TDEB Top Bar

Trending
Semiconductor Industry News — The Digital Electronics Blog
Follow Us

TDEB Masthead

The Semiconductor Hub

Semiconductors.  Chips and Systems.  Innovation.

Indexed by leading Search Engines, AI and referenced by Engineers worldwide 500+ Articles 10M+ Pageviews 30+ Reports 50+ Citations

From Bump to Final Test: Who Really Owns Each Advanced Packaging Step

Murugavel Ganesan
by
0
Share Facebook X LinkedIn WhatsApp Email
Cutaway view of an advanced chip package with logic dies and memory stacks on an interposer and substrate.

TSMC told investors in January 2026 that advanced packaging was slightly over 10% of its 2025 revenue, which on US$122.42 billion of total revenue is a little over US$12 billion. Amkor reported US$6.71 billion of total net sales for the same year, and ASE said its leading-edge packaging and test line, LEAP, reached US$1.6 billion. These categories are defined differently, so they do not rank anyone, but they do show that packaging is now a business of real size inside a foundry.

The old line, where the foundry made the wafer and an outsourced assembly and test house (OSAT) did everything afterwards, no longer describes how these parts get built. Ownership now changes hands step by step, and sometimes crosses back. This post maps each step to who does it, using what the companies themselves publish, and then shows with a small worked example why the choice matters to a buyer.

Key takeaways

  • "Foundry or OSAT" is the wrong question for a 2.5D or 3D part. The useful question is who owns each step, and who pays when a package fails at that step.
  • Wafer-level stacking is the step that stays closest to the foundry. Bumping, assembly, substrate attach and test are all offered by both camps.
  • The crossings are public. TSMC plans to contract turnkey packaging and test from Amkor in Arizona, Amkor plans to run Intel's EMIB assembly, and Samsung's turnkey page names OSAT partners.
  • In our illustrative example, materials make up about 90% of the cost of each package started. A contract term on scrap can matter more than the assembly fee.

The lines have moved in both directions

Start with what each company says it sells. All of this is company-published description, not independent verification.

CompanyWhat it publishesWhere the line crosses
TSMCDescribes SoIC as frontend 3D stacking and CoWoS as a chip-last process on an interposer, with CoWoS-S, CoWoS-L and CoWoS-R variants and InFO alongsideIn October 2024 it signed a memorandum to contract turnkey packaging and test from Amkor in Peoria, Arizona. In April 2026 its CEO said its packaging capacity is very tight and it has to work with OSAT partners
Intel FoundryLists EMIB 2.5D and 3.5D, Foveros-S, Foveros-R (stated as production ready in 2027) and Foveros Direct 3DSays chips can go through its own assembly and test or through an OSAT. Test ranges from full turnkey to support for customer-owned programmes and hardware
Samsung FoundryDescribes a package turnkey service across fab, bump, package and test, with I-Cube 2.5D derivativesStates it can enable in-house or OSAT routes, and lists Amkor, JCET, ASE Korea and Hana Micron as OSAT partners
AmkorLists wafer bump, wafer probe, back-grind, package design, packaging, burn-in, system-level and final test, and drop shipment as turnkey servicesAnnounced in April 2025 that it will implement Intel EMIB assembly in Korea, Portugal and its planned Arizona site. Its 10-K describes foundries offering turnkey services through to final test as competitors
ASEIts VIPack platform covers FOCoS, FOCoS-Bridge, FOPoP, FOSiP, TSV-based 2.5D and 3D processing and co-packaged opticsReports LEAP as a separate line. In February 2026 it guided to about US$3.2 billion for the year, split roughly 75% packaging and 25% testing, and in July it said it was tracking ahead of a US$3.5 billion figure

Two things stand out. First, the foundries openly offer an OSAT route while OSATs openly plan to run technology that began at a foundry. Second, Amkor's 10-K treats contract foundries as both customers and rivals: it says that foundry and integrated device customers keep weighing outsourced services against their own in-house capacity.

TSMC's position in late 2026 is worth reading closely. On its October call, its CEO claimed real production at 5 times reticle size, doubted that anyone else can do that today, and described work towards 9 times and 14 times. TSMC's technology page says a 5.5 times interposer solution finished certification in 2025 with volume production to start in 2026. These are TSMC's own statements, and we have not verified them independently. Asked in April whether TSMC would let another company package wafers it makes, the CEO gave no direct yes or no, saying instead that competitors offer attractive technology, that customers can have more choices, and that TSMC would not leave business on the table.

