Rosanna Prestia, MBA@RosannaInvests
🚨Everyone is still buying the chips. The bottleneck already moved.
A GPU that computes in nanoseconds and waits microseconds for data is a stranded asset. At 1.6T speeds, copper runs out of physics. The constraint on AI is no longer how fast you can think. It's how fast you can move what you thought.
Jensen has now said it twice in three months.
At GTC in March: "Is copper going to still be important? The answer is yes... Are you going to scale up optical? Yes. Are you going to scale out optical? Yes... We need a lot more capacity for copper. We need a lot more capacity for optics. We need a lot more capacity for CPO."
Last week at Computex, on Marvell's stage: "Optics where you must, copper where you can." Then he called Marvell the next trillion-dollar company and the optical complex repriced within days. The same keynote put a date on the handoff: 200G per lane is the last generation where copper is sufficient. After that, optics takes the rack.
Translation: not copper OR light. Copper now, light next, unprecedented amounts of both. 🔥
The chain is unavoidable: AI tokens are profitable → more GPUs → more bandwidth → copper hits its wall → photonics becomes the chokepoint.
And the smart money stopped debating. Follow the closed deals:
→ $NVDA has committed at least $6.5B to photonics in three months: $2B into Lumentum, $2B into Coherent, a $500M stake in Corning, and a piece of Ayar Labs' $500M round. Direct investments to secure its own light supply.
→ $MRVL paid $3.25B for Celestial AI, up to $5.5B with milestones, to build what its CEO calls a silicon photonics powerhouse.
→ $CRDO closed DustPhotonics two weeks ago. Ciena bought CPO startup Nubis for $270M.
North of $10B of strategic capital locked up one supply chain in under a year. Capital like that doesn't chase a theme. It secures a bottleneck.
LAYER 1: WAFER. Every laser starts as a crystal.
🟠 $AXTI: the InP substrate leader. The first chokepoint in the stack.
🟡 $IQE: compound-semi epiwafers feeding the laser makers. Speculative, but structurally upstream.
LAYER 2: LIGHT. Photons don't make themselves.
🟠 $LITE: revenue +90% YoY last quarter to $808M. EML shipments doubled and management says demand still exceeds supply across EMLs, pump lasers, and transceivers. NVIDIA just wired them $2B. OCS backlog past $400M plus a multi-hundred-million CPO order for 2027.
🟢 $SIVEF (Stockholm: SIVE): the external light source. CPO does not emit its own light. Every optical engine needs a continuous-wave InP laser feeding it, and that is the layer you cannot engineer around. ELS modules with POET hit production readiness end of this year. Disclosure: long.
🟣 $POET: the optical engine wildcard. Its Optical Interposer pairs with Sivers' lasers on external light sources for CPO, with a LITEON module deal stacked on top. Binary commercialization, real architecture.
LAYER 3: OPTICS AND MODULES. Where light meets the rack.
🟠 $COHR: the volume anchor in transceivers, holding NVIDIA's other $2B check.
🔵 $AAOI: Q1 revenue +51% to a record $151M, datacenter revenue more than doubled, $124M of 800G orders plus a $200M+ 1.6T order in hand. Scaling Texas capacity toward 500K+ units a month by year-end, targeting $1B+ revenue this year. Domestic supply while everyone fights over offshore. Disclosure: long.
🟠 $FN: the foundry of optics. When Fabrinet is building, the orders already exist.
THE INTERCONNECT: the layer the rack cannot route around.
🔵 $CRDO: just closed DustPhotonics. SerDes → DSP → silicon photonics → system integration, one company, 800G through 3.2T. Electrical AND optical, end to end. FY26 revenue tripled to $1.34B at 68% gross margin. The toll booth on both roads. Disclosure: long.
🟠 $MRVL: $3.25B for Celestial AI, and Jensen's trillion-dollar nod on the Computex stage.
🟠 $AVGO: switch silicon, optical DSPs, CPO engines. They define the socket.
🟠 $ANET: the AI spine. 100K-GPU clusters get stitched together in light.
LAYER 4: PACKAGING, FIBER, FOUNDRY. Where photons get industrialized.
🔵 $TSEM: the neutral silicon photonics foundry. Prints wafers for whoever wins.
🟣 $LPKF: glass-substrate packaging for glass-based CPO. Real technology, binary commercialization.
🟠 $GLW: AI racks demand several times the fiber density of legacy cloud, and NVIDIA just took a $500M stake. Corning sells density.
LAYER 5: TEST AND THE ANALOG UNDERLAYER. Complexity is a tax paid in validation.
🔵 $AEHR: silicon photonics test, ramping with the cycle. '
🔵 $VIAV: every 800G and 1.6T transceiver gets validated before it ships. The gate the market prices like an accessory.
🔵 $SMTC: the drivers and TIAs that fire the lasers. Sits directly under the LPO trade.
🔵 $MTSI: the high-speed analog behind 1.6T engines.
🟠 $CIEN: transport. Even long-haul is buying light.
💡The counter-thesis, because every map needs one. The honest debate on this stack is whether these are genuine bottleneck assets or cyclical optics suppliers enjoying peak demand at peak multiples. Lumentum's May print showed +90% growth with the stock up roughly 1,400% over the prior year at a triple-digit trailing multiple. That is a price for perfection. Most of these names live or die on a handful of hyperscaler capex lines, and one digestion quarter hits the whole stack at once. CPO timing has already slipped once. Architecture risk is real: LPO, CPO, and stretched copper are still fighting for the same sockets. The cycle is real. So is the gravity. 🔥
But the bears have to explain one thing: $NVDA, $MRVL, $CRDO, and $CIEN just spent over $10B securing this supply chain with their own balance sheets. The people with the best information are paying up for the layers.
The market owns the top of this stack. The asymmetry is at the edges: wafer, light, packaging, test.
Own the layers, not the logo.
Bookmark this for the weekend. Then tag the one investor you know who's still all compute and no interconnect. 👀
