Jeff Bezos will be on his knees

Power is the constraint that is reorganizing everything in AI infrastructure. Not capital. Not demand. Power. And that single fact is why a relatively obscure technology called Optical Circuit Switching may be the most consequential procurement shift happening inside hyperscale data centers right now.

Start with the market picture. The worldwide Ethernet switch market hit $15.4 billion in Q1 2026 alone, up 39.8% year over year, according to IDC, with the data center segment surging 61% to $10 billion as AI training and inference buildouts accelerated. Arista Networks sits squarely at the center of that spending wave. Its Q2 2026 revenue came in at $3.036 billion, up 37.7% from a year prior, and the company raised its full-year 2026 guidance to roughly $12.6 billion, implying 40% growth. Etherlink AI fabric customers have now crossed 100. The stock has rewarded shareholders accordingly.

But another procurement cycle is forming underneath that one. It involves different physics, different vendors, and a structural argument that the biggest buyers of networking hardware are starting to find compelling.

What OCS Actually Does

Optical Circuit Switching routes data by deflecting light with tiny MEMS mirrors rather than processing packets electronically. No optical-to-electrical-to-optical conversion at the spine layer. The practical consequence is striking: Google’s own internal deployment, documented in its Jupiter Evolving paper, demonstrated a 41% reduction in network power consumption when OCS replaced electrical spine switches. For a hyperscaler whose next data center is gated on available megawatts, that number is not an engineering footnote. It is a capacity argument.

Google does not sell its internal OCS technology. That leaves the external market open. Some analyst estimates for 2026 call for total OCS demand above 20,000 units, with Google representing the bulk of that volume, predominantly using its own proprietary systems. The externally procured slice is smaller but growing fast, and two companies are capturing most of it: Lumentum and Coherent.

The Backlog Story

Lumentum’s OCS trajectory is difficult to ignore. Its backlog has surpassed $400 million across multiple hyperscaler customers, with the company noting it cleared its first $10 million OCS quarter three months ahead of schedule. Management has described that backlog as shipping largely in the first two quarters of its fiscal 2027, and the company’s stated target is a run rate above $1 billion in 2027. Lumentum has also pointed to a multi-year, multi-billion-dollar OCS purchase agreement, and said it has three customers with two accounting for the majority of volume.

Coherent’s position is meaningful as well. The company has said it has received more than $300 million of OCS orders, with Google and Oracle as its main customers, and expects OCS revenue to build as production ramps.

The Arista Connection

There is a second-order read-through worth tracing here. Arista has been candid about its component availability problem. Non-cancellable purchase commitments rose to $9.7 billion as of June 30, 2026, reflecting multi-year agreements for chips and other components. The bottleneck is not customer demand. It is silicon availability.

OCS, because it steers light rather than processing electrons, requires far less silicon per port and consumes dramatically less power per bit moved. Hyperscalers that move their spine layer to optical are reducing dependence on precisely the components in shortest supply. The two procurement waves are not competing. They are linked.

Risks Worth Considering

OCS adoption is not without friction. Optical switching is far slower than electrical packet switching when reconfiguring paths, which limits its usefulness at the leaf layer where traffic patterns are less predictable. Supply chain constraints at Lumentum have required direct executive attention to ramp production. And the customer concentration risk is real: two buyers represent the majority of current volume.

Still, the structural pull is hard to dismiss. A technology that cuts spine-layer power by 41% while reducing silicon content per port is not a niche product in an era when power availability is capping new capacity builds. Lumentum and Coherent are not familiar names in mainstream networking coverage. That gap may be where the opportunity sits.