The ownership map, step by step

This is a generic flow for a logic and HBM part on an interposer or bridge. Real flows vary by technology, and some reorder these steps. The "published" columns cite what companies say they offer. The last column is our reading of what a buyer needs to pin down, and it is advice rather than sourced fact.

StepFoundry route (published)OSAT route (published)What the buyer should settle
1. Wafer fabricationFoundryNot offered by the OSATs covered hereWhether the foundry will ship wafers to an outside packager for your product, and on what terms
2. Wafer-level 3D stackingTSMC SoIC (called frontend stacking), Intel Foveros Direct, Samsung vertical integrationASE lists TSV-based 3D IC processing. Ask what bond pitch and bond type are actually qualifiedThis step is closest to the fab. Moving it later is usually a redesign rather than a sourcing change
3. Wafer probe and sortIntel lists wafer sort and die sort. Samsung lists wafer testAmkor lists wafer probeWho owns the probe programme, the data and the pass limits that define a known good die
4. Bumping and back-grindSamsung lists C4, copper pillar and microbumpAmkor lists wafer bump and back-grindBump metallurgy and pitch must match the assembly site, so fix this before choosing the next step
5. Interposer or bridge supplySamsung lists interposer under fab. Intel offers EMIB. TSMC offers CoWoS variantsASE lists FOCoS-Bridge and TSV-based 2.5D. Amkor's EMIB work uses Intel's bridge technologyWho sources the silicon, and whether a second source can use the same design
6. Die and HBM attachTSMC CoWoS and InFO, Intel EMIB and Foveros, Samsung I-Cube assemblyAmkor and ASE, plus Amkor's planned EMIB assemblyWho buys the HBM, who holds it in stock, and who pays if a good stack is lost to a bad attach
7. Substrate attachSamsung describes supply control of PCB suppliers and pre-booking capacity under long-term agreementsAmkor says it sources laminate substrates from a limited group of suppliersWho holds the substrate allocation. It can gate output as hard as the packaging line itself
8. Burn-in, final and system-level testIntel lists burn-in, final test and system-level test. Samsung lists package testAmkor lists burn-in, system-level and final testWhere test lives, and whether results feed back to steps 3 and 6 in time to change anything
9. ShipmentPart of a turnkey offer, in the terms each provider agreesAmkor lists drop shipmentWhere title passes, and who holds the finished stock

If the earlier steps are unfamiliar, our posts on how wafer probe keeps bad dies out of costly packaging and on what silicon interposers do in a 2.5D package cover them in more depth.

What this means for buyers

You are choosing an accountability model

A single turnkey owner gives you one party to call when a part fails. A split flow gives you more choice and more interfaces, and each interface is a place where a failure can land between two contracts. Neither is the right answer in every case. What goes wrong is drifting into a split flow without deciding who owns the failure analysis across the seams.

Capacity is a technical constraint

TSMC's CEO said in April that its capacity is tight and it needs OSAT partners, and in January TSMC said it was applying for permits for its first advanced packaging fab in Arizona. Amkor says construction of its Arizona facility began in the second half of 2025, and it put its 2026 capital spending at about US$2.5 billion in October. So the location of the fab and the location of the packaging may not match for some time. If US assembly matters to your programme, ask for dates in writing. Amkor's EMIB announcement gives no start dates, so confirm the qualification status at each site rather than inferring it from the press release.

Choosing a route is also choosing a design kit

CoWoS, EMIB, I-Cube and FOCoS are not interchangeable. Switching partly means re-doing floorplan, power delivery and thermal work, as our comparison of EMIB and CoWoS trade-offs shows. A second source for the assembly labour is much easier to arrange than a second source for the architecture. Our chiplet lifecycle on The Semiconductor Hub shows where these choices fall in a programme.

Test can move separately

ASE's guidance implies about a quarter of its 2026 LEAP revenue from testing, which on US$3.2 billion is about US$0.8 billion. TSMC's January capital plan set aside about 10% to 20% of the 2026 budget for advanced packaging, testing, mask making and others, a basket that its CFO put at about or below 10% in the past. For a buyer, the point is that test can be separated from assembly, and Intel explicitly offers support for customer-owned test programmes. Decide early who owns the programme and the data, because that is hard to move later.

A worked example: the fee is the small number

Every number here is an illustrative assumption, not a quote or data from any company. The two routes and their yields are invented to show the method. Nothing here implies that either kind of supplier yields better. Replace them with your own figures.

