Optical links moving from the rack onto the GPU and HBM package
Marvell is the cleanest public exposure to optical compute-to-memory fabrics
Custom silicon, XPU attach, CXL memory pooling, scale-up and scale-out optics, Celestial AI Photonic Fabric and XConn are converging on the same problem of moving data between accelerators and large shared memory. Photonic Fabric is built to connect compute to pooled memory optically, which maps onto GPU and XPU links into HBM and shared memory pools as the memory wall tightens.
| Instrument | Side | Target | Reason |
|---|---|---|---|
| MRVL | Long | Celestial AI Photonic Fabric is designed to optically connect accelerators to shared memory, which is the architecture implied by co-packaged GPU/HBM and pooled-memory roadmaps. Custom silicon, XPU attach, CXL pooling and scale-up optics sit in the same stack. |
Data movement between compute and memory is the next AI bottleneck, pulling optics onto the package and into shared memory pools
GPUs and HBM keep getting faster, but moving that data around the system is becoming the limiter. Today's optics mostly connect switches, servers and racks. A co-packaged architecture in which GPUs and XPUs reach HBM or large shared memory pools over optical links would let several accelerators share one huge memory pool at much higher bandwidth and lower power, and would widen the silicon-photonics and laser market beyond conventional datacom.
| Instrument | Side | Target | Reason |
|---|---|---|---|
| COHR | Long | If GPUs and HBM start talking over laser-based silicon photonics instead of copper, Coherent's lasers and photonic components sit in the volume path of that architecture. | |
| LITE | Long | Lumentum is a core laser and photonic-component supplier for datacom and silicon photonics. Optics moving onto the GPU/HBM package would pull more of that content into AI systems. | |
| AAOI | Long | Applied Optoelectronics is a merchant optical-module vendor that benefits if AI clusters need more laser-based links as data movement, not just raw GPU and HBM speed, becomes the limiter. | |
| SMTC | Long | Semtech's analog and signal-integrity content sits in high-speed optical and interconnect paths, so a broader AI optics build-out around scale-up and scale-out is a demand tailwind. | |
| TSM | Long | TSMC sits under silicon photonics, interposers and advanced packaging. If optics move onto the GPU/HBM package, more of that work lands on TSMC process and CoWoS-class assembly. | |
| GFS | Long | GlobalFoundries is a volume foundry for silicon photonics. CPO and optical I/O closer to compute increase the need for photonic chips it already manufactures. |
Fabrinet's Amazon and merchant ramp can support a much larger company despite NVIDIA datacom softness
Near-term NVIDIA datacom weakness looks like a gap in a longer demand pipeline. Fabrinet is ramping Amazon and merchant transceiver customers, the September quarter is running roughly 100 million dollars above prior estimates, the next factory expansion points toward about 9.8 billion dollars of annual revenue capacity, and there is a path toward about 14 billion dollars over the next three to four years.
| Instrument | Side | Target | Reason |
|---|---|---|---|
| FN | Long | 750 | New Amazon and merchant transceiver work, a September quarter well above prior estimates, and factory expansion toward about 9.8 billion dollars of annual revenue capacity with a path toward about 14 billion dollars over three to four years support a much larger manufacturing franchise even if NVIDIA datacom is soft near term. |
Themes
The content on this page is for informational purposes only and does not constitute financial advice. Stoquate is not a licensed financial advisor. Always conduct your own research and consult a qualified professional before making any investment decisions.