Optical attach is set by topology, not by shipments — linear optical TAM models are mis-specified
optical-attach-is-topological · conviction medium · status open · horizon 2026-2028 · as of 2026-08-03
Sell-side optical TAM scales optical content with accelerator and rack shipments. The published traffic decomposition says otherwise: ~97% of LLM training traffic is tensor- and sequence-parallel and therefore LOCAL, so a mesh topology reserves optics for pod-to-pod and costs under a third of flat/Clos for the same silicon. Optical content per unit of compute is a function of the topology an operator picks and the training/inference mix it runs — not of how many chips ship.
Robust to undisclosed shares. 1 derived input under this thesis; redrawing every supply weight the industry does not publish moves none of them by more than 25%. Computed from evidence at most 17 days old (oldest input: broadcom).
Exhibits
Exhibit 1Relative performance, indexed to 100How the names in this thesis have traded against SOXX.
Series available as data/optical-attach-is-topological.csv
Exhibit 2What the conviction is actually made ofEach premise and the number it composes to. A conjunction of plausible premises is far weaker than any of them.
Optical content is mechanically downstream of topology, and the desk holds both paths
90% if the 3 gates are independent, 95% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.
Weakest link: Optical Transceivers (800G/1.6T pluggables) at 0.95 — Pluggable optics; the scale-out path.
The traffic structure supports mesh for training, which is the low-optics case
69% if the 3 gates are independent, 85% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.
Weakest link: Interconnect bandwidth (scale-up and scale-out) at 0.85 — The resource both paths deliver, and the node the workload layer consumes.
Therefore linear optical TAM models are mis-specified in BOTH directions, and the tradable question is mix
38% if the 3 gates are independent, 60% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.
Weakest link: Linear-drive optics (LPO / LRO) at 0.60 — Contested architecture that changes content per port without changing port count. Not settled in the merchant market.
The variant
Consensus
AI optical demand is a monotonic mega-trend. Bandwidth per rack rises 16-45x across the roadmap, copper hits a reach ceiling, pluggables give way to co-packaged optics, and a ~$154bn rack-level optical TAM follows from multiplying content per rack by rack shipments. Optical attach per accelerator is treated as a roadmap constant.
Variant
Attach rate is a DESIGN CHOICE, and the same sell-side data shows it spanning 6-9x. Huawei's CloudMatrix 384 runs 1:18 optical modules per chip because it connects everything optically; NVIDIA's GB300 runs 1:2-3 because NVLink scale-up rides a copper backplane and optics only serve scale-out. The UB-Mesh traffic decomposition explains why the spread exists: tensor-parallel AllReduce ~52.9% and sequence-parallel AllGather ~44.1% of training traffic are highly local, so a mesh can serve them on direct-attach copper and reserve optics for the pod boundary — stated at under one third of flat/Clos networking cost. A training-heavy mesh buildout and an inference-heavy Clos buildout consume very different optical content at identical silicon volume.
Differentiator
Both sides of this argument use the same numbers and neither tracks the variable that reconciles them. The sell-side TAM already ENCODES topology dependence in its own attach table (1:18 vs 1:2-3) and then models the aggregate as if attach were fixed. The desk's edge is holding the reconciliation variable explicitly — training/inference mix and disclosed topology per platform — as the leading indicator for optical attach, and refusing to scale optical demand linearly with accelerator units in either direction.
Falsifiers
claim: Deployed capacity is overwhelmingly Clos, making topology variance immaterial · criterion: Disclosed platform topologies through 2027 show >85% of deployed AI capacity on flat/Clos fabrics, i.e. the mesh alternative never takes share · horizon: 2027-12-31 · settles: refuted
claim: Attach rate is in fact stable across architectures · criterion: Optical modules per accelerator, measured across at least three distinct platforms (NVIDIA rack-scale, a merchant Ethernet build, and a mesh build), fall within a 2x range rather than the 6-9x the current data shows · horizon: 2028-06-30 · settles: refuted
claim: The reconciliation variable moves the tape · criterion: A disclosed shift in training/inference mix or platform topology precedes a >20% revision in optical-module demand guidance at Coherent, Lumentum or Innolight · horizon: 2028-06-30 · settles: confirmed
claim: SCOPE — right mechanism, no consequence · criterion: Optical names track accelerator units within benchmark dispersion through the horizon while topology mix demonstrably shifts. The variance was real and unpriced-because-irrelevant · horizon: 2028-12-31 · settles: refuted
Open questions
The desk does NOT track the training/inference split of deployed capacity, which is the reconciliation variable this whole thesis rests on. Neither does either side of the public argument. Building that series is the prerequisite for scoring this.
The ~97% locality figure comes from one published traffic decomposition carried by a partisan source. The decomposition itself is checkable and non-proprietary, which is why it is used — but it has not been independently reproduced.
Inference topologies with wide expert parallelism favour Clos and therefore HIGH optical attach. If inference share rises faster than training, the variant points the other way. The thesis is about variance, not direction.
Reasoning chain
Optical content is mechanically downstream of topology, and the desk holds both pathsVALID
Transceivers ARE delivered interconnect bandwidth — the edge added 2026-08-02 when the pluggable layer was found to have no path into the workload layer.
Nvlink0.95 strong
The copper scale-up alternative that suppresses optical attach inside the domain.
Bandwidth can be delivered over copper inside a domain or over fibre between them. Which one carries a given link is a topology decision, so optical content per accelerator is determined at design time, not by unit volume.
The traffic structure supports mesh for training, which is the low-optics caseVALID
premises
Scale-up fabric (within the node / rack)0.90 strong
The within-node/rack domain where local collectives live.
Scale-out network (between nodes)0.90 strong
The between-node domain where optics are unavoidable.
Interconnect bandwidth (scale-up and scale-out)0.85 strong
The resource both paths deliver, and the node the workload layer consumes.
If ~97% of training traffic is tensor- and sequence-parallel and local, most links never need to leave the copper domain. Optics attach at the boundary, and the boundary's position is set by the topology.
Therefore linear optical TAM models are mis-specified in BOTH directions, and the tradable question is mixVALID
premises
Co-packaged optics0.75 strong
29%59%
REACTIVE from derived:co-packaged-optics.input_cost_pressure (today 54.4, band 21.77-87.1 -> p 0.6-0.9).
why
Input-cost pressure on CPO is a proxy for how contested the optical BoM is. It is an IMPERFECT proxy — the thesis really turns on training/inference mix, which the desk does not yet track — and it is wired with a deliberately narrow conviction band for that reason. Anchors are centred on today's reading so this makes the premise LIVE without re-rating the thesis. The high-attach path, whose penetration is ASSUMED at 25-29% in the sell-side build and carries ~59% of that TAM.
Linear-drive optics (LPO / LRO)0.60 moderate
Contested architecture that changes content per port without changing port count.
why
Not settled in the merchant market.
Credo Technology0.85 strong
Active copper — the name that profits when the copper/optics boundary moves the other way.
A forecast that scales optics with shipments is right only if mix is constant. It is not: CPO penetration is assumed, LPO is contested, and active copper extends the copper domain. The desk therefore holds attach as a dependent variable and tracks mix, rather than taking a direction on optical volume.
Sources
UB-Mesh traffic decomposition: TP AllReduce ~52.9%, SP AllGather ~44.1%, EP All2All ~1.5%, PP P2P ~0.14%, DP AllReduce ~1.34%; mesh stated at under one third of flat/Clos networking cost — linkacc 2026-08-03