The Real AI Data Center Bottleneck Isn't the GPU: Where 800G, 1.6T, CPO and OCS Meet, Through Lumentum's Capacity Lens
At Barclays' 23rd Global Technology Conference, Lumentum Holdings Inc. global business unit president Wupen Yuen gave a critical signal the market often underestimates:
The AI infrastructure shortage has shifted entirely to the optical side.
This isn't one product line running tight. Lasers, EML, silicon photonics, modules and OCS are hitting supply constraints almost simultaneously.
Below, we break down the structural industry shifts this interview really revealed, from both a technology and a supply chain perspective.
1. This Isn't 'Stronger Demand' - It's Shortage Across the Board
Lumentum's description was blunt:
Every product line is capacity-constrained
EML and InP lasers are the tightest
Customer order visibility extends all the way into 2027
EML lasers: 2026 fully sold out, most of 2027 already booked
Which means one thing:
AI optical communications has moved from a cyclical business into a structural shortage.
This differs from the early days of 400G and 800G. Back then it was a technology generation transition; now it is a system-level demand surge.
2. Pricing Logic Has Flipped: Long-Term Contracts Don't Mean Discounts, They Lock In Supply
In a typical semiconductor market, long-term contracts usually come with price concessions.
In this AI optics cycle, it's the exact opposite:
Key customers are signing 2-year supply commitments
Prices are above historical levels
Prices stay 'high and flat' through the contract term instead of declining each year
The message is clear:
This is a textbook seller's market
Lumentum even has to allocate capacity selectively rather than chase orders.
3. Capacity Allocation Logic Revealed: EML > CW Laser > SiPh, but No All-In Bet
The interview spelled out a very real, very important capacity decision rule:
🔹 Core decision metric: gross margin per wafer
The ranking is clear:
200G EML
100G EML
CW laser (for SiPh)
At the same time, Lumentum deliberately reserves CW laser capacity, for three reasons:
Silicon photonics is the long-term trend
It has to be designed in to customer designs first
Lumentum's own modules also need internal laser supply
It's a classic trade-off between short-term margin and long-term platform position.
4. 800G and 1.6T: VCSELs Are Quickly Being Sidelined
The interview was very clear about where VCSELs stand:
800G
VCSELs made a brief comeback (multimode)
But share won't return to its early peak
1.6T
200G VCSELs have seen almost no large-scale adoption
Customers have mentally moved to single-mode architectures
A 30 m reach doesn't match the physical scale of new high-power AI racks
In one sentence:
The 1.6T generation has effectively pushed multimode VCSELs off the mainstream data center stage.
5. OCS (Optical Circuit Switch): Not Replacing Switches, but Layering On Top First
OCS was another highlight of the interview worth close attention.
Two deployment scenarios today:
Scale-up (within a rack / pod)
Scale-across (across buildings / campus level)
Lumentum's view is pragmatic:
Short term: supplement packet switches
Medium term: erode spine / NVLink Switch
Long term: rewrite the system architecture
So OCS growth isn't a zero-sum game; it rides on the explosive growth of total AI cluster scale.
6. The CPO Timeline Is Faster Than the Market Thinks
This was the most striking call in the interview:
By 2028, 40-50% of switches at the first wave of CPO adopters could be CPO-based
And:
Not just scale-out
Scale-up is being planned in parallel
Nearly every hyperscaler is already committed
The real bottleneck isn't technology, it's:
Lasers
Packaging
The overall pace of capacity ramp
7. Sizing the 2026 Optical Market
Lumentum's estimate of 2026 optical ports:
Metric | Volume |
Total optical ports | ~70–75M |
800G | ~55–60M |
1.6T | ~15–20M |
And that's only hyperscaler-driven demand.
8. Why Is Lumentum in No Hurry to Add Customers?
A counterintuitive but crucial strategic choice:
Serving just 3 top-tier customers already fills capacity.
There's only one reason:
Capacity is the real moat.
With lasers, DSPs and advanced packaging all short at the same time:
Winning a new customer doesn't mean you can deliver
Expanding too fast can damage existing trust
This also explains why in-house laser supply is becoming the dividing line among module makers.
Simple Tech Trend | Takeaway
The real signal from this interview:
AI's bottleneck has shifted from compute to optical capacity.
EML, CW laser, SiPh, CPO and OCS are entering a critical phase at the same time
Technology choices have given way to supply chain reality
Whoever controls the intersection of lasers x packaging x systems will define the next five years
This isn't a single-company story;
it's the prelude to a structural reshaping of the entire AI infrastructure industry.
Comments