Optical Modules — Strategy

Where Optical Modules is heading over the next 12 months, grounded in product-axis evidence and verbatim demand from the last 90 days. The judgment column is the engine's read — operators verify and refine.

Product trajectories

800G Optical Transceiver Module (SFP/OSFP/QSFP-DD) — transceiver ↗ rising

Opportunity: NVIDIA's H100/H200 GPU fabric is consuming 800G SFP/OSFP optics at huge volume; the duopoly of Innolight + Eoptolink and Lumentum's $750M Cloud Light bet validate that this is the current cash-printing SKU of the AI buildout.

800G pluggable transceiver is the current product — concrete A5 customer-win evidence shows two Chinese vendors lock ~60% of NVIDIA's 800G supply with Lumentum chasing via Cloud Light. Risk is generational: the same evidence pool already signals 1.6T and CPO are next, so the moat is short.

Co-Packaged Optics (CPO) Switch — switch ↗ rising

Opportunity: AI bandwidth wall: pluggable optics power/density can't keep up with GPU scale-out, so hyperscalers want optics co-packaged with the switch ASIC (fact 36cca98b, fact 0ebc4a28 Broadcom Tomahawk 6 Davisson 102.4T CPO shipping).

CPO switches are the single biggest '' signal in this topic — Broadcom is shipping Tomahawk 6 Davisson at 102.4T with CPO, NVIDIA has publicly endorsed CPO for GPU fabrics, and Astera Labs is actively hiring optical validation + advanced packaging talent flagged with 'cpo'. What's missing: independent A4/A5 customer-win evidence outside Broadcom's announcement; the 'third time' framing implies execution risk.

Silicon Photonics Engine (data center) — silicon photonics PIC ↗ rising

Opportunity: Silicon photonics is being positioned as the enabling layer for both CPO and high-density pluggables; the tension between TSMC's empire-building and Jensen's 'stay on copper' quote is exactly the kind of contested narrative that signals an inflection.

Silicon photonics is the contested foundation layer — Astera is hiring optical validation with explicit silicon_photonics specialty and TSMC is building capacity, but the lead AI customer (NVIDIA CEO) is publicly hedging the timeline. Position: with binary timing risk.

1.6T Optical Transceiver Module — transceiver → steady

Opportunity: 1.6T is the explicit next-generation upgrade target on every cluster roadmap; InP supply (DFB/EML lasers) is being called out as the upstream chokepoint, framing it as both a product opportunity and a supply-chain bottleneck investors care about.

1.6T is a clear — strong A2 demand chatter and explicit chokepoint framing (InP / 200G/lane SerDes) make it the obvious successor to 800G. Missing direct A5 customer-win evidence: nobody in the dataset is yet declared the volume-supply winner for 1.6T to NVIDIA.

1 THz Optical Modulator — modulator IP/component → steady

Opportunity: A 1 THz modulator (vs ~100 GHz state-of-art) is the kind of physical-layer leap that could collapse the lane-count required for 1.6T/3.2T optics — Marvell IP ownership would re-rate the entire DSP+optics stack.

Speculative — a single retail-investor-tone A2 query is the only mention, and Marvell's $1B raise is for debt repayment, not announced as modulator capex. Worth flagging because if true the read-through to every transceiver supplier is huge, but evidence is weak.

Linear Pluggable Optics (LPO / Linear Drive) — transceiver · weak signal

Opportunity: LPO removes the DSP retimer from the optical module to cut power and cost; A2 chatter frames it as a parallel-bet to CPO that hyperscalers are actively evaluating as the next bottleneck.

LPO is a credible — it is named explicitly in A2 demand chatter as the next bottleneck alongside CPO, but evidence is thin: no committed supplier in A4/A5 in this dataset. Watch as a hedge to CPO timeline slip.

What the market is asking (last 90d)

• Why the S24 Ultra is still the absolute peak Galaxy Ultra (and why I’m skipping the S25 and S26)
• Why nobody is talking about this? Marvell now owns a 1 THz optical modulator, 10x faster than anything shipping today!
• Building infrastructure for massive #AI workloads? Traditional setups simply can't keep up. Learn how optimizing your optical architecture eliminates bottlenecks, supports next-gen 800G/1.6T clusters, and future-proofs your #datacenter.
• Silicon Photonics In The Data Center: What A CMOS Exec
• In the AI Gold Rush, What Are the Shovels?
• Would you use a system that turns your raw input into structured team training?

See the Products and Hiring modules for the full landscape and who's investing in which direction.