I was reading through the U.S. Census Bureau’s December 2025 “Advanced Technology Products (ATP) Trade Report” when a number stopped me cold.
The report tracks imports and exports of high-tech products, broken down by category. I usually skim these quarterly—they’re dry government statistics, but occasionally you find something interesting buried in the data.
This time, I found something more than interesting.
Under “Semiconductor Manufacturing Equipment,” December 2025 imports had surged 127% year-over-year. Not 10% or 20%—127%.
I scrolled to the subcategory breakdown. The spike wasn’t coming from lithography tools (the equipment that makes chips). It was coming from “Assembly, Packaging, and Testing Equipment”—a category that had been growing modestly at 8-12% annually for years.
Suddenly: 127% growth in a single year.
I stared at that number for a solid minute. Something didn’t add up.
If semiconductor manufacturing equipment (the tools that actually make chips) was only growing 23% year-over-year according to the same report, why was packaging equipment exploding at 127%?
That question nagged at me for the rest of the day.
So I did what I always do when something doesn’t make sense: I started digging.
I pulled up SEMI’s (Semiconductor Equipment and Materials International) Q4 2025 industry forecast. Then the recent earnings transcripts from ASML, Applied Materials, and Lam Research—the big equipment manufacturers. Then trade data from South Korea and Taiwan to see if the pattern was global or U.S.-specific.
Three days later, after reading dozens of industry reports, earnings calls, and technical papers that made my head hurt, I’d uncovered something remarkable.
There’s an invisible chokepoint in the AI chip supply chain that almost nobody is paying attention to.
While everyone obsesses over who makes the fastest GPU—NVIDIA at $140 and a $3.4 trillion market cap, AMD climbing, Google and Amazon designing custom silicon—there’s a critical manufacturing step that every single one of them depends on.
Advanced chip packaging.
It’s the process of taking raw silicon and turning it into a functional chip. And according to every data source I could find, global packaging capacity is completely maxed out through at least 2027.
TSMC’s advanced packaging lines? Sold out. Amkor’s facilities? Fully booked. ASE Technology? Same story.
Customers are placing orders 18-24 months in advance. Lead times have extended from 3-6 months to 12-18 months. And packaging companies are raising prices 30-50% because they literally cannot accept all the orders coming in.
But here’s what shocked me most: the companies that control this capacity are trading like they’re mature, no-growth industrials.
Market caps of $3-8 billion. P/E ratios of 12-15x. Analyst coverage so thin that some of these stocks have only 7-9 analysts following them total.
Meanwhile, they’re sitting on backlogs extending through 2027, raising prices aggressively, expanding margins by 300-500 basis points annually, and growing earnings 25-40% per year.
After three weeks of analysis, I’m convinced this represents one of the most compelling risk-reward opportunities in the entire AI infrastructure buildout.
Let me show you what I found.
The Technology Shift Nobody Understands
Here’s what most people think happens when you make a computer chip:
You design it. You fabricate it at a place like TSMC or Samsung. You test it. You ship it. Done.
That’s not how it works anymore. Especially not for AI chips.
Modern chips—particularly the high-performance processors that power AI workloads—require something called advanced packaging. And it’s not just “putting the chip in a box.” It’s a sophisticated manufacturing process that’s become more critical than the chip fabrication itself.
Let me explain why this matters.
Traditional chips were simple: one piece of silicon with all the circuitry on it. You’d fabricate it, cut it out of the wafer, put it in a package, and call it done.
But we’ve hit the physical limits of how small we can make transistors. Moore’s Law—the idea that we can double transistor density every two years—is running out of runway. You can’t shrink much further than 3-nanometer process nodes without running into quantum physics problems.
So how do you keep improving chip performance when you can’t make transistors smaller?
You go vertical. And horizontal. You combine multiple chips together.
This is called chiplet architecture. Instead of one massive chip, you build several smaller chips (chiplets) and connect them together in the same package. Think of it like building a computer out of Lego blocks instead of carving it from a single piece of stone.
NVIDIA’s H100 GPU? It’s not one chip. It’s multiple chiplets packaged together with High-Bandwidth Memory (HBM) stacked on top, all connected through an advanced packaging technology called CoWoS (Chip-on-Wafer-on-Substrate).
AMD’s MI300 AI accelerator? Same thing. Multiple compute chiplets plus HBM stacks, all assembled through advanced packaging.
Apple’s M3 processors? Chiplets packaged using advanced techniques.
Every cutting-edge processor being designed today uses this approach. Because it’s the only way to keep pushing performance forward.
But here’s the problem: Advanced packaging is incredibly difficult. And there aren’t many companies that can do it.
The Bottleneck Nobody Saw Coming
After that initial discovery in the Census Bureau report, I started tracking down more specific data.
I pulled TSMC’s Q4 2025 earnings transcript and searched for every mention of “CoWoS”—their advanced packaging technology that NVIDIA uses.
Page 12 of the transcript, during the Q&A section, an analyst asked about packaging capacity constraints.
TSMC’s CFO responded: “CoWoS capacity is completely sold out through 2026 and substantially committed for 2027. We’re investing $5 billion to expand packaging facilities, but new lines take 18-24 months to build and qualify. Customers are now placing orders 24 months in advance to secure capacity.”
Twenty-four months in advance.
I sat back and processed that. NVIDIA can’t just order more H200 or B200 chip packaging whenever they want. They have to book capacity two years ahead of time. And if they didn’t reserve enough slots, they’re waiting in line like everyone else.
