AI Networking X-Ray

Intelligence layer

MODELED · interpretation of ranking · not investment advice

MODERATE CONFIDENCE · 70%

Current reality

The ranking does not say Tomahawk-class switch silicon is what currently stops AI networking scale. TH6 is shipping in production volume. It says 800G / 1.6T optical attach is. Primary Broadcom materials tie AI cluster growth and 400G/lane DSP to module demand, while the only quantified shortfall (McKinsey via MLQ: 800G 40–60% short through 2027; 1.6T 30–40% through 2029) is secondary and partly forward-looking. That is still the strongest scarcity signal in this investigation. Lasers were previously over-ranked on secondary unique-supplier language and InP expansion stories; expansion is not scarcity, and Broadcom itself ships millions of 200G EMLs. Confidence stays moderate because present module lead times and industry WPM remain weakly disclosed.

Constraint migration

The engine relieves transceiver capacity, lead time, and concentration only, then re-ranks. Optical engines rise next because the relief path away from front-panel modules is CPO (Davisson 16×6.4T engines; NVIDIA Spectrum-X Photonics) — and engines are vendor-coupled, yield-sensitive, and not yet commodity capacity. Optical DSP sits close behind as the shared control for both pluggable 1.6T and future CPO paths (Sian3 / Taurus primary). Lasers remain under both stacks but stay below modules: primary shows production and capacity expansion, not a measured present shortfall that outranks the module attach bind.

800G / 1.6T Optical Transceivers (60) → Optical Engines / Silicon Photonics (50) · engine cascade

Who gains leverage

• 800G / 1.6T module makers who can actually deliver qualified volume into TH6-scale fabrics (Innolight-class assemblers and vertically integrated optics houses)
• Optical DSP / PHY suppliers at 200G and 400G per lane (Broadcom Sian3 / Taurus; Marvell-class second sources if they qualify)
• CPO optical-engine attach that can displace front-panel modules once hyperscaler service models clear — Broadcom Davisson and NVIDIA Spectrum-X Photonics paths
• InP / EML capacity only as the underlying laser surface under modules — watched, not treated as the observed #1 without primary shortfall series

Who loses leverage

• The story that AI Ethernet switch silicon is the binding networking constraint while TH6 ships in volume
• Treating Broadcom or Lumentum InP expansion announcements as proof the optics shortage is over
• Elevating lasers to #1 on secondary monopoly claims when Broadcom primary shows millions of EMLs shipping
• Assuming CPO already replaces pluggable module scarcity at production scale — product announcements are not hyperscaler volume

Second-order effect

Clearing front-panel transceiver slots does not create lossless AI fabrics. It moves the bind onto engine attach and the DSP+laser bill of materials those engines still need. CPO reduces module count but concentrates packaging and engine yield risk. Network power remains a parallel facility constraint (Broadcom states optical interconnect power limits cluster scalability) — component relief does not erase rack watts. The coupled trap: the same specialty lasers and PAM4 DSPs sit under both the pluggable path and the CPO relief path.

Opportunity

Worth investigating: who can deliver incremental 800G/1.6T modules into 2027 with hyperscaler quals intact; whether 400G/lane DSP (Taurus) actually unlocks 1.6T volume or only shifts the laser/driver bind; whether Davisson / Spectrum-X CPO engines clear field-service gates before pluggable shortfalls ease; and whether InP/EML lead times, if disclosed in primary filings, invert lasers above modules. Not a conclusion: Broadcom, Coherent, or Lumentum as securities, or that the interconnect shortage is over.

Watch signals — confirm

• Primary module makers or hyperscalers disclose multi-quarter 800G/1.6T allocation or shortfall language
• Secondary shortfall estimates remain directionally consistent while primary lead-time series stay opaque
• CPO remains sampling / limited shipping without large production attach that would demote pluggable scarcity
• DSP roadmaps keep tying new generations explicitly to 800G/1.6T module markets

Watch signals — invalidate

• Primary disclosures show 800G/1.6T modules with short lead times and unused capacity across major suppliers
• CPO ships at production volume that materially reduces front-panel module demand for AI fabrics
• Primary laser / InP filings show measured present shortfall that clearly outranks module attach as the binding clock
• Hyperscalers report idle 100T-class switch silicon waiting on something other than optics

Confidence

Matches the transceiver-node evidence confidence. Direction (modules over engines over DSP/lasers; switch silicon demoted) is better supported than the exact score margin. The quantified shortfall remains secondary and partly forward-looking; mid-2026 Coherent InP-constraint commentary was not verified against a primary IR filing in this pass and is treated as a watch item, not a ranking flip.

Next question

If 800G / 1.6T pluggable capacity were suddenly abundant, what becomes the next binding constraint on AI networking scale — CPO optical-engine attach, optical DSP, laser sources, or network power density?

MODELED interpretation of a CALCULATED ranking grounded in OBSERVED claims. Not a forecast and not investment advice. Areas worth investigating, not securities to buy.