Morgan Stanley: AI Hardware 3D Scaling Logic Could Quadruple Optical Module Demand
Claire Weston
Morgan Stanley breaks AI infrastructure expansion into three scaling paths — up, out, and across — and forecasts AI optical module market size will more than quadruple from 2025 to 2028, while flagging smartphones as a margin casualty of the AI supercycle.
What does "three-dimensional scaling" mean, and why do optical modules bear the brunt?
Morgan Stanley frames AI hardware expansion along three axes: scale up (making each chip more powerful), scale out (networking more chips together), and scale across (connecting different hardware types to share data).
This means → all three paths pile demand onto the same bottleneck — optical interconnects, the links that move data between chips and racks using light signals.
In plain terms = stronger chips, more of them, and more variety all require thicker, denser "pipes" between them. Optical modules are the core component of those pipes.
Beneficiary segments span GPUs, optical modules, PCBs, ABF substrates, MLCCs (multilayer ceramic capacitors), racks, interconnects, power supplies, and test equipment.
How much can optical module demand actually grow?
Morgan Stanley provides an architecture-migration dataset: legacy architecture requires roughly 144 optical engines per rack; a hybrid architecture raises that to about 1,230; full CPO — co-packaged optics, mounting optical components directly next to the chip — pushes it to roughly 2,526.
This means → per-rack optical engine demand multiplies more than tenfold. The total AI optical module market could reach nearly $100 billion by 2028, more than quadrupling from 2025.
That figure dwarfs some industry consultants' estimates of $5–6 billion. This reflects a far more aggressive view on migration speed than the market consensus.
Will CPO disrupt existing optical module makers?
The market worries that once CPO matures, today's pluggable optical module vendors will be displaced. Morgan Stanley calls that concern one-sided.
Industry migration typically follows a three-step path: pluggable → MPO/NPO (near-package transitional forms) → CPO. In the near term (2026–2027), CPO is unlikely to disrupt the market immediately.
The more realistic scenario: MPO and NPO play a major role in 2027–2028; CPO and MPO enter a competitive or coexisting phase after 2028. Morgan Stanley projects CPO switch shipments will grow at a 144% CAGR from 2024 to 2030.
Why are smartphones getting squeezed instead of benefiting?
Morgan Stanley labels the smartphone segment one of the AI supercycle's "squeezed players." The core issue: memory costs (DRAM, NAND) are being driven up by AI demand, and handset makers absorb the price hikes.
Global smartphone shipments already fell about 3% year-on-year in Q1. Customers front-loaded purchases, inflating Q1 data; Q2–Q4 shipments are expected to decline, with further downside risk in H2.
Xiaomi's Q1 gross margin slightly beat expectations, but Morgan Stanley sees it as unsustainable — high-cost inventory will eventually compress margins. Transsion already showed gross-margin deterioration in Q1. A margin recovery likely requires full cost pass-through, probably arriving in Q4 this year or next year.
How has Asia-Pacific AI hardware performed this year, and who are the winners?
The sector showed sharp internal divergence in H1: the overall average gain was about 42%, but the median was only about 19% — a handful of leaders skewed the average.
Taiwan led with an average gain of roughly 82% and a median of about 49%. Hong Kong averaged about 44%, but the median was negative at roughly -10% — most companies actually fell.
This means → Morgan Stanley's conclusion is blunt: companies with a clear AI business were decisive winners; traditional smartphone supply-chain names broadly underperformed. The market is pricing "AI exposure" with real money.
Power, interconnects, enclosures — which downstream segments matter?
Morgan Stanley identifies three beneficiary threads in data-center infrastructure: rising per-rack power favors power suppliers and power-test equipment makers; growing data throughput drives demand for network switches and data interconnects; hardware-form diversification (GPU racks, inference racks, storage racks) creates incremental demand for mechanical and structural components.
On specific names, the bank highlights Bizlink (貿聯控股), citing three growth drivers: power interconnects, data interconnects, and semiconductor-equipment sub-modules. Data interconnects center on AECs (active electrical cables), with penetration into scale-out networks beginning in H2 2025.
Bizlink's recent gross-margin dip, driven by product-mix shifts, is judged temporary. Whether each AI hardware sub-segment can sustain order flow and earnings delivery in H2 will be the critical checkpoint for this expansion thesis.
Content is for reference only, not financial advice.