Micron has confirmed that TSMC will manufacture the base dies for its next-generation high-bandwidth memory, according to the supplied BigGo Finance report. That makes TSMC part of the HBM4-era supply chain for all three major memory makers: Micron, SK Hynix and Samsung.
TSMC Watch analysis
What happened
The base die is the logic layer beneath the stacked DRAM in HBM. As HBM becomes more customised for AI accelerators, that layer has to handle more interface and control work. The report says memory makers are increasingly turning to TSMC's logic-manufacturing expertise, while continuing to make the DRAM core dies themselves.
This is not simply a story about outsourcing to save money. The report says TSMC-made HBM4 base dies can cost three to four times more than comparable in-house core-chip production, but customers are paying for performance and customisation. The exact commercial terms remain attributed claims rather than independently verified figures.
Key findings
- TSMC is now in all three supply chains
The report says Micron joins SK Hynix and Samsung in using, or planning to use, TSMC for HBM base dies. - HBM is becoming more bespoke
Wider interfaces and customer-specific designs are pushing more logic into the base die. - The DRAM stays in-house
The memory makers still manufacture the stacked DRAM core dies themselves, according to the report. - AI infrastructure is spreading its bottlenecks
The shift adds foundry capacity and advanced logic processes to the list of things AI memory supply depends on.
Why it matters
AI hardware is often described as a contest between accelerators, but the memory feeding those accelerators is becoming just as strategic. TSMC is gaining a role beyond conventional chip production, while memory makers are splitting the work between their own DRAM processes and an external logic specialist.
For system builders, that could mean better room for customised HBM designs, but also a more complicated supply chain and another place where scarce advanced capacity can constrain AI deployments. The memory stack is acquiring more plumbing, and, naturally, the plumbing now has a roadmap.
Our read
This is a meaningful infrastructure shift, not a flashy product launch. Treat the broad direction as credible on the supplied account, but treat the detailed pricing, market forecasts and performance figures as reported claims until the underlying company announcements and independent data are available.
The practical takeaway is simple: when assessing future AI-server capacity, watch the base die as well as the GPU, DRAM and packaging queue.
What to watch
- Micron's original announcement and its detailed HBM4 or HBM4E manufacturing plans.
- Whether TSMC confirms production scale, process technology and customer-specific HBM arrangements.
- How Samsung and SK Hynix balance TSMC production with their own foundry capacity.
- Whether outsourced base dies improve HBM performance enough to justify their reported premium.
Discussion spark: Does putting the HBM logic layer in a specialist foundry make AI memory more resilient, or simply move the bottleneck into another scarce part of the chip supply chain?
Sources and evidence
Independent WittyWires tracker for public updates about TSMC. Not affiliated with or endorsed by TSMC; this is not an official account.