TSMC’s CoWoS-L remains the dominant advanced-packaging technology for AI chips, but Samsung Electro-Mechanics and LG Innotek are preparing alternatives rather than waiting for one winner to rule the wafer. The important shift is that AI hardware competition is moving beyond chip design and fabrication into the packaging that joins processors, memory and multiple dies together.
TSMC Watch analysis
What happened
The Korea Herald reports that TSMC’s CoWoS-L is expected to remain the mainstream AI-packaging technology through at least 2028, backed by its manufacturing experience, customer base and expanding production capacity. Nvidia and other AI-chip developers rely on TSMC’s advanced packaging to combine logic chips with high-bandwidth memory.
Korean substrate makers are nevertheless developing several paths. Samsung Electro-Mechanics is working on 2.1D and 2.5D substrate technologies, alongside approaches related to Intel’s EMIB and EMIB-T bridge-based packaging. LG Innotek is expanding high-end FC-BGA substrate capabilities aimed at AI servers and data centres.
Why it matters
An AI accelerator is no longer simply a chip that gets faster when its designer adds more silicon. Larger packages must connect more compute dies and memory while controlling power delivery, heat, circuit accuracy and warpage. Packaging has become part of the performance equation, and a constraint on packaging can slow an AI build-out even when chip designs and factory capacity are available.
The practical consequence is a more complicated supply chain. TSMC’s lead is substantial, but customers and suppliers have reasons to support alternatives for cost, capacity, architecture and geopolitical resilience. The industry is not replacing CoWoS-L tomorrow. It is making sure tomorrow does not depend on a single answer.
Our read
This is a useful infrastructure story because it moves the AI conversation away from the usual parade of model launches and towards the physical details that decide whether those launches can scale. TSMC still has the strongest position described here, but the preparation by Korean firms suggests the market expects several packaging architectures to coexist.
For readers tracking AI hardware, the sensible question is not simply who makes the fastest accelerator. It is who can package the whole thing reliably, at volume, without turning every new generation into an engineering bottleneck. The glamorous bit remains the model. The less glamorous bit is where the models get assembled.
What to watch
- Whether CoWoS-L remains the clear mainstream choice through 2028.
- Whether Samsung Electro-Mechanics or LG Innotek secure named AI-server customers for advanced substrates.
- Whether Intel’s EMIB or EMIB-T wins meaningful commercial adoption beyond specialist designs.
- Whether larger AI packages make power delivery, heat and warpage bigger constraints than raw compute.
Discussion spark: Should AI hardware buyers prioritise the best-performing packaging technology, or deliberately support several competing routes to reduce supply-chain dependence?
Sources and evidence
- Beyond CoWoS: Korean firms eye next AI packaging race – The Korea Herald (25 September 2026, 07:00 UTC)
Independent WittyWires tracker for public updates about TSMC. Not affiliated with or endorsed by TSMC; this is not an official account.