From Handsets to Hyperscalers, One Package at a Time

CoWoS: what it is, why it became a bottleneck, where it sits

For most of the chip industry's history, packaging - taking the finished die and wrapping it into something that can be mounted on a board - was, compared to the fab, relatively routine. That is starting to no longer be the case. Modern AI accelerators have developed from single pieces of silicon into chiplet designs: several compute dies and stacks of high-bandwidth memory (HBM) wired together inside one package at very high density. Getting those parts to communicate at the bandwidth the chip requires has become its own engineering problem, making assembly increasingly difficult.

CoWoS is the process that does this at the high end. The name stands for Chip-on-Wafer-on-Substrate, TSMC's advanced-packaging platform for mounting compute dies and HBM onto a shared interposer. The most advanced variant, CoWoS-L, is what today's leading accelerators depend on. This is not capacity that can be added quickly. It requires equipment with long lead times and engineering effort to qualify each process, which is why packaging supply cannot simply scale up on demand.

The production chain runs in a fixed order: a chip is designed, fabricated into wafers, packaged and tested, then shipped as a finished part to whoever builds it into a system. Packaging and test is the final manufacturing step. Once a chip has been packaged, it does not go back to the fab - it moves onward to the customer. That ordering is worth keeping in mind.

TSMC's chokehold and the memory delay

TSMC owns the frontier tier of packaging almost entirely, and the demand numbers show how completely. Of Nvidia's projected CoWoS demand for 2026 - somewhere around 595,000 wafers on current estimates - roughly 510,000 are set to run through TSMC, and Broadcom's custom-ASIC demand of about 150,000 wafers sits largely there as well. Nearly all of the frontier work funnels to one company. TSMC is expanding to meet it, lifting CoWoS capacity from around 75,000 wafers a month in 2025 towards 115,000 to 140,000 by the end of 2026, yet even that ramp trails what its customers are asking for. The shortfall is not only physical. TSMC runs advanced packaging as an extension of its fab business for its most important customers, and its priority lies upstream, in protecting the wafer capacity where the returns are richest. Packaging demand it cannot comfortably serve, it is content to let go elsewhere.

For now, the packaging constraint is masked by the one ahead of it. A packaging house marries compute dies to stacks of HBM, so until the memory is there, there is nothing to assemble. So while memory is the binding shortage, packaging cannot bite as hard as it eventually will. The effect is already visible in the sector's results, where some packaging revenue has been pushed out by customers waiting on memory. As that shortage eases, the constraint does not disappear. It moves downstream, onto packaging.

Where does the overflow go?

That overflow has a natural home: the OSAT, or outsourced semiconductor assembly and test, a company whose entire business is packaging and testing chips that other firms have designed and fabricated. The customer's finished wafers arrive, the OSAT assembles the dies and memory stacks into a package, tests the result, and ships it onward. It is packaging and test done for hire.

Two firms dominate this independent layer, and both are full-service, selling to broadly the same set of chip designers. ASE, out of Taiwan, is the larger; Amkor, headquartered in the United States, is the second.

For Amkor in particular, the relationship is already formalised. TSMC has signed a memorandum of understanding routing packaging from its Phoenix fabs - its own CoWoS and InFO formats - to Amkor's Arizona facility, still under construction. The OSATs are not merely catching what overflows; on the frontier work, they are collaborating with TSMC directly.

Amkor: the company and where it sits

Amkor is the largest outsourced packaging firm headquartered outside Asia, and the second-largest in the world behind ASE. Founded in 1968, it was among the companies that turned outsourced packaging and test into a business in the first place. Amkor designs no chips and fabricates no wafers. It sits downstream of the foundry, supplying the designers - Nvidia, AMD, Apple, Qualcomm - by packaging the silicon they have already made.

Of $6,708M in 2025 net sales, advanced packaging, the higher-complexity, multi-die and integrated-module work, accounted for 82.8% against mainstream's 17.2% in older, simpler packaging for commodity parts. The largest slice of the advanced side is System-in-Package, or SiP, which combines processors, memory and sensors into a single module rather than packaging each separately. At roughly $3.08bn it is close to half the company.

