Increasing processing ability will prepare data centers for explosive growth in 2026
Brian Rhoney
Published: January 29, 2026
If 2025 was the year hyperscalers started to scale out to support growing AI demand, 2026 will be the year they take it to the next level by scaling up and across and up into new territory. The approach of scaling out — adding more nodes to a distributed network — will evolve into a scale-across approach, connecting multiple data centers into a powerful single unit. Additionally, we’ll start to see more operators scale up inside the centers themselves, adding more GPUs to a single existing node.
We’ll see that shift show up in three distinct aspects: data center interconnect, network and switch-to-switch, and in the server racks themselves. Each one of these areas is facing infrastructure challenges driven by density and deployment thresholds as compute demands from AI and other applications increase. As we progress into 2026, here’s what I’m keeping an eye on.
Data center interconnect: The metro era evolves
Long-haul data center interconnect (DCI) remains the backbone that ties states and regions together, requiring high-performance technology to maintain signal integrity across vast distances. Looking ahead, we’ll continue to see it growing steadily with deployments at a higher rate.
But this year, DCI on a smaller, metro scale will be an important growth area as hyperscalers prepare for gigawatt-scale AI campuses. This is the idea of “scaling across,” linking multi-hundred-thousand GPU clusters across multiple buildings and campuses within a single region, then treating the connected facilities like one AI factory. The shift is already driving a surge in outside-plant cabling requirements.
The biggest differentiator? Density. Long-haul cables have always been dense, but advanced networks will require cables to take it to the next level. Instead of sub-1,000 fiber runs, we’ll see thousands of fibers per run and multi-conduit deployments that stack cable counts into the hundreds of thousands. Attenuation will become less important because distances are shorter in these massive complexes, making density the top priority.
Expect to see more innovation aimed at squeezing even more fibers per duct, newer cable designs that push higher counts, and bend sensitivity tightening. For many in the industry, this will serve as the first step toward multicore fiber becoming a mainstream option.
Network and switch-to-switch: The year CPO and lens connectivity leaves the lab
Two big shifts will define switch-to-switch infrastructure in 2026. The first is co-packaged optics (CPO), which will finally hit actual AI cluster deployments after years of demos and pilot concepts. CPO packages optical transceivers alongside integrated chips on the same silicon substrate. As a result, there’s a shorter electrical path, which unlocks unprecedented bandwidth capacity and reduces power consumption.
Currently, data centers use pluggable optical transceivers, which will likely remain in place initially to offer operators a backup plan as operators test drive CPO. No one falls behind by not deploying CPO in 2026, but early adopters will carry a strategic advantage into the next cycle.
Additionally, a new technology, lens connectivity, is expected to come into play — and may help mitigate one of the industry’s biggest obstacles. Over the past few years, hyperscalers have needed to hire technicians at such a fast pace that training has become challenging. Physical-contact connectors require meticulous handling, which doesn’t mix well with compressed timelines and rotating crews. Even when only a small percentage of links chasing those failures can double the total commissioning time. And in some cases, connectors that aren’t properly cleaned can cause permanent damage forcing expensive and slow splice-in replacements.
Lens-based “expanded beam” connectors offer a practical answer. They trade a small increase in average loss for a dramatic reduction in the long tail of bad links. Early estimates suggest deployment speed improvements in the 30–35% range. And they behave more like the connectors technicians are used to: plug them in and they work.
Server racks: Rethinking how racks get built
NVIDIA’s Vera Rubin GPU platform is slated to arrive in the second half of the year, and the connectivity requirements follow the same trends we’ve been seeing.
Fiber density will continue to climb to more than 1,000 fibers per rack — and future designs may push toward 5,000. That density is unmanageable with a traditional on-site integration model, where every rack is effectively built like a custom house.
Now, hyperscalers are shifting toward factory-built integration. These thousands of fibers and hundreds of connectors get installed, labeled, tested and documented in controlled environments. Then the entire rack is shipped ready for final turn-up, moving labor off the critical path so site ready is achieved in a more expedite manor.
Another change within server racks is continued shift from copper to fiber. Unlike the scale-out network which is dominated by fiber interconnects, the scale-up network inside data center racks still run on copper. I believe starting next year, we’ll see the first field trials of fiber-fed scale-up fabrics linking multiple GPUs across multiple racks into a tightly synchronized, single computing “brain.” These fabrics will lean heavily on lens-based connectivity from day one to make those connections.
In summary, 2026 will not be the year that everything completely changes — but it will be the year that groundwork is laid which will usher in the next large shifts in AI connectivity. The operators that lean into these early shifts will enjoy the leanings and be more confident in their plans when the real scale arrives in 2027 and beyond.
