Where Optics Win in AI Data Centers, and Who Approves the Change
AI clusters need more bandwidth than copper can carry over the distances and power budgets they run at. The question for an optics supplier is not whether AI needs more optics. It does. The question is where your optics create enough system value to replace what is there today, and who has to approve that change.
The optical opportunity in numbers
A $19 billion market today
Datacom optical component revenue passed $19 billion in 2025, with more than 42 million 400G and faster modules shipped. Almost 10 million 1.6T modules are forecast for 2026.[1]
A switch cycle approaching $1 trillion
AI back-end switch sales are forecast to approach $1 trillion from 2026 to 2030 combined, and scale-up networks become more than half of that market by 2030.[2]
Optics moving into the compute package
AMD, Broadcom, Meta, Microsoft, NVIDIA, and OpenAI formed the OCI MSA to write an open specification for optical scale-up links, with a multi-vendor supply chain as the stated goal.[3]
Power is the selling point
NVIDIA rates its co-packaged optics switches at 5 times the power efficiency of pluggable transceivers. Every watt saved per link is multiplied across tens of thousands of links.[4]
Approach the operator at the right point
An optical product passes through several companies before it reaches an operator. Good technology fails commercially when it approaches the operator at the wrong point in this chain, so we start by placing your product in it.
- 1Laser, modulator, or detector
- 2Photonic integrated circuit
- 3Optical engine
- 4Transceiver or co-packaged assembly
- 5Switch or accelerator platform
- 6System maker or contract manufacturer
- 7Operator
Where optics compete in an AI cluster
Each layer of the cluster has its own bottleneck, and optics answer a different question in each.
Compute
Constraint
Accelerators wait on data held by other accelerators.
Options
Larger scale-up domains and rack-scale systems.
What we check
Whether throughput per accelerator is limited by the network or by the math.
Scale-up links
Constraint
Copper loses reach and costs more power per bit as lane rates rise.
Options
Co-packaged optics (CPO), near-packaged optics (NPO), linear pluggable optics (LPO), and the OCI MSA optical link.
What we check
Power per bit, reach, serviceability, and multi-vendor supply.
Memory
Constraint
Model weights and conversation memory outgrow the memory attached to each accelerator.
Options
CXL memory expansion and pooled memory over optical links.
What we check
Added latency, bandwidth per node, and which workloads accept a remote memory tier.
Scale-out and reliability
Constraint
Link failures and failed transceivers stall large training and serving jobs.
Options
Linear optics, connector cleaning discipline, and link telemetry.
What we check
Failure modes by link type and the operating procedures behind them.
Power and facility
Constraint
Watts per link add up across tens of thousands of links.
Options
Lower-power optics and fewer retimed links.
What we check
Watts per link and per switch against the facility power budget.
What your account plan has to answer
- Where does the product live: scale-up, scale-out, memory fabric, or between data centers?
- At what data rate and reach?
- What does it improve: watts per terabit, cost per gigabit, density, latency, or reliability?
- Which switch and accelerator reference architectures support it?
- Who has to qualify it before the hyperscaler sees it?
What the industry is working on
At the Optica and APC Photonic-Enabled Cloud Computing (PECC) Summit (October 2026, Sunnyvale, California), the agenda centered on:
- Scale-up AI networks that span more than one rack
- Memory access and disaggregated memory
- Optical circuit switching
- Co-packaged optics manufacturing, packaging, and specifications
More from this practice
Executives sizing a market
AI Infrastructure Market Map
Where the trillion dollars goes, layer by layer, and which part of it your product can compete for.
Business development and sales leaders
Winning Hyperscaler Design-Ins
The five gates between a first meeting and a design win, and who inside a hyperscaler decides at each one.
Memory and CXL companies
AI Memory Wall
Why large models run out of GPU memory, which suppliers that creates room for, and what to prove.
Power, cooling, and facility suppliers
Power and Cooling Go-to-Market
How time to power and rack density became buying criteria, and how to sell into facility design.
Frequently Asked Questions
Datacom optical component revenue passed $19 billion in 2025, according to Cignal AI, with more than 42 million 400G and faster modules shipped and almost 10 million 1.6T modules forecast for 2026.
The Optical Compute Interconnect multi-source agreement, formed in 2026 by AMD, Broadcom, Meta, Microsoft, NVIDIA, and OpenAI, is writing an open specification for optical links that connect accelerators, so operators can buy them from more than one supplier.
By proving system-level value in the metrics the buyer judges, such as watts per terabit and reliability, getting supported in a switch or accelerator reference architecture, and qualifying with the module makers and system makers upstream of the hyperscaler.
References
Figures checked against these sources on October 22, 2026.
- [1] Datacom optical component revenue surpasses $19B in 2025 (Cignal AI, April 2026). https://cignal.ai/2026/04/datacom-optical-component-revenue-surpasses-19b-in-2025/
- [2] AI back-end switch sales to approach $1 trillion over the next five years (Dell'Oro Group, July 2026). https://www.delloro.com/news/ai-back-end-switch-sales-to-approach-1-trillion-over-the-next-five-years/
- [3] Optical Scale-up Consortium established to create an open specification for AI infrastructure (TelecomTV, March 2026). https://www.telecomtv.com/content/next-gen-telco-infra/optical-scale-up-consortium-established-to-create-an-open-specification-for-ai-infrastructure-55073/
- [4] NVIDIA silicon photonics networking. https://www.nvidia.com/en-us/networking/products/silicon-photonics/
Bringing an optical product to AI data centers?
Bring the product and the architecture you are targeting. We place it in the supply chain, find who has to approve it, and build the proof they need.