The race to build AI data centers has hit an unexpected wall: the wiring connecting GPUs together is becoming just as important as the chips themselves. As artificial intelligence systems grow larger and more distributed across multiple buildings and campuses, the speed and efficiency of data center networking has shifted from a minor detail to a major competitive advantage. Three distinct trends are reshaping how companies think about AI infrastructure, from fiber optics to custom silicon to geographic strategy (Source 1, 2, 3).

Why Is Data Center Networking Suddenly Critical?

When AI training happened on a single rack of graphics processing units (GPUs), the physical distance between chips barely mattered. But as projects scale to train the largest language models, compute is now distributed across multiple data center campuses, sometimes hundreds of miles apart. This creates a fundamental problem: the time it takes for data to travel between GPUs can become a bottleneck that slows down the entire system.

One startup, Relativity Networks, is betting that faster fiber optics could reshape the entire geography of data center buildouts. The company raised $22 million in seed funding and secured a $40 million follow-on order from a major hyperscaler to deploy hollow-core fiber technology, which transmits data through a vacuum chamber instead of traditional glass. The result is a 50 percent reduction in latency, or the time it takes a signal to travel one kilometer.

"The largest systems are distributing the compute across multiple campuses to reach the power that exists. They're moving to where the warm shell is, but they still need to operate as one synchronized machine," said Jason Eichenholz, CEO of Relativity Networks.

Jason Eichenholz, CEO at Relativity Networks

In practical terms, reducing latency from five microseconds to three and a half microseconds per kilometer might sound trivial. But when you're coordinating billions of calculations across a sprawling campus, those microseconds compound. By cutting latency in half, developers can span 50 percent larger distances before the network becomes a constraint. For companies building multi-campus AI factories, this could unlock entirely new locations for data center development.

How Are Companies Solving the Networking Problem?

The networking layer of AI infrastructure has become the most reliable profit center in the entire buildout. While headlines focus on GPU shortages and hyperscaler spending, the switches, optics, and custom silicon that stitch GPU clusters together are growing at rates that rival the chip makers themselves.

Three companies have separated from the pack with strong financial results and raised guidance:

• Arista Networks: Specializes in Ethernet-based AI fabrics and has become the default choice for hyperscalers. The company reported revenue of just over $3 billion in Q2 FY26, up 37.7 percent year over year, with multi-year purchase commitments nearly tripling to $9.7 billion. CEO Jayshree Ullal noted that the company now serves over 100 cumulative customers with its EtherLink switches, up from just four to five customers in 2024.
• Broadcom: Sits at the intersection of custom AI accelerators and the switching silicon that clusters them together. The company reported AI semiconductor revenue of $10.8 billion in Q2 FY26, up 143 percent year over year, with networking making up almost 40 percent of that total. The company plans to ship 10 gigawatts of capacity in fiscal 2027.
• Marvell Technology: Focuses on optical interconnect and custom silicon, with data center revenue of $1.83 billion in Q1 FY27, up 27 percent year over year. The company raised full-year FY27 revenue guidance to approximately $11.5 billion, representing 40 percent year-over-year growth.

These three companies are not household names, but they are quietly dominating a niche that is becoming central to the AI infrastructure buildout. Each posted earnings beats in their most recent quarters and raised guidance, signaling confidence in sustained demand.

Where Are Data Centers Moving, and Why Does Networking Matter?

Geography is reshaping the data center industry in real time. The Nordic region, which includes Finland, Norway, and Sweden, has emerged as one of the most attractive locations for AI infrastructure development. Companies are drawn to the region because of abundant access to renewable power, available land, and a naturally cool climate that reduces cooling costs.

Nvidia, the dominant GPU maker, is playing an active role in this geographic shift. The company is connecting GPU customers with data center operators in the Nordics that have available capacity, effectively acting as a matchmaker between companies with chips and companies with infrastructure. This reflects Nvidia's broader strategy to expand its influence beyond just selling chips, into shaping the entire AI infrastructure ecosystem.

"How can we help them obtain land, power, shell? How do we help them in terms of standing up the compute as fast as possible for what they need to do?" said Colette Kress, Chief Financial Officer at Nvidia.

Colette Kress, Chief Financial Officer at Nvidia

The scale of development in the Nordics is staggering. Multiple multi-hundred-megawatt facilities have been announced in recent months, including a 110 megawatt campus in Finland with potential to scale beyond 550 megawatts, and a facility in Norway with up to 500 megawatts of capacity. There is currently 2.3 gigawatts of data center capacity queuing for future connections to the power grid in Norway alone.

How to Evaluate Data Center Infrastructure Investments

For investors and companies tracking the AI infrastructure buildout, understanding the networking layer is essential. Here are the key factors to monitor:

• Backlog Growth: Watch for multi-year purchase commitments and backlog expansion. Arista's backlog nearly tripled to $9.7 billion, signaling sustained demand from hyperscalers.
• Guidance Raises: Companies that raise revenue guidance after beating earnings are signaling confidence in the underlying demand. All three networking leaders posted beat-and-raise quarters.
• Geographic Concentration: Monitor where new data center capacity is being built. The Nordics are attracting large-scale AI projects because of power access, land availability, and climate advantages.
• Latency Innovation: Emerging technologies like hollow-core fiber could unlock new geographies for data center development by reducing the distance constraint on distributed computing.

The third era of AI optimization is now underway. The first era optimized for raw compute power, the second optimized networking inside the data center, and the third is optimizing geography itself. Companies that can solve the networking and latency puzzle will have a significant advantage in determining where the next generation of AI infrastructure gets built.

The data center power crisis that dominated headlines in 2025 and early 2026 was only half the story. The other half is networking, and that story is just beginning to unfold.