Here’s the thing about the AI boom: it’s not one race. It’s a relay race where each leg creates a new set of winners.
First, everyone needed GPUs—Nvidia won that round so hard it became a verb. Then hyperscalers realized they needed servers to hold all those GPUs. Then they hit a thermal wall and suddenly liquid cooling companies were printing money. Then the power grid started sweating. Then memory became the constraint.
Now? The bottleneck is hiding inside the data center itself: the cables and chips that move data between GPUs.
Think about it. You’ve got hundreds of thousands of processors in a cluster, each one costing tens of thousands of dollars. If the data can’t move fast enough between them, these expensive chips just sit there idle. That’s like owning a Ferrari but only driving it in a school zone.
The Copper vs. Optics Showdown
Here’s where it gets interesting. There are two ways to move data: copper cables and fiber optics.
Copper’s the incumbent. It’s cheap, fast, and power-efficient for short distances. But there’s a physics problem: at today’s speeds (800G), copper cables only work reliably for about three meters. That’s basically one rack. As speeds push toward 1.6T, copper’s range shrinks even more.
Optical fiber doesn’t care about distance. Light travels light-years (literally). But it’s expensive, power-hungry, and adds latency because you have to convert electrical signals to light and back again.
Enter the middle ground: active electrical cables. These are copper cables with embedded chips that extend the range to 7-10 meters while using way less power than optics. It’s copper’s last stand before physics wins.
Broadcom’s CEO recently made it clear: copper dominates for connecting GPUs within a cluster (scale-up), while optics handles longer distances between clusters (scale-out). But here’s the kicker—Broadcom’s entire AI architecture is optimized for copper. So take that opinion with a grain of salt.
The Real Play: Two Phases
This isn’t a one-time trade. It’s a sequence.
Phase 1 (Now through 2027): Copper wins. Companies like Credo Technology, Marvell, and Amphenol are printing money right now. Credo’s basically the pure-play copper interconnect stock—explosive growth, high risk, high reward. Marvell’s the diversified hedge that plays both copper and optics. Broadcom’s the Ethernet switching giant. Amphenol’s the boring, reliable connector play.
Phase 2 (2027-2030+): Optics takes over. Co-packaged optics (CPO) technology—which puts photonics directly on the chip—eliminates optics’ power and latency penalties. Nvidia just dropped $4 billion on Lumentum and Coherent to lock down supply. That’s not a casual investment. That’s a supply chain takeover. Fabrinet’s the contract manufacturer that assembles these things. Applied Optoelectronics is the speculative small-cap with maximum leverage.
The Bottom Line
The AI infrastructure boom isn’t slowing down. It’s just moving to the next constraint. The hyperscalers have identified the problem. They’re throwing billions at it. The supply chain is tight. The winners are already visible.
Get positioned before the market figures it out. Because the biggest gains always go to the people who see the next phase coming before everyone else does.