3 & 2nm high yields create abundance. Packaging next opportunity.
Packaging delays hinder AI data center rollout
This shortage has direct consequences for AI data center deployment timelines. The bottleneck preventing hyperscalers from deploying AI data centers at their desired speed is the packaging step, not fabrication itself. In 2025, the four largest AI chip designers consumed about 90% of global CoWoS packaging and HBM supply, leaving little available for others. CoWoS demand exceeded supply by an estimated 20–30% in 2025, with lead times stretching to 6–12 months.
Why AI's packaging bottleneck is slowing growth and how firms are responding Copilot: 2. Will transistors in the Z-dimension increase as a result? Yes — this is already happening, and packaging constraints accelerate the shift.
Industry data shows:
- Traditional scaling (shrinking transistors) is no longer the bottleneck — foundries can produce 3nm/2nm logic dies with high yields.
- The bottleneck is advanced packaging, especially 2.5D/3D integration with HBM.
Because horizontal scaling (bigger interposers, more HBM stacks) is constrained, hyperscalers and chip designers are moving toward:
Z-dimension scaling:- Hybrid bonding (TSMC SoIC, Intel Foveros Direct)
- 3D stacked SRAM
- 3D DRAM (HBM4/HBM4E roadmap)
- 3D logic tiles
- Vertical interconnect density increases (TSVs, microbumps ? hybrid bonding)
This is the industry’s response to packaging bottlenecks: If you can’t scale out (CoWoS area), you scale up (Z-dimension).
3. Is this why Intel’s Tan says packaging is the next growth driver? Yes — and the search results directly support this.
Intel’s EMIB-T is explicitly positioned as an alternative to CoWoS:
- EMIB-T capacity expected to reach 40,000–45,000 wafers/month by 2028.
- EMIB-T is “more suitable for custom AI chips such as those designed by Google and Amazon.”
- Intel can capture demand that TSMC “may not be able to meet quickly enough.”
This is exactly the strategic point Intel’s leadership (including Tan) has been making:
Why packaging is the next growth driver:- Logic is not the bottleneck TSMC can produce 3nm/2nm wafers; the constraint is packaging.
- AI accelerators require massive interposers + HBM Every generation increases interposer area and HBM stacks.
- Packaging capacity is booked 85%+ through 2027 Even with tripled output, demand still exceeds supply.
- Hyperscalers need alternatives Intel’s EMIB-T and Foveros give them a second source.
- Packaging is where differentiation now happens Performance per watt, bandwidth, latency, and memory proximity are packaging-driven.
Intel doesn’t need to beat TSMC at 2nm. Intel needs to provide packaging capacity that hyperscalers cannot get elsewhere.
This is why Tan says packaging is the next growth driver — and the data supports him.
4. So does XT:260 “fix” the bottleneck? No — it reduces it.
TSMC plans to double CoWoS capacity to 260,000 wafers/month by 2028, but:
- Demand is rising faster than supply.
- Each GPU generation uses larger interposers and more HBM stacks.
- Hyperscalers (NVIDIA, Google, AMD, Amazon) consume nearly all available capacity.
XT:260 is necessary, but not sufficient.
Final synthesis- XT:260 helps but does not eliminate the CoWoS/HBM bottleneck.
- Z-dimension scaling will increase because horizontal scaling is constrained.
- Intel is right: packaging is the next growth driver, and EMIB-T/Foveros are strategically positioned to capture overflow demand from TSMC.
- The bottleneck has shifted from transistors -> packaging, and this is the defining semiconductor trend of 2025–2030.
|