Perspectives Research Terminals Weekend RWA Board

VIEW

››Haelangdal's View

EDITORIAL VIEW

How AI Gets Connected — Reading the Connectivity Value Chain as Layers

Dissecting AI infrastructure's third bottleneck, 'connectivity,' from the basics (scale-up and scale-out) to the 8-layer value chain — where copper's time and light's time split

HHaelangdal·Founder AnalystJune 7, 202622 min readHaelangdal's View

Global Investment through Themes

Content

Themes Ideas Reports Earnings

Explore

Guide ETFs Perspectives Telegram NVIDIA Portfolio Weekly Review Credit & Leverage

Community

Legal

Privacy Policy Terms of Service Disclaimer

© 2026 NewMoneyMoves. All rights reserved.

Bottom Line

AI connectivity is a theme bought as a 'layer,' not a single stock. Today's center of gravity is copper, connectivity silicon, substrates, and assembly; light comes later. And light arrives not 'inside the rack' but 'between racks' first.

Reader's Brief — 30-second TL;DR

Intermediate

Why Now

At Computex in June 2026, NVIDIA's Jensen Huang formalized connectivity as the next bottleneck and called Marvell the 'next trillion-dollar company,' and connectivity began to be priced as an independent investment axis.

Winners ?? Losers

Copper and connectivity-silicon center of gravity — Credo (AEC ~73%), Astera Labs (connectivity silicon), Amphenol (connectors). Substrate and assembly — Isu Petasys (Korea, global big three), Fabrinet (optical assembly). Light's leverage (beyond 2028) — Coherent, Lumentum (external light sources), Marvell (copper and light). Triggers — how NVL576 links rack to rack, whether Feynman uses co-packaged copper, external-light-source supplier lock-in.

Watch For

Reading depth

1. A Single GPU Can Do NothingAI's power comes not from the speed of one GPU, but from how well many GPUs are bound together.Jump to section
2. Two Words to Know First — Scale-up and Scale-outScale-up is close, tight, fast, mostly copper. Scale-out is far, wide, large, mostly light. The two are not rivals but a division of roles.Jump to section
3. Inside a Real Data Center — Dissecting the NVL72 RackInside the NVL72 rack is copper; between racks is light. In one data center, copper and light split roles and work at once.Jump to section
4. Why Copper Gives Way to Light — A Matter of DistanceWhat splits copper and light is distance, not speed. Copper hits no cliff; it hands the baton to light at its distance limit.Jump to section
5. The Ladder of Connectivity — DAC, AEC, AOC, Transceiver, CPOCloser means copper (DAC to AEC); farther means light (AOC to transceiver to CPO). The data center picks from this ladder by distance.Jump to section
6. The Bottleneck Moves — From Compute to ConnectivityConnectivity is no longer a sub-theme attached to AI chips. It has begun to be priced as an independent investment axis.Jump to section
7. Where the Market Got It Wrong — The Real Optical-Transition TimelineThe '2027 intra-rack optical transition' conflicts with NVIDIA's roadmap. Light enters between racks, not inside them, tested in 2027, at volume past 2028.Jump to section
8. Why Copper Won't Die — And 'Co-Packaged Copper'Co-packaged copper is the single biggest variable in the optics/copper mix. It extends intra-rack copper past 2028; if light reaches the GPU, light-source demand pulls forward.Jump to section
9. The Investment Timeline — Copper's Time and Light's TimeIt is not 'just watch optical names.' Copper grows too now. Weighting scale-up optics can wait until NVL576's 2027 CPO adoption is confirmed.Jump to section
10. Map of the 8-Layer Connectivity Value ChainThe connectivity chain runs eight layers from chip to assembly. Korea's only top-three layer is substrates; the rest center on the US, Taiwan, and China.Jump to section
11. Detailed 8-Layer Mapping and Korea's CoordinatesKorea's coordinate is clear at one layer, substrates. Needed whether light or copper wins, this layer is a center of gravity neutral to the optical debate.Jump to section
12. Who Stands at Which LayerThe connectivity center of gravity is the US, Thailand, Taiwan. Credo near-monopolizes 73% AEC, Coherent and Lumentum hold light sources, weighted past 2028.Jump to section
13. What to TrackThree triggers decide when and where light descends. Watch the order in which NVL576's link, Feynman's co-packaged copper, and light-source lock-in switch on.Jump to section
14. Conclusion — A Theme Bought as LayersConnectivity is a theme bought as layers, not one stock. Now copper, substrates, assembly; later light, which comes between racks, not inside them.Jump to section

1. A Single GPU Can Do Nothing

AI news always features GPUs. Yet a large model like ChatGPT does not run on a single GPU. The model is too big to fit in one chip's memory. So data centers bind tens to thousands of GPUs into one giant computer.

Chips must then exchange data ceaselessly. This act of 'linking each other' is . No matter how fast a chip is, if data cannot move in time, the chip idles. Jensen Huang put it plainly: "If data cannot get out, a GPU is just an isolated island." Connectivity is the bridge between islands.

AI's power comes not from the speed of one GPU, but from how well many GPUs are bound together.

Takeaway

AI's power comes not from the speed of one GPU, but from how well many GPUs are bound together.

Full access requires 🥉 Bronze tier

Sign in with Google — your tier will be checked automatically and access granted if eligible.

Sign in with Google

This site runs on ads — the tier system rewards community contributions.

Comments

This report is provided for informational purposes only and does not constitute a recommendation to buy or sell any financial instrument. Investment decisions should be made based on your own judgment and responsibility. The analysis and opinions contained herein are based on information available at the time of writing and are subject to change.