AI Data Center and Silicon Photonics
Advisory, product, and go-to-market work for the optical interconnect and memory systems that large AI clusters depend on, from scale-up links to CXL memory pooling.
Direct answer
AI capacity is now limited by the links between chips and by memory.
Large mixture-of-experts models spread their weights and KV cache across many accelerators, so every layer moves data between chips. Copper links are reaching their power and reach limits, which is why the industry is moving scale-up and memory traffic onto optics and CXL. Tactical Edge helps the companies building and buying that infrastructure decide what to build, how to prove it, and how to sell it.
Photonics and interconnect companies
Product definition, standards strategy, and go-to-market for optical engines, CXL optical links, and optical switching.
AI data center operators and neoclouds
Architecture reviews of scale-up domains, memory tiers, and link reliability before capacity is committed.
Hyperscaler and enterprise infrastructure teams
Vendor evaluation and pilot design for optical interconnect and pooled memory in existing clusters.
How to start
Send a short request
Tell us the outcome you want through the contact form. We reply within one business day.
Scope it together
One conversation covers the use case, systems, data, team responsibilities, and operating environment.
Review a written proposal
You get deliverables, assumptions, customer inputs, milestones, fees, and how scope changes are handled.
From compute to interconnect
Each layer of an AI cluster now has its own bottleneck. The Optical Compute Interconnect (OCI) MSA, launched at OFC 2026 by AMD, Broadcom, Meta, Microsoft, NVIDIA, and OpenAI, is one sign of how far scale-up links are moving toward optics.
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 fabric 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 PHY.
What we check
Power per bit, reach, serviceability, and multi-vendor supply.
Memory
Constraint
Model weights and KV cache outgrow the memory attached to each accelerator.
Options
CXL memory expansion and pooled, disaggregated memory over optical links.
What we check
Added latency, bandwidth per node, and which workloads tolerate a remote memory tier.
Scale-out and reliability
Constraint
Link flaps 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 We Do
Technology and architecture assessment
Review a product or cluster design across compute, interconnect, and memory, and find where the fabric sets the ceiling.
- Scale-up and scale-out topology review
- Memory tier and CXL fit by workload
- Power, reach, and reliability trade-offs
Product definition
Turn a photonics or CXL capability into a product requirements document buyers can evaluate.
- Target workloads and form factors
- Standards and interoperability plan
- Specifications and proof points
Go-to-market with hyperscalers and AI data centers
Plan how a product reaches design-in at cloud providers, neoclouds, and system makers.
- Target account and partner mapping
- Benchmark and reference design story
- Launch, standards body, and partner programs
Market intelligence
Track the AI data center supply chain with the Prospectory AI Infrastructure Leaderboard.
- Capital spending from SEC filings
- Funding, hiring, grants, and patents
- Standards and consortium membership
The spending behind the build-out
Capital spending over the latest four reported quarters, through June 30, 2026 (Oracle: August 31, 2026), from SEC filings as of October 16, 2026. The figures cover all property and equipment, not only AI data centers.
See all 377 companies on the leaderboard| Company | Four quarters |
|---|---|
| Amazon Web Services | $173.0B |
| $132.4B | |
| Microsoft | $115.9B |
| Meta | $89.3B |
| Oracle | $75.7B |
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
Our team's experience
- Led AI and machine learning go-to-market at AWS for Worldwide Public Sector and Regulated Industries, including the AWS Inferentia and Trainium accelerator chips.
- Worked inside two hyperscalers: AWS, and Microsoft, where we directed Azure Data and AI work with US federal agencies.
- Held a vice president role at Luminous Computing, a photonic AI computing company.
- Advised Lightelligence from 2021 to 2024 on its optical fabric and the launch of Photowave, CXL optical interconnect hardware for disaggregated AI memory.
- Authored AWS whitepapers and contributed to the AWS Well-Architected Machine Learning Lens.
Photowave specifications are published on Lightelligence's product page.
Frequently Asked Questions
Tactical Edge reviews AI cluster and product architectures across compute, interconnect, and memory, writes product requirements for photonics and CXL products, and plans go-to-market with hyperscalers, neoclouds, and system makers.
Yes. We assess which workloads, such as KV cache for long-context inference, can use a CXL memory tier, what latency and bandwidth that tier adds, and whether optical links are needed to reach pooled memory across racks.
The leaderboard, built on Prospectory, tracks companies across the AI data center supply chain using SEC filings, job boards, standards bodies, grants, and patents. We use it to size markets, map target accounts, and follow competitors.
Building or buying AI infrastructure?
Bring one product, cluster design, or launch plan. An assessment maps where the fabric and memory set your limits and what to do first.