Investment thesis · Live

The AI bottleneck moves from chips to the network: all that 1.6 Tb/s of bandwidth hangs on a handful of InP laser makers

The AI boom is shifting the bottleneck from silicon to the network: 1.6 Tb/s modules rest on indium phosphide EML lasers, made by a handful of companies.

Published August 7, 2026 · 180-day horizon · technology

Basket return

-9.9%

equal weight, since publication

The market over the same window

-1.1%

benchmark for this basket

Edge over the market

-8.8%

in percentage points

Causal chain: Step-change growth in AI clusters → Network grows faster than GPU count → Shift to 1.6 Tb/s modules (200G per lane) → Bottleneck: EML lasers from InP epitaxy → Coherent → Lumentum → Fabrinet

What this thesis rests on

Each one is a statement that has to be true. When a filing says otherwise, the thesis is in trouble, and this is where we say so.

Coherent's datacom transceiver revenue grows at least 20% year over year in each of the two quarterly reports filed after 7 August 2026.
financial · Challenged
Lumentum's reported gross margin expands by at least 3 percentage points versus its June 2026 quarter in a 10-Q filed before 3 February 2027.
financial · Unverified
No Chinese or Japanese supplier brings an InP epitaxy line into volume production for 200G-per-lane EML chips before 3 February 2027.
competitive · Unverified
Nvidia's AI rack platforms shipping through 3 February 2027 keep using pluggable optical modules for the majority of GPU links instead of co-packaged optics.
operational · Unverified
Fabrinet's optical communications revenue grows at least 15% year over year in every quarterly report filed through 3 February 2027.
operational · Challenged

Each new generation of AI training clusters increases not so much the number of processors as the density of the links between them. Scale-up and scale-out topologies mean that a single GPU today comes with a dozen-plus optical links, and eventually several dozen. It is the network, not the silicon itself, that is becoming the fastest-growing line item in a data center budget, and it is the network that sets the pace of the build-out.

To keep up, the market is moving from 400G modules through 800G to 1.6 Tb/s. What matters is not the headline speed but the jump from 100 to 200 gigabits per optical lane. At that rate, cheap directly modulated DFB lasers no longer suffice and you have to reach for externally modulated lasers, i.e. EMLs. Those in turn are grown by epitaxy on indium phosphide (InP) substrates, a process that is technologically far harder and far more capacity-constrained than making ordinary silicon chips.

This thesis, over time

What has happened since publication. Every entry is computed as of its own date, from price history.

August 19, 2026

Basket fell below -10% since publication

This basket is at -10.5% since publication, 12 days after the thesis was posted. The benchmark (S&P 500) returned -0.5% over the same window, so the basket trails its market by 10.0 percentage points. The strongest contributor is LITE (-0.8%), the weakest is FN (-16.6%). This is not a change of status: the horizon keeps running and the verdict only lands once it expires.

August 18, 2026

Basket crossed +10% since publication

This basket is at +10.3% since publication, 11 days after the thesis was posted. The benchmark (S&P 500) returned -0.2% over the same window, so the basket is ahead of its market by 10.4 percentage points. The strongest contributor is LITE (+15.6%), the weakest is COHR (+5.1%). This is not a change of status: the horizon keeps running and the verdict only lands once it expires.

Open the interactive thesis

Interactive causal map, the basket charted against its market, and the full timeline.

Open in Invesaro →

This is analysis, not investment advice and not a recommendation to buy or sell anything. We publish it and track it in public, mistakes included. Any decision is yours and yours alone.