AI CLUSTER PHOTONIC FABRIC PANEL REPORT
4/18/26
SEQH CAPITAL RESEARCH – TEAR SHEET
AI CLUSTER PHOTONIC FABRIC PANEL – PHOTONICS FABRIC TIGHTNESS PREVIEW
WHAT THIS REPORT ANSWERS
The report uses SEQH’s 23‑cluster AI Photonic Fabric Panel to quantify how “photonics‑tight” today’s and roadmap AI clusters are, via a 0–100 Photonics Fabric Tightness Score (PFTS), and then maps that directly into where value is accruing across photonics, switch/ASIC, connectivity, and precision‑timing equities.
It shows a clear architecture break between Rubin NVL144 (PFTS 40) and Rubin Ultra NVL576 (PFTS 90) as CPO, OCS, lane‑rate doubling, and rack‑level thermals force a shift from DSP pluggables to CPO and external‑laser fabrics.
CORE THESIS & PFTS FRAMEWORK
Scale‑up bandwidth is outgrowing what DSP‑based pluggable optics can handle on power, thermals, and jitter, so the next GPU generations (Rubin, Feynman) must move scale‑out and parts of scale‑up to photonics, with CPO and OCS as the critical tools.
PFTS blends five weighted factors: CPO share (30%), optical ports per GPU (20%), inverse pJ/bit (20%), OCS presence (15%), and rack thermal density (15%), and classifies clusters into Low (<40), Medium (40–79), and Elite (≥80) photonics tiers.
Hopper/Blackwell live in the Low band (PFTS 11–24), Rubin NVL144 sits around 40, and Rubin Ultra/Feynman hit Elite at 90 and 98, where copper physics fails and CPO becomes architecturally mandatory.
KEY ARCHITECTURAL FINDINGS
The three‑tier interconnect model, scale‑up (<1 m), scale‑out (<2 km), scale‑across (2–4,000 km), now has distinct photonic economics, with:
Scale‑up still copper‑heavy (NVLink + DACs) but trending toward CPO at Rubin Ultra and beyond.
Scale‑out transitioning from 800G DSP pluggables to 1.6T LPO and then CPO, driven by a ~4–5× pJ/bit improvement (25 pJ → 5.6 pJ and toward ~4 pJ).
Scale‑across remaining coherent‑optic dominated (CIEN/COHR/MRVL).
Rack‑density vs optical‑ports‑per‑GPU plots isolate Rubin Ultra (~600 kW, ~3 ports/GPU) and Feynman (~800 kW, ~3.5 ports/GPU) as the first clusters where scale‑up itself becomes photonic, and DSP pluggable power density ceases to be viable.
OCS is a major discriminator: Google’s Palomar/Jupiter‑style OCS fabrics earn a 15‑point PFTS bump and clearly separate Google‑family TPU pods from Ethernet‑only peers.
WHO BENEFITS – EQUITY MAP
The component layer is anchored by Lumentum (LITE) and Coherent (COHR) as NVIDIA’s disclosed Quantum‑X photonics suppliers; both elevate from “Major” to “Anchor” starting at Rubin NVL144 in the dataset.
The ASIC/switch layer is driven by Broadcom (AVGO) (Tomahawk 6, TH6‑Davisson CPO) and NVIDIA (NVDA), with Marvell (MRVL) supplying DSPs and custom AI silicon as CPO and 1.6T optics move into volume.
Connectivity names Credo (CRDO) and Astera Labs (ALAB) appear across nearly every Blackwell‑and‑later cluster via AECs and PCIe/CXL retimers, while Fabrinet (FN), AAOI, and MTSI sit in the manufacturing/analog stack with rising content per port as lane rates and CPO penetration increase.
SiTime (SITM) is surfaced as a hidden precision‑timing beneficiary: MEMS oscillator content per optical module roughly doubles moving from 800G to 1.6T, and SiTime’s Renesas timing acquisition plus Super‑TCXO/TimeFabric stack give it Anchor status on the timing layer.
WHAT PAID MEMBERS GET IN THE FULL REPORT
Upgrade to access the full AI Cluster Photonic Fabric Panel report, including:
The full 23‑cluster dataset (36 fields × 10 sheets) with PFTS construction, optical mix, thermal/power, economics, vendor‑exposure matrices, and per‑cluster topology detail.
Visuals and tables showing how optical ports/GPU, pJ/bit, and CPO share evolve from Hopper/Blackwell through Rubin, Rubin Ultra, and Feynman, and how that maps into specific component, ASIC, connectivity, and timing names.
A 14‑ticker photonics/AI‑interconnect snapshot (LITE, COHR, SITM, MRVL, AVGO, CRDO, ALAB, FN, AAOI, ANET, NVDA, INTC, CIEN, MTSI), with condensed catalyst book and a thematic exposure matrix (CPO, LPO, OCS, timing, 1.6T).
An appendix deep‑dive on Sivers (SIVE) as an upstream CPO external‑laser (DFB array) supplier and a detailed methodology section so institutional clients can re‑weight PFTS and vendor exposures to fit their own process.
Q2 / EASTER PROMO – 20% OFF YEARLY FOR LIFE
For investors who want ongoing access to this photonic‑fabric series plus SEQH’s broader AI‑infrastructure and nuclear/energy work, we’re running a limited Q2 / Easter promotion on our yearly subscription:
20% discount on the annual plan.
Locked “for life” as long as the subscription remains active.
Applies to all premium photonics, AI‑interconnect, and energy research going forward.
You can activate the promo and lock in the lifetime discount here:
CLICK FOR Q2 PROMO LINK AND FULL SUBSTACK ACCESS
FULL 19-PAGE FABRIC PANEL REPORT BELOW:
