Morgan Stanley: AI Could Require Up to 70 Optical Engines per GPU, Optical Content Becomes a Core Variable
Alina Collins
Morgan Stanley's latest report argues the number of optical engines per GPU can rise from today's 2–4 to as many as 70, making optical content per chip — not shipment volume — the variable that matters most for the photonics trade, regardless of which packaging architecture wins.
How many optical engines does one GPU need?
Today each GPU pairs with roughly 2–4 optical engines. Under the NVL576 hybrid architecture that rises to about 17; a full CPO design reaches about 35; a bull-case CPO scenario pushes it to roughly 70.
This means → the question is no longer whether a GPU needs optics — it is how many engines fit on each chip. Density itself is the growth curve.
Lumentum's content value per optical engine runs about $150–200. Total value per system grows on two axes simultaneously: more units shipped and more engines packed per GPU.
How big is the CPO demand opportunity?
Morgan Stanley forecasts 2027 CPO optical-engine demand closer to 6–7 million units, not the 20–30 million some investors expect.
In plain terms = the bank is using a conservative number, yet even at that pace the CPO switch market's 2024–2028 CAGR is roughly 144%.
The report sees higher certainty for mass adoption after 2028–2029. Whether actual 2027 demand lands at 6–7 million or moves toward 20–30 million is the key checkpoint for whether the thesis plays out early.
What is the real relationship between NPO and CPO?
NPO — near-package optics, which keeps the optical engine slightly separated from the GPU package — lowers production, assembly, and maintenance complexity. Optical engines can be sourced independently from GPUs, opening the supply chain.
This means → NPO is better understood as a transitional step toward deeper optical interconnects, not a rejection of the CPO direction.
The real bottleneck for near-term CPO ramp is not the photonic IC process itself. It sits in back-end packaging, optical alignment, module-level testing, and system assembly — SoIC yields currently run about 50–60%, and downstream CPO switch assembly yields only about 20–50%.
Where exactly are the bottlenecks?
Fiber-array-unit alignment, single-mode fiber coupling, and test-equipment maturity all remain constraints.
In plain terms = these are engineering-yield and supply-chain-maturity problems — no one disputes the direction; what's stuck is the manufacturing detail.
The pace at which SerDes speed advances from 200G to 400G will directly determine whether optical-engine counts per GPU can keep doubling. Morgan Stanley flags this as a key verification point to track.
Which companies benefit? The logic is not tied to "pure CPO" alone
Lumentum (LITE): benefits from optical engines, lasers, and OCS opportunities in AI datacom. If the 1.6T AI ramp arrives faster than expected, margin upside is more pronounced.
Coherent (COHR): benefits from rising AI datacom demand, pricing improvement, and margin leverage.
Corning (GLW): benefits from data-center fiber and glass-material demand. If CPO adoption exceeds expectations, the foundational optical-materials layer stands to gain further.
This reflects Morgan Stanley's core call: the market is over-indexing on the architecture debate while overlooking the fact that optical-content growth is still in its early innings — after the recent pullback, valuations for all three names look more attractive.
Content is for reference only, not financial advice.