Why constructing two information facilities per week received’t repair AI’s bottleneck
The AI race is now firmly underway, and those lacing up their running shoes have become fixated on two variables – power and compute, but no one is talking about the third and it’s the one that decides who gets there first.
By the end of the year, the top five hyperscalers – Meta, Apple, Microsoft, Amazon, and Alphabet (Google) – will have collectively invested around $600 billion on AI-enabling infrastructure like data centers, with some estimates suggesting that as many as two data center facilities are coming online every week to keep pace with demand.
In the US alone, there are now more than 3,000 data centers in operation, with a further 1,500 already in development.
Latest Videos From
Founder and CEO of Taara.
And almost none of them may deliver on its promise, because we are building cities without roads. The conversation around AI infrastructure still tends to start and end inside the data center, yet the real work of AI happens between them.
Training clusters need to exchange massive volumes of data. Inference is moving closer to users, spreading workloads across edge environments, cloud platforms, and regional hubs. None of that works without fast, reliable, high-capacity connectivity – and that’s where the real bottleneck is starting to emerge.
The networks built to move data simply aren’t keeping pace with the compute being built to generate it.
The world’s most expensive paperweights
AI chips represent some of the most valuable assets in modern infrastructure, with billions poured into GPUs and accelerators designed to power training and inference at scale. But unlike traditional capital investments, their value doesn’t hold.
In some cases, these chips can lose up to 90% of their value within 48 months, as the pace of innovation renders them obsolete almost as quickly as they’re deployed. That puts enormous pressure on organizations to extract value as quickly as possible.
And that’s where the dependency really reveals itself. Compute on its own doesn’t generate outcomes. It needs to be fed, continuously and at speed, by the data pipelines that connect systems, environments, and users.
As AI workloads become more distributed and data centers proliferate, that dependency will only intensify. Because inference is so latency-sensitive, it can’t run in big centralized locations – it needs to happen across edge environments, regional hubs, and cloud platforms, often within the same workflow.
At the same time, machine-to-machine traffic is surging as models and devices communicate, update, and respond in real time. All of this places unprecedented strain on the connections between systems, and when those connections fall short, performance suffers, latency rises, and expensive compute sits idle.
The chips don’t fail. The network does.
The infrastructure gap we can’t dig our way out of
For decades, fiber has been the backbone of global connectivity, and it will remain vital for many years to come. It’s reliable, high-capacity, and proven. But it’s also slow to extend, expensive to deploy, and dependent on construction timelines that bear no resemblance to the pace of AI innovation.
Even with increased funding and streamlined permitting, the gap is simply too wide. According to the Fiber Broadband Association, as of 2025 the US has deployed more than 159 million miles of fiber, yet an additional 213 million miles is still needed to meet the performance, scalability, and security requirements of AI-driven workloads.
AI demand evolves in months; fiber expansion happens in years. The future of connectivity won’t be buried underground.
From electrons to photons: a new model for connectivity
Every major technology shift has been defined by how information moves. Copper moved electrons through wire. Fiber moved light through glass. Now, as AI pushes infrastructure to its limits, a new approach is emerging that removes the medium altogether.
Wireless optical communication uses narrow, invisible beams of light to carry data between fixed points, establishing direct, high-capacity links that behave like fiber without the limitations that come with it.
Rather than broadcasting signals broadly, like a satellite, these systems create precise, point-to-point connections that can carry data over multiple kilometers with the consistency and throughput that AI workloads demand.
Where traditional infrastructure is defined by geography, permits, and multi-year construction cycles, light-based links can be deployed in hours and reconfigured as requirements evolve.
That opens up a fundamentally different model of network expansion, where capacity can be added exactly where it’s needed, whether that’s connecting data centers across a city, bridging infrastructure across difficult terrain, or increasing density and resilience around concentrated clusters of demand.
For hyperscalers racing to extract value from depreciating compute assets, for operators under pressure to expand without the time to lay new cable, and for enterprises that can’t wait years for capacity to arrive, that level of flexibility can’t come soon enough.
Connectivity will define the next growth phase of the AI era
The scale of AI investment shows no signs of slowing. If anything, it’s accelerating, with governments, hyperscalers, and enterprises all racing to secure the infrastructure needed to compete.
But the next phase of this buildout won’t be defined by how many data centers come online, how intelligent the next AI model is, or how powerful the next generation of chips are. It will be defined by how effectively those resources can be connected, coordinated, and put to work.
Because without the ability to move data quickly, reliably, and at scale, even the most advanced AI systems will struggle to deliver on their promise
Connectivity is no longer a supporting layer sitting beneath compute and storage. It’s the foundation that everything else depends on.
As the limits of traditional infrastructure come into sharper focus, a different approach is beginning to take hold where data travels not just through cables beneath our feet, or via latency-prone satellites in orbit, but through harnessing something we’ve had all along – light. We just need to follow it.
We’ve featured the best AI tool.
This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit