Verdict: Demand-side narrative CONFIRMED at four mechanically independent layers. Positioning now CONTESTED (Wave 1). The "why did they pop" question has a coherent four-layer answer rooted in physics (GPU-to-GPU bandwidth scaling outrunning copper backplane), corroborated by primary-source Q3 FY26 prints (COHR May 6: $1.81B record revenue, datacomm +37% YoY, backlog into calendar 2028), and validated by an external voice arriving at the same framework independently (NuttyCLD's Last Micrometer series, March–April 2026). The system did this work in March–May 2026; the right artifact today is a verdict that consolidates the demand-side narrative for the Ciena/Coherent "why did they pop" backlog question and points at the active perspective for continued tracking, without re-doing the perspective's work.

What we're asking

The open question (Priority 1, semi-blind):

Source: CIEN +753% 1Y, COHR +450% 1Y. Our agents identified them as bottlenecks but we haven't traced the specific DEMAND catalyst. Thesis: Something specific changed in optical networking demand in late 2025/early 2026. What?

Specific questions the source asks:

• What specific hyperscaler buildout decisions drove optical demand?
• When did 800G start shipping in volume? Who was first?
• What's the 1.6T timeline and who's leading?
• DCI (data center interconnect) vs intra-DC — which is growing faster?
• Did a specific NVIDIA/AMD GPU launch change the interconnect math?

This is a semi-blind item: we have the price data but not the demand narrative. The artifact is the demand narrative.

What we already had before today

This investigation is post-perspective consolidation, same shape as the HBM and tungsten verdicts:

So the perspective + trace + Q3 print already answer the backlog question; what this investigation does is consolidate the demand-side narrative in one place for closeout.

The "WHY did they pop" answer — four mechanically independent layers

The CIEN +753% 1Y / COHR +450% 1Y move is not one thing; it's four mechanically separable layers stacking on the same calendar.

Layer 1 — GPU-cluster scale-out math (the physics)

GPU-to-GPU bandwidth requirements scale superlinearly with cluster size. NVIDIA Blackwell GB200 (Q4 2025 ramp) → GB300 (Q1–Q2 2026 ramp) → Vera Rubin (2027) each requires more inter-GPU bandwidth than the prior generation:

• Per-GPU NVLink bandwidth climbed from H100 (900 GB/s NVL4) to GB200 (1.8 TB/s NVL72). Inside-the-rack copper-backplane works at this density.
• Inter-rack scale-out fabric at NVL72 → NVL576 → cluster of clusters needs orders of magnitude more optical throughput than the prior generation. 800G optical transceivers per GPU socket is the baseline; 1.6T is the next step.
• Energy budget per bit transmitted on copper traces is ~2–3 pJ/bit at >100Gbps line rates; on optical is <1 pJ/bit at 200G/lane. Power-budget breaks copper at the high end.

NuttyCLD's "light moves into the package wherever electrical signaling becomes too power-hungry" is the one-liner version. The shift point is the GB200 generation — that's when optical I/O became mandatory for the scale-out fabric, not optional.

Layer 2 — 800G ramp + 1.6T pull-forward (the merchant ASP/volume story)

The 800G optical-transceiver generation (using 200G-per-lane modulation) started shipping in volume during 2H 2025 alongside the GB200 ramp. The 1.6T generation (200G-per-lane × 8 lanes) is on the Q3 2026 commercial-shipment timeline per AAOI's stated guidance ($200M+ order book for 1.6T noted in the optical-supercycle catalyst calendar).

Per-customer math: a 100K-GPU cluster running on 800G optics consumes ~100K transceivers per side, ~$2K ASP each, ~$200M optical-transceiver capex per cluster. There are a half-dozen named US/Taiwanese/Korean hyperscaler buildouts at this scale or larger across 2025–2027. The merchant TAM that COHR/LITE/CIEN/FN address roughly doubled inside 18 months, with the next doubling already on the order book.

The "why pop now" specifically: the 800G ramp moved from sample/pilot to volume across 2H 2025 → 1H 2026, exactly the window where the prices ran. CIEN's role is on the WAN/DCI side (datacenter-interconnect, AI-cluster east-west fabric between sites); COHR's role is at the laser-source level (DFB lasers, EML lasers, InP epi) plus precision optics. Both layers ramped together because both are needed.

Layer 3 — DCI vs intra-DC (the answer to the source question)

The source asks: "DCI (data center interconnect) vs intra-DC — which is growing faster?"

