$WKEY/USD H2 #Indicator #Signal #Trading #GlobalOneAI

Who lost money in crypto this week?

$SPCX $RKLB $BKSY $RDW $PL | Los Angeles startup Orbital secured a $5M pre-seed round led by a16z speedrun to build autonomous AI data centers in space.

The capital funds a 2027 tech demonstration flight to test server cooling in orbit.

🚨 A GPU is only as fast as the slowest cable it's plugged into.

That's the wall the entire AI buildout is about to hit → bandwidth. Copper runs out of road as clusters scale past 100K GPUs, and light is the only way through.

So I mapped every route around it, and who gets paid on each. 🧵

First, why there's a wall at all.

A modern AI cluster isn't one chip → it's tens of thousands of accelerators that have to talk at the speed of the slowest hop. Scale the GPUs and the network has to scale with them, or the expensive silicon sits idle.

Copper can't reach. Light can. That's the whole game.

BAND 1 → Scale the Pluggable. The roadmap everyone's already on.

800G today → 1.6T ramping now → 3.2T next. Each step doubles the lanes and adds optics per port.

Who gets paid: transceiver makers ($AAOI $COHR $LITE ) and the DSP/retimer layer routing it ($CRDO $MRVL ). The base case, already re-rating. 👀

BAND 2 → Strip the Power. Because energy per bit is the real ceiling.

→ LPO: drop the DSP, cut power and latency ($AAOI ships 1.6T LPO)
→ CPO: move optics onto the switch package. This is where NVIDIA put $4B → into $COHR and $LITE , not the transceiver names.
→ External laser sources: pull the laser out as its own module.

BAND 3 → Light Into the Chip. The long game, mostly still private.

→ Optical I/O: light to the edge of the die (Ayar Labs, Lightmatter)
→ Optical circuit switching: replace electrical switches entirely
→ Photonic compute: do the math in light itself (2030+)

Huge if it lands. Years out. Watch, don't chase.

BAND 4 → here's the part the crowd skips, and it's the whole thesis.

Every single route above needs two things it cannot route around:

→ InP lasers. Silicon can't lase. Every optical engine on earth needs a III-V laser.
→ Glass substrates. The packaging foundation for CPO and optical I/O.

The bottleneck inside the bottleneck.

Stack it up and the pattern is obvious.

The crowd buys the top → the GPU, the logo, the headline. Crowded, re-rated, lower margin.

The scarce inputs sit at the base → InP lasers and glass. Hard to copy, capacity-constrained, and where the margin quietly pools.

Own the layers, not the logo.

The other side, because there always is one:

→ Timelines slip. CPO and optical I/O have moved before.
→ It rides on hyperscaler capex staying vertical. If that cools, optics cools.
→ The base-layer names are smaller and more volatile.

But the wall is real, and light is the only way through.

Long $AAOI $SIVEF $CRDO $LPKFF 🔵

First it was GPUs. Then memory. Now it’s PHOTONICS → and demand exceeds supply until mid-2027. (sound familiar?👀)

Every AI rack is a chain: compute → memory → the optics that move the data between them. As copper hits its physical wall at 1.6T, the bottleneck moves to light. And the light layer just went scarce.

The signal isn’t subtle. The 6-inch indium phosphide wafer → the foundation every optical laser is built on → has gone from ~$1,400 to ~$5,000 since China imposed export licensing in early 2025. A ~250% move. Effective global capacity is roughly 600-750K wafers against 2.6-3.0M of demand. A supply gap north of 70%, with 90%+ of production held by three names.

This is no longer a thesis. It’s a Reuters headline: $COHR CEO flew to China with a U.S. delegation to raise the export-license delays in person. When a CEO gets on the plane, the constraint is real.

Here’s the layered map of who supplies the light, why each layer matters, and what three management teams just told us about the next 18 months.

A thread on the optical stack. 🧵

The optical supply chain runs in one direction:

Raw materials → InP crystal → blank substrate → epitaxy → device fab (lasers, modulators) → optical component → module → AI rack.

Squeeze any upstream layer and the whole chain feels it. Right now the squeeze is at the bottom: the substrate and the laser.

The edge isn’t owning the rack. It’s owning the layer the buildout can’t route around. Three companies sit on exactly those layers, and all three are vertically integrated enough to control their own fate.

🔵 $AAOI → “top 3 laser supplier worldwide” by end of 2027 (the CEO’s words) 🔥

The integrated bet. Founder Thompson Lin has run AOI for 29 years, and it’s anchored by its own InP laser fab in Sugar Land, Texas → the exact layer everyone else is scrambling to secure.

Q1 2026, verified: revenue $151.1M, up 51% YoY. Data center revenue up 154% to $81.4M, outgrowing cable for the first time in years. Full-year guide raised to $1.1B+. Capacity, not demand, is the constraint → management says demand exceeds supply through mid-2027.

The strategy is the tell. AAOI makes the high-power CPO laser (300-400mW, ~5x an 800G laser) in-house and won’t sell it to rivals. It hoards the scarce part and captures the full module margin. Amazon already took a customer warrant.

Long $AAOI 😎

What “top 3 by 2027” actually requires:

$AAOI is moving 4-inch → 6-inch wafers by end of 2027 (a >4x chip-yield jump), and models ~$471M/month of data center transceiver revenue by mid-2027 → split roughly $90M 100/400G, $217M 800G, $164M 1.6T, with ~40% U.S.-made.

The counter-thesis, because it’s real: $AAOI is not yet profitable. Non-GAAP gross margin is ~29% against a 40% target that doesn’t arrive until 2027. The stock has already run ~800% in a year and missed Street consensus last print. You’re underwriting 2027 execution, not buying it today.

🔵 $CRDO → the layer the rack literally cannot route around

If $AAOI is the laser, $CRDO is the electrical nervous system: SerDes, retimers, active electrical cables → the interconnect that keeps signals alive between GPUs before, and alongside, the optics. It just closed DustPhotonics, adding silicon photonics in-house. Now it spans electrical AND optical, 800G through 3.2T, end to end.

And unlike most of the stack, it’s already profitable. FY26 revenue tripled to $1.34B at 68% gross margin, with $1.4B in cash.

This is the both-roads play. Copper now, light next → $CRDO sells the connection either way. The toll booth doesn’t care which road the traffic takes.

Long $CRDO

🟢 $SIVEF (Stockholm: $SIVE) → the light source CPO cannot engineer around 💡

The most upstream of the three, and the purest expression of the bottleneck. Co-packaged optics doesn’t emit its own light. Every optical engine needs an external continuous-wave InP laser feeding it. That’s the layer Sivers sits in.

Its external-light-source modules with $POET are targeting production readiness by end of 2026, into a market analysts size above $1B annually. This is the earliest-stage, highest-conviction, and highest-risk of the three → a small-cap laser-source bet on the exact layer the whole CPO transition depends on.

Long $SIVE $SIVEF

Three layers, one chain, same constraint:

🟢 $SIVEF → the light source (earliest, binary)
🔵 $AAOI → the integrated laser + module (scale, founder-led)
🔵 $CRDO → the interconnect toll (profitable, both roads)

The incumbents you already know → $LITE, $COHR, $GLW → are forward-sold into 2028 and flying CEOs to Beijing. That’s not a story about a good quarter. It’s a structural undersupply that prices through every layer above the wafer.

Own the layers, not the logo.

The question this map answers: which layer re-rates first?

The substrate names are known. The module names have run. But the light source $SIVEF and the integrated laser maker $AAOI are still priced like accessories to a transition that doesn’t work without them. 👀