Where that work is done is becoming its own advantage. Most packaging capacity sits in Asia, an increasing friction as supply chains and the big demand-driving end markets shift towards the US - wafers fabricated there would have to travel to Asia for packaging and back. Amkor is building a large advanced-packaging campus in Arizona, secured on commitments from Apple, Nvidia and AMD and under construction since the second half of 2025, next to TSMC's new US fabs. The aim is to keep the whole sequence domestic: a wafer made in Arizona, packaged in Arizona, built into a US data centre without ever leaving the country. No other OSAT of Amkor's size can offer that - ASE is in Taiwan, and the next tier down is in China.

How Amkor compares

Set beside each other, Amkor and ASE are close substitutes - both full-service OSATs selling to broadly the same designers, which is why large customers like Apple and Qualcomm deliberately split their work across the two rather than commit to one. That splitting is what gives the OSATs their volume.

Where they differ is more specific. ASE's advantage is scale: it is the larger firm. Amkor's are narrower - its US domicile and the Arizona build, its depth in SiP co-design, and a packaging partnership with Intel.

That last one is built around EMIB, or Embedded Multi-die Interconnect Bridge - Intel's own approach to high-density packaging, and a genuine alternative to TSMC's CoWoS. Intel developed the technology; Amkor is the partner bringing it to commercial scale in its plants. The value to Amkor is specific: a route into frontier packaging that bypasses the CoWoS queue entirely, and a more diversified revenue base with it.

A different kind of demand

For most of its history, Amkor has been a cyclical business. Packaging volumes tracked the consumer-device cycle, with mobile at the centre of it: strong handset demand filled the factories, and when it softened, utilisation and margins fell with it. That cyclicality is the biggest single reason a firm of Amkor's scale and history has traded the way it has.

AI demand behaves differently, and it runs against that cyclicality. The data-centre build-out is funded by multi-year capital commitments rather than annual refresh cycles, so the volume pulling chips through the line is steadier and far less hostage to the consumer-hardware swing. It is also higher-value work: the multi-die accelerators and integrated modules going into AI systems are more complex to package than the commodity parts that defined the older business.

The Arizona campus and the TSMC overflow agreement further strengthen the change. They go beyond a capacity or a nearshoring story, and act as demand being structurally relocated and locked in. Large customers in the AI build-out (Nvidia, AMD, Apple) committed to packaging done in the US before the capacity was finished, and TSMC is routing frontier work there directly.

Put together, the change turns a cyclical, mobile-led packager into a business with a steadier, higher-value, more domestic demand base. That is part of the real value story.

It is worth being precise about how the re-rating would actually come, because it is not the mechanism the memory players are re-rating on. Amkor re-rating, as opposed to the memory players, comes less through pricing power, as downward price pressure is a standing feature of the OSAT business.

Much of that demand, though, is not landing as orders yet, and the reason sits one step upstream. Packaging comes after memory in the sequence: a packaging house marries compute dies to stacks of HBM, so until that memory exists there is nothing for it to assemble, and a share of the AI packaging demand the shift implies is, for now, doing exactly that - waiting on memory rather than arriving as orders. As the constraint eases, that held-back work converts to orders; and because Amkor's economics are set by how full its factories run, the conversion is the inflection that matters.

The easing reaches Amkor from two directions at once. Its own customers - the designers held up waiting on HBM to finish their accelerators - release the packaging orders they had deferred. At the same time the pool itself widens: as total demand climbs past what TSMC will absorb, the frontier work it turns away routes to the independent layer it has already partnered with for exactly that. The easing does not simply hand Amkor back its deferred book; it enlarges the work the company is positioned to catch.

The risk in the customer book

The same customer relationships that give Amkor its volume are also where its sharpest risk sits.

The large designers it sells to - Apple and Qualcomm in particular - deliberately split their packaging across both Amkor and ASE rather than commit to one. That splitting is what gives the independent OSATs their volume, but it cuts the other way just as cleanly: a customer already running work through both firms can move a block of it from one to the other over capacity or price, and Amkor has little leverage to hold it.

Apple is where that concentrates. At roughly 30% of revenue it is comfortably the largest customer, so the rotation risk is not spread across the book - it is loaded onto a single name. And it is not theoretical. In 2024 Apple moved a meaningful share of its packaging volume to ASE, the clearest demonstration that the risk is live and that it runs through the one relationship large enough to make it hurt.

The shift in demand does more than strengthen the base; it also loosens this grip. As Nvidia, AMD and hyperscaler-driven volume grow into a larger share of the book, Apple's weight in it falls, and so does the leverage any single customer can exert. The same change making the business stronger is what gradually de-risks its customer base.

