Photonics: The AI Bottleneck
---*@henryvzero
· 1
−11.5%
since Sep 30
The idea
I believe copper hits its limits as AI clusters scale, so moving data with light becomes the next bottleneck in the buildout, and the next couple of years favor whoever is already shipping optical gear. The basket spans the whole chain: laser and transceiver makers, the contract assemblers running near capacity, chipmakers pushing optics onto silicon, and the networking vendors selling finished systems, weighted toward the names with the most direct optical revenue today.
What it owns
- The chipmakers moving light closer to the processor, which is where the next wave of design wins land. · Broadcom Inc.Furthest along at putting optics right on switch silicon, which means it wins whether the industry shifts to co-packaged optics or stays with pluggable transceivers., Marvell TechnologyMakes the digital signal processor chips inside most high-speed optical transceivers, a near-unavoidable toll on every fast link., Credo TechnologySells active cables and retimers that stretch copper as far as it can go, the exact pressure point that forces the move to light., Astera LabsSells connectivity chips that tie GPUs, memory, and networks together inside AI clusters, the plumbing that scales before optics even matter., Taiwan SemiconductorBuilds silicon photonics and the advanced packaging that co-packaged optics require, so it gets paid whichever chipmaker wins., NVIDIA CorporationPushes photonics into its own switches and defines what the whole cluster needs, so its buildout pace sets demand for everyone else in this take.31%
- These companies make the lasers and optical parts every AI data link needs, the raw material of the whole buildout. · Coherent Corp.Makes lasers inside datacenter transceivers, the exact part that becomes scarce if AI data has to move as light instead of copper., Lumentum HoldingsSells both lasers and finished transceivers that connect AI racks, so it gets paid twice if optics become the bottleneck., MACOM Technology SolutionsMakes the drivers and amplifier chips inside nearly every optical link, so it wins across transceiver makers rather than betting on one., Corning IncorporatedMakes the glass fiber that all this light travels through, inside datacenters and between them, the steadiest way to own the bottleneck., Applied OptoelectronicsSmaller transceiver maker that wins if the supply squeeze gets real and buyers reach further down the supplier list to find capacity.26%
- The factories that build the transceivers and optical gear; when demand outruns supply, they run full and pricing improves. · FabrinetAssembles a huge share of the world's high-end optical transceivers and is already running near capacity, so utilization and pricing show the squeeze first., Celestica Inc.Builds the AI networking switches and systems hyperscalers order directly, so its backlog is a live read on buildout acceleration., Amphenol CorporationSells the connectors and cabling in every rack, copper and optical alike, so it collects on the buildout no matter how fast the shift to light happens., Jabil Inc.Brings big contract manufacturing scale to photonics and AI hardware, the extra capacity the chain leans on when dedicated assemblers fill up.19%
- The vendors selling the finished switches and optical transport boxes that datacenters buy to move AI traffic. · Ciena CorporationMoves light between datacenters, the other half of the bottleneck once traffic leaves the building., Arista NetworksSells the finished switches that AI clusters plug into, and its AI networking orders are the clearest sign buildout money is flowing.9%
- Every new optical link gets tested before it ships, so test gear orders track the pace of the buildout. · Keysight TechnologiesSells the test gear that verifies every new faster optical link before it ships, a picks-and-shovels way to own the transition., Viavi SolutionsSells instruments that verify optical links in labs and live networks, so demand tracks how many new links get built and deployed.5%
- Cash10%