The data center network has reached a critical point. As companies race to build massive XPU clusters for AI, traditional technologies simply can't keep up. Competition is intensifying, with multiple chip manufacturers now vying for position in the AI infrastructure market, each pushing the boundaries of what's possible.

The success of these XPU deployments depends on infrastructure that can support demanding AI workloads – and that starts with the network. To sustain performance at scale, the network must deliver far higher capacity and lower latency between servers, racks, and clusters. Traditional intra-data center optics (IMDD/PAM4) have supported these connections for years, but as data rates climb beyond 800 Gb/s (800G) toward 1.6 Tb/s (1.6T) and 3.2T, the reach they can achieve diminishes.

Power and space constraints are also forcing compute distribution across multiple locations, increasing the distances between facilities. As facilities spread out, the network challenge shifts from connecting racks within a building to connecting entire data center campuses.

Why traditional optics can’t keep up

For years, intensity modulation direct detect (IMDD) technology has been the go-to for short-reach data center connections. It’s relatively simple, low-power, and cost-effective, which is ideal for connecting servers within a single facility. But data center sites are expanding rapidly. Large metro campus data centers now span five or six massive buildings with cable runs approaching 20 kilometers. IMDD wasn't designed for this scale at next-generation speeds.

The real challenge isn't just distance, it's power. AI infrastructure consumes unprecedented amounts of electricity, and you can't concentrate all that compute in one location because the stress on the power grid would be too great. This means distributing compute functions across multiple locations, sometimes 1,000 kilometers or more apart, creating networking requirements that IMDD just cannot meet.

How coherent delivers more at scale

The optical networking industry has seen this before. In the wide area network (WAN), coherent optical technology replaced IMDD at 100G and became the de facto standard. Since then, coherent has become so dominant that you don't hear about IMDD in the WAN anymore. The same change is now starting to be seen in the data center, where coherent will be deployed for high capacity, longer reach applications while IMDD continues to serve shorter distances where it provides adequate reach

Coherent optics use sophisticated modulation techniques like quadrature amplitude modulation (QAM) and manipulate the amplitude and phase of light waves to carry more data per wavelength. A digital signal processor compensates for signal degradation over longer distances, providing significantly better performance than IMDD can achieve. The result is higher data rates, longer reach, greater spectral efficiency, and lower power consumption at scale.

Equipment vendors are now developing solutions that tune down WAN-grade coherent for shorter distances, delivering 1.6T capacity in “coherent-lite” pluggable form factors suitable for metro campus deployments between two kilometers and 20 kilometers.

Technology ready, infrastructure must be prioritized

However, technology transitions take time because having the technology available is only half the battle. Pluggable coherent modules are only valuable if you have the network infrastructure that can support them. Leading hyperscalers and operators have already made the leap, with several upgrading to 800G routers and moving to 1.6T, highlighting the critical role of these advancements for AI-ready networks.

Despite early adopters of 800G and 1.6T, the reality is that even cutting-edge vendors are still selling significant volumes of technology that's three generations old. Previous-generation equipment continues to dominate sales for 12 to 24 months after new generations launch, as operators upgrade infrastructure at their own pace based on actual need rather than what's technically possible.

New partnership models

The demand for AI infrastructure is creating new kinds of partnerships. Even though cloud providers have built their own networks, they can't build fast enough to meet current demand. Building a new data center usually takes between 18 and 30 months due to the complexities of site selection, permitting, construction, and commissioning.

Because of these timelines, cloud providers cannot rely solely on constructing new facilities to meet immediate demand. Instead, they partner with infrastructure and network providers to secure wavelengths and network capacity on a short-term basis. These partnerships provide rapid access to the connectivity needed to support AI demands today while new data centers are still under construction.

Looking ahead

The question isn’t whether coherent will play a major role in data centers, as the WAN precedent makes the answer clear. The questions are when and how? For metro campus applications, the transition is happening now. For intra-data center applications, the timeline depends on GPU buildout pace and infrastructure upgrade cycles.

What's certain is the scale. This buildout is unlike anything we’ve seen, reimagining how compute infrastructure is distributed, connected, and powered. Operators who strategically plan their upgrades, evaluate coherent options, and explore partnerships for rapid capacity will be the best positioned for this shift. AI isn't just changing what we compute – it's changing how we connect.