But today, Apple is still the dominant name, the concentration is real now, and a rotation would still hurt. But the direction matters: the risk is at its sharpest at the start of the shift and dilutes as the shift runs. It is a genuine vulnerability in the present, narrowing across the length of the build-out.

The price of the idea

Amkor's earnings move with how full its factories run, and they currently sit well below their last-cycle peak. The business has been cyclical: revenue eased from $7.1bn in 2022 to $6.3bn in 2024 before recovering to $6.7bn in 2025. Earnings swung harder. Diluted EPS fell from $3.11 in 2022 to $1.43 in 2024, and at $1.50 for 2025 it remains less than half its last peak. Getting back to prior earnings power alone would roughly double EPS, before any structural growth on top.

That recovery is already underway, and the data centre is leading it. First-quarter 2026 revenue rose 27.5% year on year to $1.68bn, with the strength concentrated in AI data centres, and management lifted second-quarter guidance to around $1.8bn. Beneath the headline, the mix has been turning for two years: the computing and data-centre share of revenue has climbed from 16% in 2023 to 20% in 2025, communications - the smartphone-and-tablet core - has eased from 50% to 46%, and advanced packaging now makes up 83% of sales, against 77%.

The projections carry the trend forward. Consensus puts 2026 at roughly $7.5bn of revenue and $1.74 of EPS; management's own targets reach much further - above $11bn of revenue by 2030, gross margins north of 22%, and EPS near $5. Those are management's numbers, not consensus, and read better as ambition than forecast. Either way, the company is acting with confidence - still paying a dividend and running a $300m buyback while funding the Arizona build.

How earnings travel from here to there is operating leverage. Most of Amkor's cost base is fixed, so as utilisation rises the extra volume drops through at high margin without the company charging more. The last upcycle shows the scale: gross margin ran near 20% with the factories full, against 14% today. It is the engine behind both the recovery headroom and management's 22%-plus margin target.

None of this is cheap. Amkor trades at around 41 times trailing earnings and 15 times EV/EBITDA, easing to roughly 31 and 11 times on forward estimates, and ASE - the larger peer - sits at much the same level, near 14 times trailing EV/EBITDA and 10 times forward. This is the AI re-rating of the whole packaging space, not one name running ahead. But it is worth recalling the frame this series opened on. Valuations across the AI complex are stretched. If you wait for these names to look cheap on a conventional multiple, you will sit in cash and watch them run. Amkor here is an AI idea as much as a financial one - a position in the build-out as the narrative turns towards it and the numbers begin to confirm it. Bought now, and more so on a pullback, it fits what this series is looking for: exposure to the idea and the momentum behind it.

The bet rests on the shift arriving. The valuation already embeds it - cyclical-and-mobile giving way to steadier-and-AI - so if it stalls, if data-centre demand proves lumpier than structural or the mix stops moving, the recovery is just an ordinary cyclical upswing, and a stock priced for a re-rating has little underneath it.

Packaging the edge

SiP is already close to half of what Amkor does. It is also the capability the edge build-out needs.

System-in-Package is the method for fitting a lot of silicon into a small, power-constrained space - a processor, memory, an AI accelerator and sensors combined into one module rather than scattered as separate chips on a board. That is exactly the packaging problem an edge device poses. On-device inference has to run in a phone, a wearable, an embedded product, where there is no room for a sprawling board and no power budget for one either. The chip that makes that work has to be packed tight, and packing it tight is SiP.

The relationships line up the same way. The edge build-out runs through the devices people already carry, and the silicon in those devices belongs to the mobile designers - Apple and Qualcomm above all. Amkor already packages for both, with deep SiP co-design work on how those modules are built. The designers who will lead edge AI are the designers Amkor has spent years packaging for. When that silicon needs assembling into a shippable module, the capability and the relationships are already in place.

The point that ties it together is that Amkor is not a single-wave story. It catches the data-centre packaging wave first - the accelerators, the structural demand, the order book that loads as memory eases. And it is positioned for the edge wave that follows, on the same SiP capability and the same mobile relationships. The business pointed at the data centre and the business pointed at the edge are the same business.

It is worth being precise about the timing, because the edge wave is not the data-centre one. Edge is further out - the destination this series is building towards, not the stage the market is pricing now. This is not a reason to own Amkor for the edge angle today. It is the reason the business stays relevant across the whole length of the build-out, into the wave it ends on rather than only the one in front of it.