Intra-DC is the bigger TAM but DCI is the higher-growth rate in 2025–2026. The reason:

• Intra-DC (within a single datacenter) scaled with 800G transceivers per GPU socket. Volume is huge but ASP is mid-range and the architecture is already mature.
• DCI / WAN scaled with multi-site AI-clusters — the GB200/GB300 cluster sizes outgrew what one datacenter site can power (the AI-power bottleneck, see ai-power-bottleneck perspective). Customers run AI-workloads across multiple sites linked by coherent DCI. This was a structural change, not an incremental capacity add.
• CIEN specifically is overweighted to coherent DCI / WAN (WaveLogic 6 platform); that's the layer growing fastest.
• COHR/LITE/FN address both layers; the intra-DC volume is the bigger bucket of revenue, but DCI is what changes the ASP mix.

Layer 4 — CPO endgame + power-budget crossover (the future-state thesis)

Co-Packaged Optics (CPO) is the post-pluggable architecture where the optical engine sits inside the GPU package rather than at the rack-edge. This is the Wave 4 thesis in the perspective.

• NVIDIA Quantum-X Photonics + Spectrum-X Photonics (H2 2026 announced timeline) are the first NVIDIA-branded CPO switch platforms. They're the catalyst that confirms CPO as the architectural direction.
• The driver is power, not bandwidth. Pluggable optics at 1.6T+ blow through the per-port power budget; CPO collapses the laser into the package and saves ~30–50% of the optical I/O power per bit.
• NuttyCLD's "Last Micrometer" thesis (April 25 2026 article) arrives at the same conclusion from the power-delivery side: the same physics that drives 800V DC into the rack also drives the laser into the package. Two paradigms (power-delivery + bandwidth-density) converging on the same architectural answer.
• LWLG (polymer modulators), POET (optical interposer), and SIVEF (TFLN modulators) are the pre-revenue speculative names for this layer. Wave 4 stays binary.

Did a specific NVIDIA/AMD GPU launch change the interconnect math?

Yes — the GB200/NVL72 generation. NVIDIA's GB200 platform, with NVLink 5 and 1.8 TB/s of per-GPU bandwidth, was the inflection where optical I/O moved from "nice-to-have at the WAN edge" to "mandatory for the scale-out fabric." Prior generations (H100 / A100) had enough headroom on copper traces; GB200 didn't. GB300 (Q1–Q2 2026) extended the demand curve; Vera Rubin (2027) and Feynman (2028) compound it. AMD's MI300X / MI325X had a similar inflection but smaller volume so didn't dominate the demand curve.

State-of-coverage by ticker

Pricing as of the 2026-05-23 scan summaries (no fresh tape pulled per the producer-only constraint).

Wave 1 — transceivers + optical components

Wave 3 — momentum layer (high-speed photonic / foundry photonic)

Wave 4 — CPO endgame (still binary)

Verdict and reasoning

Demand-side narrative: CONFIRMED at four mechanically independent layers.

The system already did this work. The optical-supercycle perspective documents the demand-side mechanics; the 2026-04-11 trace surfaced the right names; the COHR Q3 May 6 print is the primary-source confirmation; NuttyCLD Part 4 is the independent external corroboration. The CIEN +753% / COHR +450% magnitude lines up with: (i) GPU-cluster scale-out physics making optical mandatory at GB200, (ii) 800G volume ramp through 2H 2025 → 1H 2026, (iii) DCI/WAN structural step-change driven by multi-site AI-clusters, (iv) CPO architectural convergence on the future-state power-budget thesis.

Wave 1 (LITE/COHR/CIEN/FN): positioning CONTESTED, fundamentals CONFIRMED. R20 wrote the full contested entry (Aschenbrenner Q4 2025 exit + Druck initiation, opposite-direction smart money, textbook positioning bifurcation). R31-A propagated the cohort-state language to the README. The fundamentals are the fundamentals; the trade is no longer a fresh discovery.

Wave 3 (TSEM/ALAB/CRDO): MOMENTUM. Same physics chain, different tier. Demand-confirmed by the merchant ramp; positioning is the momentum-extension trade, not the structural-fresh trade.

Wave 4 (LWLG/POET/SIVE/AAOI): STILL BINARY. CPO is the architectural direction; merchant winners are not yet determined. The Q4 2026 NVIDIA Spectrum-X/Quantum-X Photonics ramp is the catalyst that picks the winners.