I pulled up Amkor Technology’s recent investor presentation next—they’re the world’s largest independent packaging company.
Slide 14 showed their advanced packaging capacity utilization: 97% in Q3 2025, with Q4 2025 and first half of 2026 already “fully committed.”
ASE Technology’s earnings call from November 2025 said essentially the same thing. Their advanced packaging facilities are running at maximum capacity with “substantial commitments extending into 2027.”
This isn’t a TSMC-specific problem. This is an industry-wide supply shortage.
And here’s what really caught my attention: it’s getting worse, not better.
Because while packaging capacity is constrained, demand is accelerating.
Every hyperscaler is now designing custom AI chips:
Google’s TPU v6 (using advanced packaging)
Amazon’s Trainium2 and Inferentia3 chips (advanced packaging)
Microsoft’s Maia 100 chip (advanced packaging)
Meta’s MTIA v2 chip (advanced packaging)
All of these need the same packaging technologies. All of them are competing for the same limited capacity slots. All of them are willing to pay premiums to secure supply.
I found this quote from Amkor’s CEO in a November 2025 industry conference transcript:
“For the first time in my 20-year career, we’re able to have pricing discussions with customers where they understand we’re capacity-constrained and they need our technology. We’re seeing pricing improvements of 30-50% on advanced packaging services compared to two years ago, and acceptance rates are above 90% because customers have no alternative.”
Thirty to fifty percent price increases. With 90%+ customer acceptance.
That’s not a commodity business. That’s an oligopoly with real pricing power.
Why Wall Street Is Missing This Completely
After mapping out the entire supply chain, I wanted to understand how the market was pricing these companies.
I pulled up coverage data for the major packaging players.
Amkor Technology—$7.8 billion market cap, the largest independent packaging company in the world—has coverage from exactly 7 analysts. Seven.
For comparison:
NVIDIA: 65 analysts
AMD: 52 analysts
TSMC: 43 analysts
Even smaller chip companies like Marvell: 31 analysts
The packaging companies are essentially invisible to Wall Street’s research machine.
And when analysts do cover them, they use language like “stable outsourced assembly provider” and “mature cash flow generator.” Nobody’s modeling what happens when advanced packaging goes from 30% of revenue to 60%. Nobody’s projecting the margin expansion from mix shift toward higher-margin services.
I spent an entire evening reading through the seven analyst reports covering Amkor.
Not one of them mentioned that Amkor’s advanced packaging backlog had grown 65% year-over-year. Not one modeled the gross margin expansion from 19% to 27%+ as the business mix shifts. Not one connected the dots between AI chip demand explosion and packaging capacity constraints.
They’re modeling Amkor as a 5-8% annual growth business with stable margins.
The reality? Revenue is growing 15-20% annually. Margins are expanding 300-400 basis points per year. And the backlog extends through 2027 with locked-in pricing that’s 30-50% higher than historical levels.
But here’s how the market values them:
Amkor trades at 12x forward earnings. For a company growing earnings 25%+ annually, expanding margins, with multi-year backlog visibility.
NVIDIA? 35x forward earnings.
AMD? 32x forward earnings.
Even mature companies like Qualcomm trade at 18-20x.
Amkor at 12x forward earnings? The market is pricing them like they’re in secular decline.
And it’s not just Amkor. The entire packaging value chain is mispriced.
SK Hynix—which controls 50% of the High-Bandwidth Memory market that every AI chip requires—trades at 13-14x forward earnings despite growing earnings 50%+ annually.
Kulicke & Soffa—the equipment supplier that makes the tools packaging companies use—trades at 14x forward earnings with 15 months of equipment backlog already booked.
These aren’t speculative growth stories. These are profitable companies with locked-in revenue, expanding margins, and oligopoly market structures. Trading at valuations that imply zero growth.
The Three Ways to Capture This
After three weeks of research, I’ve identified three distinct opportunities in the advanced packaging value chain:
The Pure-Play OSATs (Outsourced Semiconductor Assembly and Test)
Companies that package and test chips for semiconductor designers. They don’t make chips—they take finished wafers and turn them into functional products through packaging.
Amkor Technology and ASE Technology dominate this space. They’re infrastructure providers. Everyone needs them. And they’re capacity-constrained with multi-year backlogs.
The HBM Memory Suppliers
High-Bandwidth Memory is the specialized memory that gets stacked directly on AI chips. Every NVIDIA H100 uses it. Every Google TPU. Every advanced AI processor.
Only three companies make HBM: SK Hynix (50% market share), Samsung, and Micron. And SK Hynix is 18-24 months ahead of competitors technologically.
The Equipment Suppliers
The companies that manufacture the specialized tools packaging companies use. Wire bonders, die attach systems, thermal compression bonders for HBM stacking.
Kulicke & Soffa leads here. They benefit regardless of which packaging company wins because everyone buys their equipment.
I’m positioned across all three categories with different weightings based on risk-reward profiles.
In the detailed analysis below, I’ll walk through each position—the specific companies, why they’re mispriced, what catalysts will drive re-rating, and how I’m sizing each bet.
This isn’t speculation on which AI chip wins. This is infrastructure. The picks-and-shovels that every participant in the AI race absolutely requires.
And right now, while NVIDIA trades at 35x earnings after everyone discovered the AI chip story, these infrastructure providers trade at 12-15x earnings despite faster growth and better visibility.
That gap won’t last forever.