What is the part between an AI chip and its memory?
An AI chip does not work alone. It sits beside stacks of memory, and a thin slab of silicon called an interposer connects them.
TSMC's name for the packaging that does this is CoWoS, short for Chip-on-Wafer-on-Substrate. Think of the interposer as the street grid between a processor and its memory. The chip and the memory can both be perfect, and the system still won't run without the grid between them.
At TSMC's third-quarter call this month, CEO C.C. Wei said CoWoS "remained customers' top choice for integrated circuit assembly," as Focus Taiwan reported it. Most buyers of AI servers never see this step. It can still hold up a delivery.
What did GlobalFoundries and TSMC actually agree to?
GlobalFoundries will make silicon interposers for TSMC's CoWoS packaging at its Malta, New York plant. The initial term is five years.
Volume production is expected to ramp in the first half of 2028. Embedded Computing Design, which reported the announcement on October 13, says the agreement also covers embedded deep trench capacitor components and sets a framework for adding capacity as customer demand grows. Press coverage puts the value at about $2 billion. The companies describe Malta as the first U.S.-based source of these interposers.
GlobalFoundries senior vice president Ed Kaste said "advanced packaging is becoming increasingly critical" and called the deal "a secure, scalable source of essential advanced-packaging elements." What the companies left out matters as much as what they said.
| Item | What was reported | What was not disclosed |
|---|---|---|
| Value | About $2 billion | How much is plant spending and how much is prepaid wafers |
| Term | Five years initial, with room to expand | How much capacity the expansion adds |
| Volume timing | Ramp in the first half of 2028 | Actual volumes and TSMC's share of Malta's output |
| Customer | Not named | Who ends up using these interposers |
| Scope | Interposers for TSMC's CoWoS | Whether every CoWoS variant is covered |
Read the table from a buyer's side. Every blank in the right-hand column is a question you could put to a vendor, and probably can't get answered yet. The customer isn't named, so you can't tell whether your vendor's chips are in line for this supply. Volumes aren't given, so you can't tell how much of the 2028 ramp is already spoken for. That is normal for a deal announced this early. It also means the headline number tells you very little about your own order.
Does this create a fully U.S.-made AI chip package?
Not by itself. Coverage of the deal says final assembly, testing and the memory involve separate steps and separate suppliers.
Embedded Computing Design describes only the interposer. It does not mention U.S. assembly. In September, TrendForce passed along a Taiwanese press report, which TSMC has not confirmed, that CoWoS capacity could double from about 130,000 wafers a month at the end of 2026 to about 260,000 by the end of 2028. The report centers that build on Taiwan and TSMC's Arizona campus.
Nexairi Dispatch
Get the next AI move before it turns into a workflow problem.
Join the free newsletter for concise AI news, practical checklists, and the decisions practitioners need to make next.
Free. No spam. You will also get the Nexairi Dispatch.
Why does the date matter more than the dollar figure?
The first half of 2028 is when the ramp starts, not when the shortage ends. A company ordering AI servers in 2027 is ordering before this supply exists.
TSMC CFO Wendell Huang said the company will "ensure sufficient CoWoS capacity to meet their demand," per Focus Taiwan. That is a promise without a date. TSMC also kept its 2026 capital budget at $60 billion to $64 billion and expects higher spending in 2026 through 2028 than in the prior three years. Money is going in. The coverage we found does not say when the extra packaging capacity comes out.
Which part of the AI chip sets your delivery date?
An AI accelerator needs four parts to arrive together. A shortage in any one of them stalls the whole rack.
- The chip. The processor itself, made at a foundry such as TSMC.
- The memory. Stacks of high bandwidth memory next to the chip. Our October 12 piece covers why SK Hynix's CEO calls 2027 the worst supply year on record.
- The part between. The interposer. GlobalFoundries becomes a new U.S. source here from 2028.
- The assembly. The CoWoS step that joins the other three, run by TSMC.
Picture a procurement lead who gets a server quote with a delivery date 14 months out. The vendor says "supply constraints" and nothing else. That is an illustration, not a reported case. The point is that the quote doesn't say which of the four parts is the slow one.
Reading the 2028 date
This is interpretation, not reported fact. The memory layer and the packaging layer are both tight, and neither has a dated fix before 2028. A second source for interposers is good news for the market. It is not news for a buyer whose purchase order is due next spring. If a vendor tells you this deal will free up your 2027 order, ask them to show where the capacity comes from and when.
What This Means for You
The upside: a second packaging source in the U.S. is coming, with a framework to grow. Over a longer horizon that should loosen one of the tighter spots in the AI buildout.
The risk: it does not help a 2027 order. If your plan assumes faster deliveries because of this deal, the sources do not support that.
The next question: ask your server vendor which step sets your delivery date. Then ask whether the 2027 quote assumes any new packaging capacity, and whether your allocation is contracted or first come, first served.
Here is the plain read. The headline says a U.S. plant will make a part for AI chips. The date in the fine print says the help comes after most 2027 budgets are spent. Pull out your current quote this week and find the line that explains the lead time. If it isn't there, that is your first question.
Sources
- Embedded Computing Design: GlobalFoundries and TSMC agreement on U.S. silicon interposers
- AI Weekly: GlobalFoundries, TSMC sign $2B U.S. interposer deal in Malta, NY
- Let's Data Science: GlobalFoundries and TSMC sign $2 billion U.S. CoWoS interposer deal
- Focus Taiwan: TSMC third-quarter results and capex
- TrendForce: TSMC reportedly targets CoWoS capacity boost by 2028
Related articles on Nexairi
Jim Smart is the founder and editor in chief of Nexairi. A Business Intelligence Developer with experience building data systems for Verizon, U.S. Army operations, and enterprise finance teams, Jim spent years turning complex data into decisions that executives could act on — dashboards, forecasting models, and automation pipelines across telecom and government contracting. He founded Nexairi to apply that same clarity to AI: making emerging technology understandable and actionable for the operators, accountants, and business owners who need it most. Jim holds GenAI certifications from the University of South Florida Bellini College of AI and completed Springboard's Data Science Career Track.
