← All sectors / The AI transformation
179 · Advanced packaging for AI accelerators
Stacking silicon into one part
Curve position
Binding constraint
Assembly capacity, which has been the industry bottleneck for accelerators.
Why this sector sits at Liftoff
Advanced packaging for AI accelerators is in clear Takeoff: TSMC CoWoS capacity is fully booked with ~1M wafers of 2026 demand (up from ~370K in 2024), Samsung is mass-producing its 3D advanced packaging for AI chips in Q2 2026, HBM4 is ramping, and the sector has repriced sharply with memory and packaging now representing 60–70% of AI accelerator COGS — capital is flooding in, the story is widely known, and new entrants are racing to add capacity.
Reviewed on a two month cycle. The position moves only when a dated, verifiable change in the binding constraint justifies it.
Heatmap / How this sector's stocks are moving
| Stock | Price | 1 Day | 1 Week | 1 Month | Year to date |
|---|---|---|---|---|---|
| $AMKR | $35.66 | −5.21% | −15.52% | −33.21% | −9.68% |
| $ASX | $37.51 | −3.94% | −14.59% | −19.82% | +132.98% |
| $TSM | $395.75 | −2.24% | −9.96% | −17.95% | +30.23% |
| $BESIY | $152.82 | −11.02% | −15.96% | −22.43% | −2.51% |
| $KLIC | $68.44 | −5.52% | −13.85% | −30.11% | +50.22% |
| Sector average | −5.59% | −13.98% | −24.70% | +40.25% |
Prices as of November 5, 2026. The sector's move is the equal-weighted average of these listed companies, a market proxy for context, not recommendations. See it on the full heatmap →
Sentiment / The mood in this sector's stocks
Since 2018, Advanced packaging for AI accelerators has been in Capitulation and steady 9 separate times. After those, its stocks beat their usual three months 5 times (56%), by a median of +1.9 points.
Confidence: low, a split like this is common by chance, so no edge is called.
Performance gains increasingly come from how several pieces of silicon are combined into one package rather than from the transistors on any single die. That shift made packaging a bottleneck for the entire accelerator supply chain.
Historically packaging was the low value step at the end of manufacturing, outsourced and largely ignored by anyone analysing the industry. Chiplet architectures and high bandwidth memory changed that completely.
The structural driver is that accelerator supply has been limited by packaging capacity rather than by wafer capacity, which is an unusual inversion and one that persists.
The technology layer spans silicon interposers, organic substrates, hybrid bonding that connects die directly without bumps, high bandwidth memory stacking, thermal solutions for very high power packages, and the test that verifies known good die before assembly.
Adoption economics are dictated by yield. A packaging failure destroys several expensive die simultaneously, which makes process control worth more than throughput.
The beneficiaries include outsourced assembly and test providers, substrate manufacturers, packaging equipment makers, bonding equipment suppliers, and the memory makers whose stacks go into the package.
The value chain runs from substrate and interposer through assembly to test. Substrate supply has been a specific and persistent constraint.
The overlooked layer includes substrate manufacturers, bonding equipment makers, underfill and material suppliers, and the inspection firms verifying assemblies.
Competitive dynamics are concentrated among a few assembly providers with the required capability, giving them unusual leverage in a business that was historically commoditised.
Risks: capacity additions could overshoot demand, accelerator demand may normalise, substrate capacity is concentrated geographically, and technology transitions can strand equipment.
What to watch: packaging capacity additions, substrate lead times, hybrid bonding adoption, and accelerator supply commentary from the companies constrained by it.