Per package started (illustrative)Route A: one turnkey ownerRoute B: split flow
Tested compute diesUS$300US$300
HBM setUS$500US$500
Interposer and substrateUS$150US$150
Assembly and test feeUS$100US$80
Extra logistics between sitesUS$0US$5
Cost per startUS$1,050US$1,035
Assembly and test yield98%96%
Cost per good packageUS$1,071.43US$1,078.13

The model assumes a failed package is a total loss, with nothing recovered. Route B starts US$15 cheaper and finishes about US$6.70 dearer per good package, because two points of yield are worth more than the fee saved.

Three checks are worth doing on your own version:

  1. Break-even yield. Route B matches Route A when its yield reaches 1,035 × 0.98 ÷ 1,050, which is 96.6%. If Route B reached 98% too, it would cost about US$1,056.12 and win.
  2. Material share. Materials are US$950 of Route A's US$1,050, or about 90%. The party doing the assembly earns the small part of the bill and touches all of it.
  3. Scrap terms. For 100,000 good packages, Route A needs about 102,041 starts and Route B about 104,167. That is about 2,041 and 4,167 scrapped packages, carrying about US$1.94 million and US$3.96 million of materials. Total spend is about US$107.14 million against US$107.81 million, a gap of about US$0.67 million.

Now suppose Route B's contract refunds half of the scrapped materials, and Route A's refunds nothing. Route B's refund is about US$1.98 million, its total falls to about US$105.83 million, and it now beats Route A by about US$1.31 million. In this example, that one clause moves the result by more than the fee and yield differences, which net to about US$0.67 million. Real terms are private and will differ, so the lesson is about what to negotiate rather than what to expect: get the scrap and consigned-material terms on the table before you compare fees.

Questions to ask before you award the work

  1. Who owns a failed package at each step? Ask for it in the form of the table above, with a named party in every row.
  2. Who buys and holds the HBM and the substrates? If it is you, the allocation risk is also yours.
  3. What defines a known good die, and who signs it off? A dispute after attach is expensive when the pass limits were never agreed.
  4. Is this site qualified for this technology today? Ask for the qualification report, and treat a future date as a target.
  5. Who runs failure analysis across a seam? Agree how data moves between companies, and how quickly.
  6. Will your foundry release wafers for this flow? Get the answer for your product and your volume, rather than relying on general statements.
  7. Do your location rules cover packaging and test? A US or European wafer may still be assembled elsewhere.

FAQ

Are OSATs losing advanced packaging to foundries?

The published numbers do not show a simple loss. TSMC's packaging share rose from about 8% of revenue in 2024 to slightly over 10% in 2025, and it expects low teens in 2026. Over the same period, ASE's LEAP line went from US$0.6 billion to US$1.6 billion. Both can grow while capacity is short.

Can I compare these revenue figures directly?

No. TSMC's figure is its own advanced packaging category. Amkor's Advanced Products line was US$5.56 billion, or 82.8% of its sales, but it includes flip chip, memory and wafer-level processing with related test, which is far broader than 2.5D and 3D. ASE's LEAP is a separate line, and its definition is not spelled out in the releases we read.

Does a turnkey deal remove the risk?

It moves the risk and makes it easier to trace, but only as far as the contract says. Read the scrap, consigned-material and capacity clauses as closely as the price.

Sources

The worked example uses illustrative numbers. Vendor figures and statements are the companies' own claims and targets, as published on the dates shown, and dollar figures derived from a stated share are our arithmetic. This post is general information for engineering and sourcing decisions and is not investment advice.

Murugavel Ganesan
About the Author
Murugavel Ganesan
Product Director | SerDes · IVN · Edge AI · Compute | Ex Intel · Infineon

A semiconductor product leader working across product architecture, definition, strategy, and silicon and software execution, from concept through volume production. He has worked with teams across Asia, Europe and the USA, including at Intel, Infineon and ArrayComm.

His work covers ASICs, SoCs, FPGAs and chiplets for wireless, wireline, automotive, IoT and AI, including PCIe, CXL and DDR interfaces alongside RF and mixed-signal design.

Share Facebook X LinkedIn WhatsApp Email

Post a Comment

0 Comments

Your comments will be moderated before it appears here.

Post a Comment (0)

#buttons=(Ok, Go it!) #days=(20)

Our website uses cookies to enhance your experience. Learn more
Ok, Go it!