TSMC plans to have CoWoS silicon interposers manufactured at GlobalFoundries, according to a report by the German tech publication igor’sLAB. Neither company has confirmed the arrangement publicly, and the report does not specify volumes, timelines or financial terms. If it holds up, it would be the first time the world’s largest contract chipmaker hands production of its most advanced packaging step to an external foundry.
CoWoS, short for Chip-on-Wafer-on-Substrate, is the packaging family behind virtually every flagship AI accelerator on TSMC silicon, including Nvidia’s data center GPU lineup and AMD’s top-end Instinct processors. The process mounts logic dies and high-bandwidth memory on a silicon interposer, a passive wiring layer that carries thousands of connections between the chips. Without that layer, today’s large AI accelerators cannot be built at all.
Why this step matters
The context is simple: advanced packaging has become the scarcest resource in the semiconductor industry’s AI buildout. TSMC has warned investors repeatedly that advanced packaging capacity, not wafer fabrication, is the binding constraint on AI chip shipments. Nvidia and AMD have both cited CoWoS capacity as the limit on how many accelerators they can ship in a given quarter. The company has run an aggressive expansion program across its Taiwan sites, but analysts and supply-chain reports have consistently described the ramp as tight relative to hyperscaler orders.
Outsourcing interposer manufacturing would give TSMC a second external leg of CoWoS supply. It would also amount to a striking reversal of practice. TSMC has historically kept every step of its most advanced manufacturing in-house or within its own contracted ecosystem in Taiwan. The company’s Taiwan cluster is one of its strongest defenses; moving even the interposer step, the least transistor-intensive part of the stack, to a US-headquartered specialty foundry signals how badly the capacity crunch has bitten.
Why GlobalFoundries is a logical pick
GlobalFoundries makes sense as the partner for a specific reason: interposers do not require bleeding-edge transistors. They are passive wiring layers, and the work plays to a specialty foundry’s strengths rather than to a scale race TSMC has already won. GlobalFoundries exited leading-edge logic at the 7nm node in 2018 and has since built its business around mature and specialty nodes, automotive chips, and RF and embedded work.
A TSMC arrangement would give GlobalFoundries a direct position in the AI supply chain, a segment where it currently has no flagship product exposure. For a company that has spent six years being described as the foundry that stepped back from the frontier, that is a meaningful shift. Interposer volumes also scale with accelerator output, which is the fastest-growing segment in the industry by revenue.
The geopolitical angle is hard to miss. TSMC has faced years of pressure from Washington to expand US manufacturing footprints, and its Arizona fab program is already the largest single foreign investment in US chipmaking history. Partnering with a fab that operates in Malta, New York and Singapore diversifies the packaging supply chain outside Taiwan, which matters to hyperscalers and governments alike after years of concentrated risk discussion.
What the report does not say
Chip Dispatch, which covered the report on October 10, notes it could not independently verify the claim at press time. No filing, press release or executive statement backs the arrangement yet. The igor’sLAB article does not say whether the deal covers existing CoWoS capacity or output from planned expansions, and neither company has commented.
That caution matters because the stakes are high. CoWoS interposers go into products with per-unit prices in the tens of thousands of dollars. Any quality problem in the interposer layer shows up as a dead accelerator, and Nvidia and AMD qualification cycles for a new packaging supplier would take months of validation before shipping volumes ramped. Even if the deal is real, shipping products built on GlobalFoundries interposers is a 2027 story, not a 2026 one.
Competitive context
The move would also land in a competitive landscape that has shifted over the past year. Intel Foundry has pitched its own 2.5D and 3D packaging services to AI customers, and Samsung has done the same with its advanced packaging lines. Neither has made a real dent in the high-end accelerator market, where TSMC’s CoWoS remains the default choice. Licensing or contracting out interposer production would function as a pragmatic capacity release valve for TSMC: cheaper and faster, perhaps, than building new packaging lines of its own, and it keeps the customer relationship fully inside TSMC while the extra wafers come from outside.
For GlobalFoundries, the arrangement would not put it back in the race for leading-edge logic, and it does not need to be. The foundry has been profitable on the mature-node strategy, but its stock has traded at a discount to peers on the argument that it lacks AI leverage. An AI supply-chain role, even a narrow one, gives the market a different story to price.
What to watch
The verification path is straightforward. TSMC’s quarterly earnings call, due later this month, would be the first credible confirmation point if the company plans to acknowledge the arrangement. GlobalFoundries’ own investor communications would be the second. Beyond that, watch for supply-chain evidence: analysts who track accelerator teardowns would eventually identify a GlobalFoundries-fabricated interposer in a shipped Nvidia or AMD product, which would settle the question more reliably than either company’s press shop.
If the report does not hold up, nothing much changes. TSMC continues expanding its own Taiwan packaging capacity, and GlobalFoundries continues to look for AI-adjacent revenue from other angles. If it does hold up, it reshapes the competitive dynamics of advanced packaging by turning TSMC’s worst bottleneck into new business for a foundry it otherwise displaced from the leading edge. Either outcome will be visible well before the next generation of accelerators ships.