This article pulls together Bernstein’s perspective that connectivity is shaping up to be the next big bottleneck in AI deployment. It digs into why copper interconnects, co-packaged optics, and the entire supply chain matter so much for data-center infrastructure—and, honestly, for investors trying to figure out where the real value lies.
As AI workloads keep demanding faster data speeds and less latency, the hardware that shuttles and processes all that data becomes a real battleground. Bernstein isn’t really picking a single winner here. Instead, the focus is on mapping where value gets created across the ecosystem—from raw materials and substrates to test equipment and how everything gets integrated.
Connectivity as the next AI bottleneck and battleground
From intra-rack copper backplanes to multi-node optical integrations, today’s AI workloads are pushing existing infrastructure to its limits. Bernstein points out that the copper-versus-co-packaged optics debate misses the bigger picture.
Investors, he suggests, should look at the whole timeline of connectivity upgrades, not just bet on a single tech outcome. That’s a pretty broad way to look at things, but it makes sense given how fast everything’s moving.
Copper-based solutions are expected to stick around as the main choice for intra-rack, scale-up deployments through at least 2026–28. Co-packaged copper (CPC) might even stretch copper’s usefulness by letting suppliers like Luxshare delay those expensive system-level refreshes.
The debate around co-packaged optics (CPO) is all about tighter optical integration with accelerators. But Bernstein sees some big hurdles: manufacturing yield, testing complexity, fiber coupling, and doubts about serviceability and vendor concentration for cloud providers.
Copper interconnects vs. co-packaged optics: where value lies
On the copper side, upgrades look like they’ll happen gradually instead of all at once. Copper interconnects should keep powering intra-rack, scale-up deployments for the next few years.
CPC could help copper hang on longer by enabling denser, faster links—no need for a total hardware overhaul. This is good news for a lot of suppliers in the copper and related components space.
Co-packaged optics, on the other hand, promise much tighter integration between optics and accelerators. That could mean real performance gains. Still, the technology faces some stubborn obstacles: getting high manufacturing yields at scale, making testing simpler and more reliable, keeping fiber coupling robust, and tackling cloud-provider worries about serviceability and vendor lock-in.
Bernstein stresses that while CPO paints a compelling long-term picture, its near-term hurdles—like yields and ecosystem readiness—remain pretty big question marks for both investors and operators.
Market signals are showing just how complicated this transition is. In 2026, shares of optical and fiber companies have jumped—Lumentum, Coherent, and Corning are all riding the AI connectivity wave.
Lumentum especially has seen big year-to-date gains and more attention in the indexes, which shows just how excited investors are about optics tied to AI acceleration at scale.
Supply‑chain beneficiaries and notable partnerships
Bernstein singles out several downstream winners across the connectivity value chain. These companies are in a good spot to capture more value as AI data centers keep evolving:
- Fiber array units: TFC and Senko are key for building high-density fiber arrays, which are critical for both CPO and advanced copper interconnects.
- Packaging and test equipment: Chroma ATE, All Ring, and Teradyne are leading in packaging and testing, making sure devices are reliable and “known good die” get assembled right.
- ABF substrates and related films: UniMicron and Ajinomoto stand out in substrate and film supply, which are foundational for the growth of large-scale AI servers.
Strategic partnerships are also shaping the field. TFC has a multi-year collaboration with Nvidia on CPO development, which really ties silicon leadership to optical packaging.
Senko has worked with Nvidia and Marvell, and Teradyne supplies test equipment to verify “known good die” in CPO setups. The ABF substrate market has taken off since 2018, thanks to bigger server fleets and more complex chip designs. UniMicron could be heavily dependent on AI by 2026 if this keeps up.
What this means for investors and AI infrastructure planning
Bernstein’s core message feels pretty practical: you don’t have to pick a single technology winner just to get in on the connectivity megatrend. Investors can spread out across the value chain—fiber, packaging, test, substrates, and all those related materials—so they can catch growth as AI data centers roll out faster interconnects in waves.
If you’re leading a data center, the real takeaway is to plan for a blended path. Lean on copper where it makes sense right now, but keep an eye on CPC advances as manufacturing and the whole ecosystem start to mature and improve.
Here is the source article for this story: Here’s the next AI ‘battleground’ — and how investors can get in on the action