BofA: Taiwan Supply Chain Sell-Off on Kyber Delay Is an Overreaction
Alina Collins
After SemiAnalysis flagged a roughly one-year delay to Nvidia's Kyber NVL144 rack, Taiwan CCL/PCB/substrate stocks dropped 8%-9%; BofA's Taiwan tech team called the selloff overdone, maintained buy ratings, and framed the issue as a supply-constrained volume deferral, not a demand reversal.
What actually happened?
SemiAnalysis posted six threads on X claiming Nvidia's Kyber NVL144 rack faces a major delay, adding that Rubin Ultra was cut from a 4-die to a 2-die design and the co-packaged optics (CPO) timeline is uncertain.
Taiwan stocks in CCL, PCB, substrate, and test-interface segments sold off 8%-9% on average.
Nvidia's spokesperson told Bloomberg the Kyber roadmap is "intact." Taiwan supply-chain sources told DIGITIMES that even if the timeline shifts, it will not affect Rubin Ultra's overall shipment cadence.
Why does BofA say the drop went too far?
BofA's Taiwan tech team (analyst Yang Ming et al.) labeled the potential delay a "supply-constrained volume deferral", not a demand-trend reversal. This means → the pie is not shrinking; it may just arrive later.
The precedent: similar fears over DRAM demand in Vera CPUs previously triggered a selloff, but the demand trend ultimately held.
Key logic: CCL/PCB/substrates account for only a low-to-mid single-digit percentage of an AI server's total bill of materials (BOM). In plain terms = the lower the cost share, the less likely customers cut these items first — high-value components like memory get squeezed before basic materials do.
Q-glass vs. PTFE — which route wins?
The manufacturing challenge in Kyber's PCB mid-board points to M9-grade CCL, where two material routes compete: Q-glass (quartz cloth) and PTFE (polytetrafluoroethylene).
BofA notes the bottlenecks differ: PTFE is constrained on the CCL production side; Q-glass is constrained on the PCB fabrication side (yield ramp). This difference means Q-glass-based M9 CCL will penetrate AI servers faster.
This reflects a counterintuitive dynamic: the slower Q-glass PCB yields improve, the wider the high-end CCL supply-demand gap grows — and the higher the scarcity premium, which actually benefits CCL suppliers.
Which sub-segments are most resilient?
BofA ranks resilience explicitly: test interfaces / substrates > CCL > PCB. The closer to the die, the deeper the tie to the AI-server upcycle, and the stronger the through-cycle defense.
Within CCL, two suppliers diverge: TUC has more near-term upside from share gains and pricing power; EMC holds a long-term scarcity advantage rooted in substrate-material capability and M9-Q technology barriers.
On substrates, the market worries that a 2-die design cuts per-die substrate content — but BofA argues higher total shipment volumes offset this. The concern that memory customers will pressure substrate pricing "does not hold" — BT substrate supply is shrinking, glass-cloth costs are elevated, and memory demand remains strong. All three factors support substrate pricing resilience.
How tight is supply, really?
High-end CCL (M8-grade and above) and ABF substrates are expected to remain undersupplied through at least the end of 2027.
The root cause: equipment lead times are long, while server CPU- and accelerator-driven demand growth is structural.
This means → even if Kyber's timeline slips, the supply bottleneck does not ease — capacity expansion is paced by equipment cycles, not customer orders.
What should investors watch next?
The supply chain's near-term focus has shifted to Vera Rubin, expected to ship in H2 2026 on the Oberon rack. Volume is projected to ramp from late Q3, making it the core revenue driver for server suppliers in the second half.
DIGITIMES sources say Kyber NVL144 was originally expected for H2 2027 mass production, but many technical details remain unverified — Nvidia has never explicitly told suppliers that Kyber is delayed. Market reports inferred the delay from supply-chain conditions, not from an official Nvidia statement.
In plain terms = two verification milestones will decide confidence in Nvidia's AI-infrastructure supply chain: whether Kyber can hold its revised timeline, and the actual shipment volume of Vera Rubin in H2 2026.
Content is for reference only, not financial advice.