Net. The "WHY did they pop" question is answered, resolving the open backlog item, and continued tracking runs through the active perspective. No new tasks need to be filed.

Coverage gaps / follow-ups

Zero new tasks filed. The active perspective is the tracking layer for everything this verdict surfaces. Specifically:

• COHR Q4 earnings (Aug 2026) — already in the perspective's catalyst calendar.
• AAOI 1.6T volume shipments (Q3 2026) — already in the perspective.
• LWLG SilTerra tapeout (mid-2026) — already in the perspective.
• NVIDIA Spectrum-X Photonics CPO (H2 2026) — already in the perspective; also cross-referenced in ai-power-delivery.
• ONTO cross-thesis exposure — already noted in the 2026-05-25 HBM verdict's "deliberately not filed" list; surfaces in the optical perspective's next log touch.

Already-filed adjacent tasks (not opened by this investigation but worth flagging here so the next reader knows the surface):

• ALMU binary-outcome deep-dive (TASKS-RESEARCH.md:19) — Wave 4 speculative leg.
• KOPN contract-claim verification (TASKS-RESEARCH.md:69) — watching.json promotion gated on independent verification of four contract claims.
• SIVE/SIVEF fundamentals divergence + dual-listing catalyst (TASKS-RESEARCH.md:110) — Wave 4 momentum/fundamentals divergence watch.

What didn't get done (transparency)

• No fresh OHLC fetch. Prices and cohort-state language are pulled from the May 23 scan summaries and the active perspective.
• No COHR Q3 transcript line-extraction. The May 6 print's revenue numbers and the backlog-to-2028 figure are cited from the perspective README; primary-source transcript extraction is not in scope for a verdict-consolidation artifact.
• No NVIDIA Q1 FY27 print (2026-05-28) preview. The forthcoming print is in the catalyst calendar but post-dates this investigation's data cutoff. The right way to handle the print is an update on the perspective once it lands; not this verdict's job. (Correction 2026-06-11: the 2026-05-28 date was a cadence estimate — NVDA's official IR date was 2026-05-20, so the print did NOT post-date this investigation; it had already landed.)
• No fresh perspective-refresh edits. The perspective is the tracking surface; this verdict points at it. Editorial refreshes are the perspective owner's job.
• No CPO-supplier deep-dive on which specific photonic-foundry / laser-source / TFLN-modulator combination wins the H2 2026 NVIDIA ramp. That's Wave 4 binary-resolution work; tracked in the perspective's existing thread.

Sources

Price, RSI and trend figures read from the desk's pre-computed scan summaries (summaries). Source backlog item:

• thesis-ideas:54` — Priority 1 Semi-Blind #5 (Ciena/Coherent Optical — WHY Did They Pop?). To be updated to RESOLVED with verdict pointer.

Cited and consolidated:

• README — wave-curve rotation model, Q3 print summary, catalyst calendar, cohort state.
• log — R20 contested-downgrade entry (Aschenbrenner exit + Druck initiation).
• 2026-04-11-optical-networking-bottleneck — bottleneck-side trace.
• inputs/articles/2026-04-25-nuttycld-ai-power-crisis-part-4-last-micrometer.md — external power-budget corroboration.
• 2026-05-25-hbm-memory-supercycle-blind-test-verdict — cross-reference for ONTO/COHR cross-thesis exposure.
• optical-supply-chain — 30-ticker watchlist (Wave 1–4 coverage).

Filed follow-ups (this investigation):

• None new. The active perspective is the continued tracking surface; its own task layer in TASKS-PERSPECTIVES.md is the right home for any new work this verdict surfaces. Existing adjacent tasks (ALMU/KOPN/SIVE) remain untouched.

Commands run for this artifact:

the desk's own tooling
  --topic "Optical supercycle WHY — what specifically drove CIEN/COHR demand acceleration in late 2025 / early 2026" \
  --slug "optical-supercycle-why-did-they-pop-demand-side-verdict" \
  --tickers "CIEN,COHR,LITE,FN,AAOI,CRDO,ALAB,TSEM,LWLG,ONTO,NVDA" \
  --perspectives "optical-supercycle,ai-power-bottleneck,ai-power-delivery" \
  --watchlists "optical-supply-chain,supply-chain-traces" \
  --cites "2026-05-25-hbm-memory-supercycle-blind-test-verdict"