AI’s growth is increasingly constrained not by chips but by infrastructure, from data storage and energy demand to connectivity bottlenecks. As data centers expand, the weakest link shifts, threatening to limit AI’s future scale and performance.
Frontier AI companies are working with exabytes – billions of gigabytes – of data to train models and run real-time inference, and they need somewhere to store it all.
This is driving an astronomical surge in data centre development across the United States, with more than 1,500 projects now in the pipeline. If built, the number of operational data centres in the US would double. Even that would be unlikely to satisfy AI’s needs. McKinsey estimates global demand for data storage capacity could triple by 2030, with approximately 70% coming from AI workloads.
Each of these facilities is home to countless rows of servers that require vast amounts of electricity to run, and equally power-hungry cooling systems designed to manage the heat generated.
Collectively, the power demands of these proposed projects total 780 gigawatts of electricity. To put that into perspective, if all of these plans were approved, it would more than double the current strain on the US electrical grid.
We often view AI progress as a compute race. Who has the chips to build, train, and run the most advanced models? In reality, it’s an infrastructure race, and the ecosystem continues to lag behind ambition.
AI’s energy problem
Constrained by local pressures and regulatory scrutiny, only 25% of those projects are likely to reach fruition. The grid won’t suddenly collapse under a strain it can’t support. However, with consumption in high-performance servers used for AI growing by 30% annually, and with energy infrastructure lead times far slower than data centre construction, there will come a point where a power deficit halts growth in the AI space.
We’re already seeing construction timelines falling behind as a result. Building a new data centre, which used to take six months, now takes up to 18 months. And connecting them to the grid currently takes five to seven years.
Given the scale of these facilities, local grids need to expand their infrastructure with transformers, breakers, cables, and more before they can safely bring them online. Planning and reviewing changes can take years. Likewise, rising demand coupled with skilled labour shortages has seen procurement lead times skyrocket. Manufacturing high-voltage transformers, which once took six months, now takes eight times as long.
As energy scarcity grows, without the means to increase capacity quickly and at scale, supply and demand come into play. In areas near data centres, wholesale electricity costs have soared by as much as 267% in the past five years.
As AI develops, its main restraint won’t be computing or the data centres that support it, but the infrastructure that powers it.
Connectivity bottleneck
Data centres require power to get them online, but it’s connectivity that makes them operational.
Much of AI’s connectivity demand is invisible to end users. While real-time inference and API traffic require lower-latency networks, a significant proportion of AI traffic consists of moving enormous volumes of data between GPU clusters, storage systems, and geographically distributed data centres. Operators must also synchronise models, replicate datasets for resilience, and distribute updates globally.
These workloads are less sensitive to latency but place enormous demands on network capacity, creating a different type of bottleneck as AI infrastructure scales.
Beyond backend training and replication workloads, AI platforms must also support continuous user-facing and operational traffic. Inference requests, API calls, telemetry, logging, backups, and model distribution all contribute to sustained high-capacity data flows across global networks.
Together, they are creating pressure on networks in very different ways: one driven by sheer bandwidth demand, the other by responsiveness and low latency.
AI-driven traffic almost tripled last year alone, and continues to grow eight times faster than other traffic. Likewise, as consumers and enterprises adopt AI at scale, token consumption – the units of data processed by an AI model during training and inference – is expected to grow by 2,300% by 2030.
Reliable, high-capacity networks are essential for moving those vast volumes of data. If they fall short, training will slow, and inference will lag. While the focus is rightly on the readiness of energy infrastructure, the telecom ecosystem faces similar pressure to scale. Without investment in the connectivity backbone, even if a solution is found to AI’s energy constraints, the network bottleneck will continue to limit its expansion.
Capitalising on AI
Network service providers’ foremost task is to prepare operational, resilient, and connected infrastructure – and it should have been ready yesterday. While AI has accelerated demand, surging traffic is not a new challenge for the industry. Digital services, from social media to video streaming, have been steadily pushing networks to improve for years now. Subsequently, those providers best positioned to support AI are those that have already adapted, scaled capacity, and prepared their networks for the future.
That said, clear growth opportunities remain for the networking sector. More than half of data centre decision-makers expect AI to place a significant strain on interconnect infrastructure over the next few years, driving demand for fibre-heavy connections that link multiple centres together, enabling vast volumes of data to be shared seamlessly between them.
Likewise, edge computing is emerging as a natural extension of the data storage ecosystem, shifting processing closer to where data is generated and consumed to reduce latency. From AWS Wavelength to Google Distributed Cloud Edge, hyperscalers are turning to telcos to provide connectivity and enable edge deployment.
These opportunities, however, will not be evenly distributed. Those capable of supporting these demands will be the agile network service providers willing and able to scale quickly. Traditional, rigid operations that are slow to react, on the other hand, will have a hard time making up ground.
Keeping up with demand
AI’s growth will depend not on the ambition shown but on the speed at which the underlying infrastructure can be built. Ultimately, it will only ever be as strong as its weakest link.
Yesterday, that constraint was chips. Today it’s power. And tomorrow – without continued investment in resilient, high-capacity networks built to handle the projected, exponential growth in AI-driven data flows – it will be connectivity.
Tony is Chief Executive Officer at global network services provider RETN. He is instrumental in the company’s pricing, interconnection and network strategy and he likes to take a very active involvement in all other aspects of the business to truly understand any pressures and issues within the company. Tony is an F.C.C.A accountant. He joined RETN in 2007 and was initially tasked with establishing RETN’s market presence in Western Europe. Tony took a leading role in expanding RETN from a multinational carrier selling in 2 countries to a Pan-European player. In 2016 he took on the Chief Operating Officer position, transforming the company’s supply chain, pricing and operational strategy, moving to Hong Kong in 2017 to successfully establish the company’s Asian operations, turning it into the Eurasian network it is today. He took full responsibility as CEO at the beginning of 2021.
