up 340% over 12 months precisely because the market figured it out in 2024-2025.
The US-actionable name is Micron (MU). SanDisk (SNDK) is the NAND adjacency that's run even harder. Our Micron vs SanDisk piece explains why pricing power held this cycle. But it's all in the price now.
Basket 3 — Custom Chips (ASIC): the quiet compounder
Mild add if you have no exposure. Cloud companies (Google, Amazon, Meta) increasingly design their own chips instead of just renting GPUs. They hire Broadcom (AVGO) and Marvell (MRVL) to build them.
The math is simple — at $50B+ a year in AI spending, designing a chip that saves 30% on your specific workload eventually pays for itself. AVGO isn't cheap, but this basket gets less attention than GPU, which is what makes it interesting.
Basket 4 — Optical Interconnect: add here first
This is the underpriced basket today. Add here before anywhere else.
Here's the simple version: today's training clusters use copper wires to connect thousands of GPUs. Push past 100,000 GPUs and the wires literally can't keep up — too much heat, too much delay, too many cables. Optical chips (using light instead of electricity) fix this. It's the same setup HBM had in 2024 — went from theoretical concern to actual constraint.
The US-actionable names: Lumentum (LITE) and Coherent (COHR) — both already shipping optical components to cloud data centers, with broad customer bases. POET Technologies (POET) is the pure-play, but our prior why POET crashed flagged it as the cautionary tale of betting on one customer. The lesson: in optical, pick diversified names (LITE / COHR), not single-customer pure-plays.
Where this could be wrong: if cloud companies slow their cluster build-out past 100K GPUs (real risk — power grid limits), or if optical deployment slips another 2-3 years, this basket dead-money-ifies through 2027. It's a 2026-2028 bet, not a 2026 trade.
Basket 5 — Foundry: TSMC, that's it
Own TSM. End of basket. Every chip in baskets 1-4 has to be physically built somewhere. At the leading edge, TSMC has no real competitor — Samsung Foundry struggles, Intel Foundry is years away.
TSMC (TSM) is fully booked through 2026. Apple, NVIDIA, AMD have all reserved capacity. The 2026 problem isn't demand — it's how fast they can build the next round of factories (TechNode).
If you don't have TSM, buy it. If you do, you're done with this basket.
The 5 baskets at a glance
| # |
Basket |
What it does |
Lead names |
My call today |
| 1 |
GPU |
The compute engine |
NVDA, AMD |
Hold, don't add |
| 2 |
HBM Memory |
Feeds the GPU |
MU, SNDK |
Hold; already ran |
| 3 |
Custom ASIC |
Cloud-company-specific chips |
AVGO, MRVL |
Mild add if no exposure |
| 4 |
Optical Interconnect |
Connects clusters at scale |
LITE, COHR |
ADD HERE FIRST — underpriced |
| 5 |
Foundry |
Builds everything |
TSM |
Core hold; one name |
What this map is honest about
It's not stock-picking — naming who plays in each basket is different from saying any single stock is mispriced today. Run valuation separately. The map tells you what universe to run it on. It's also not comprehensive — we left out networking silicon, liquid cooling, power infrastructure (see our AI physical infrastructure overview for those layers). Basket boundaries will also move over the next 12 months — score this against reality in May 2027.
If you disagree with the Basket 4 call, the disagreement should be specific: about deployment timing, about LITE / COHR execution, about how fast optical actually deploys. Specific disagreement is how the framework gets better.
This is research and education, not personalized investment advice. Edgen Research is the institutional byline for cross-coverage synthesis pieces. Every quantitative claim is sourced in-line. Verify against current data before acting. Edgen and contributors may hold positions in the securities discussed.
