AI Silicon Stack — 4 Bottlenecks Dissected: Micron, Amkor, Broadcom, Marvell

Before an Nvidia GPU even gets racked in a data center, four other components matter more than the GPU itself. And all four are supply-constrained right now.

This piece dissects the four bottlenecks of the AI silicon stack — HBM memory, CoWoS packaging, custom AI silicon, and optical transceivers — one layer at a time. For each layer, who's the leader and why that leader is asymmetrically positioned today.

The reason I bunch these four together: block any one of them and Nvidia stops working too. Nvidia's margin expansion story is, in practice, the same story playing out across these four names at the same time.

Layer 1: HBM Memory — Micron's Asymmetry

Three companies make all the world's HBM. SK Hynix 62%, Micron 21%, Samsung 17%. There is no fourth option.

HBM is the memory that physically stacks on top of every Nvidia AI chip. The speed at which the chip reads and writes data during training and inference is almost entirely a function of HBM bandwidth. That's why GPU specs always quote things like "HBM3e 192GB."

All three suppliers are sold out for 2026 production, with waitlists already running into 2027. Demand exceeds what three companies can produce. The entire AI build-out waits on HBM availability.

This is where the Hormuz variable changes the picture. If Qatar's helium exports wobble, Korean fabs pay up for helium. SK Hynix and Samsung run heavily on Korean fabs. Micron runs on US fabs with domestic helium reserves.

The result is one of two outcomes — Micron picks up share, or the entire market price expands. Both expand Micron's margins.

Revenue grew 56% last year, EPS grew 175%. A PEG ratio of 0.25 says the stock is irrationally cheap relative to growth. This is the fastest-growing name on my AI list.

Layer 2: CoWoS Packaging — Amkor's Overflow

CoWoS (Chip on Wafer on Substrate) is the final assembly process for Nvidia AI chips. The compute die, HBM memory, and interconnects all get bonded into a single high-density module. Without CoWoS, the chip doesn't exist.

TSMC does most of this work. But Nvidia has already locked up more than 50% of TSMC's CoWoS capacity. Which means Broadcom, AMD, every custom ASIC builder needs a second source.

That second source is Amkor. Their Arizona facility, partly funded by the CHIPS Act, sits adjacent to TSMC's US fabs. Advanced packaging revenue is expected to nearly triple this year alone — overflow demand from TSMC flows directly to Amkor.

CoWoS demand is growing 80% annually. Supply only grows 50%. That gap doesn't close in 2026.

Layer 3: Custom AI Silicon — Broadcom's Two Layers

Broadcom entered 2026 with a $73B AI backlog — locking in 18–24 months of revenue with the five largest AI spenders before the year even started.

What makes Broadcom unusual is owning two pieces of the networking silicon layer simultaneously.

First, custom AI accelerator design. When Google, Meta, Amazon, and Microsoft decided they wanted chips optimized for their specific workloads — instead of buying Nvidia off the shelf — they all went to Broadcom. Google's TPU, Meta's MTIA, Amazon's Trainium are all co-designed with Broadcom's teams. That's 60–70% of the custom AI chip market.

Second, Ethernet switching. The fabric connecting every server inside the data center. The competing-standards battle is settled — Open Ethernet won. Broadcom's Tomahawk 6 controls 80% of that market. When a hyperscaler builds a million-GPU cluster, every GPU communicates through Broadcom silicon.

AI revenue last quarter was $8B — 43% of total. Guidance is up another 27% next quarter. AI is no longer a footnote at Broadcom; it's the core business.

Layer 4: Optical Transceivers — Marvell's Optical DSP

63 million 800G+ optical transceivers will ship this year. A 2.6x jump from last year. There's a Marvell chip inside every single one.

That chip is an optical DSP. It's the processor that converts electrical signals from a server into light pulses traveling through fiber. The market is effectively a Marvell-Broadcom duopoly, and Marvell holds the #1 spot.

Why is every hyperscaler suddenly going optical? Physics. At 224 Gbps signaling speeds, copper cable can't reach 1 meter. And it consumes 30% of total cluster power. Optical (photonics) cuts that by 70%. The data center industry doesn't get a choice — they have to make the shift.

Four years ago, Marvell was just an Ethernet chip company. They spent $10B in 2021 acquiring Inphi and built the optical silicon position from scratch. EPS tripled in a single year on a business segment that barely existed when that deal closed.

Putting the Four Layers Together: Where's the Highest Beta?

Stock	Market position	Key trigger	Beta as I see it
Micron	21% HBM	Helium/Korean fab cost asymmetry	Very high
Amkor	#2 CoWoS	TSMC overflow demand	High
Broadcom	60–70% custom chips	Hyperscaler capex	Medium
Marvell	#1 optical DSP	800G+ transition speed	High

My personal read — Micron is the most undervalued at PEG 0.25, Broadcom is the safest with $73B in backlog, Marvell carries the most asymmetric upside as the optical cycle just begins, and Amkor is the highest-volatility bet of the four.

Risks / Counterarguments

The two counterarguments I take most seriously:

First, cyclicality. Semiconductors are a cycle industry. HBM sellouts and CoWoS shortages don't last forever. New capacity coming online in 2027–2028 normalizes pricing. My base case is that the next 18–24 months hold a margin-expansion window between now and that normalization.

Second, Nvidia dependence. All four names are partially tied to Nvidia revenue. If Nvidia itself wobbles, they wobble together. That said, Broadcom's custom chip mix and Marvell's optical mix both show Nvidia dependence trending down structurally.

Accept those two risks and I'd argue all four are sitting where the highest growth happens over the next five years.

FAQ

Q: Could a fourth HBM supplier emerge? A: Highly unlikely in the short term. HBM mass production has very high capital and technical entry barriers. Standing up a new fab and ramping yield takes 4–5 years minimum. Don't expect a fourth option before 2030.

Q: Can Amkor's CoWoS revenue really triple? A: That's company guidance. But the base is small, so absolute revenue is still a fraction of TSMC's. The bigger the percentage growth, the bigger the proportional execution risk.

Q: Won't optical transceivers eventually return to a normal cycle? A: Yes — but the 800G → 1.6T → 3.2T transition, once it happens, is hard to reverse. And every step lifts both unit volume and ASPs simultaneously. I see normalization coming after 2028.

Q: Doesn't Broadcom look expensive right now? A: Agreed. But run the math on the $73B backlog as forward revenue and the PEG actually looks reasonable. If the multiple still bothers you, dollar-cost averaging or waiting for a pullback is fine.