AI’s Next Arms Race Isn’t Just in GPUs — It’s in Optical Interconnects
3.2T and 6.4T solutions are showing up at ECOC 2026. The next AI bottleneck may be shifting from raw compute to connectivity.
As AI clusters scale from thousands to hundreds of thousands of GPUs, one problem is becoming increasingly important:
How do you move massive amounts of data between those GPUs fast enough — and without burning too much power?
That is exactly why optical interconnects are becoming a bigger part of the AI infrastructure story.
At ECOC 2026, Coherent showcased a 3.2T OSFP optical module as well as a 6.4T NPO optical engine designed for next-generation AI scale-up and scale-out networks.
The direction is becoming clear:
800G → 1.6T → 3.2T
And the upgrade is not just about higher headline speeds.
The bigger shift is that optics are moving closer to the chip.
As GPU density rises, traditional electrical interconnects face growing limits in bandwidth, distance and power efficiency. That is pushing the industry toward architectures such as NPO and CPO, where optical components sit much closer to switches and compute chips.
Marvell is also showcasing 400G/lane PAM4 technology built on 2nm, targeting future 3.2T interconnects, while 1.6T solutions are moving closer to broader deployment.
This matters because AI infrastructure spending is expanding beyond GPUs.
The first bottleneck was compute.
Then came HBM.
Then power.
Now, networking is increasingly part of the constraint.
A powerful GPU is only as useful as the system around it. If data cannot move efficiently between accelerators, expensive compute ends up waiting.
That means AI capex is gradually spreading into:
optical modules, DSPs, lasers, silicon photonics, switches, CPO/NPO and data-center interconnects.
Tiger View
The key takeaway from ECOC 2026 is not that 6.4T is suddenly ready for mass deployment.
Many of these technologies are still in demonstration or validation stages.
What matters more is the pace of innovation.
Optical networking is evolving quickly from 800G to 1.6T and toward 3.2T, while the architecture itself is changing from pluggable optics toward NPO and CPO.
The goal is simple:
move more data, faster, with less power.
If GPUs are no longer the only bottleneck in AI infrastructure, then the companies that help connect compute more efficiently could become increasingly important in the next phase of the AI buildout.
AI’s next arms race may not only be inside the GPU.
It may be in the light.
$Coherent(COHR)$ $Lumentum(LITE)$ $Marvell Technology(MRVL)$ $Broadcom(AVGO)$ $NVIDIA(NVDA)$
Disclaimer: Investing carries risk. This is not financial advice. The above content should not be regarded as an offer, recommendation, or solicitation on acquiring or disposing of any financial products, any associated discussions, comments, or posts by author or other users should not be considered as such either. It is solely for general information purpose only, which does not consider your own investment objectives, financial situations or needs. TTM assumes no responsibility or warranty for the accuracy and completeness of the information, investors should do their own research and may seek professional advice before investing.
I am particularly interested in the optical side because bandwidth and power efficiency will become increasingly important as AI data centers scale. Companies like $COHERENT(COHR)$ , $Lumentum(LITE)$ $Marvell Technology(MRVL)$ and $ASML Holding NV(ASML)$ could benefit if optical interconnects take a larger share of AI capex.
I would still separate demonstrations from mass deployment. I am not chasing the 6.4T headline, but I want to follow NPO and CPO closely. I am bullish on AI and semiconductors, and the next opportunity may be not only in the chips, but also in connecting them.
@Tiger_comments @TigerClub @TigerStars @WallStreet_Tiger
ECOC 2026 is showing why: Coherent is demonstrating 3.2T pluggable optics and a 6.4T NPO engine, while Marvell is pushing 400G/lane technology toward 3.2T.
The deeper story is power per bit. As electrical links become harder to scale, optics are moving closer to the chip through NPO/CPO.
For investors, this expands the AI infrastructure map beyond GPUs into DSPs, lasers, silicon photonics, optical modules and packaging.
3.2T and 6.4T are still largely technology demonstrations—not mass deployment. But the direction is clear: AI may increasingly be limited by how fast machines can talk, not how fast they can compute.
The next AI arms race could be fought in photons, not transistors.
@Tiger_comments [思考]