Every AI query, every model training run, every real-time inference request has to happen somewhere. That somewhere is a data center—and the explosive growth of artificial intelligence is pushing data center infrastructure to its limits.
Data centers have always been the physical foundation of the digital economy, but AI is a different tenant than what most facilities were built to serve. Traditional cloud workloads are power-hungry; AI workloads are power-intensive at a scale that strains existing infrastructure at every level—from the chip to the rack to the utility grid. A single rack of Nvidia's latest GPUs can draw nearly 100 kilowatts of power. Cooling systems designed for conventional servers can't handle the heat. And demand for this kind of capacity is accelerating faster than the industry can build to meet it.
At the same time, AI is not a monolithic workload. Training and inference have fundamentally different infrastructure requirements, which means AI growth is reshaping data center demand across geographies, not just concentrating it in the same places as before. New markets are emerging. New facility designs are required. And operators who built their businesses around yesterday's compute standards are now facing costly decisions about how to retrofit, rebuild, or reposition.
This analysis examines what AI means for data center deployments, how it is changing the operating environment inside facilities, and what challenges the industry must confront to meet demand that shows no signs of slowing.
A primary driver in surging AI use is the quick maturation of large language models (LLM). LLMs such as OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude use huge quantities of data and parameters to analyze and generate content. Other key applications driving AI's development include natural language processing, computer vision, and robotics.
Much of this explosion of AI has been made possible with Nvidia’s chips, and Nvidia's sudden market domination highlights the explosion of machine learning requirements. While the company began as a manufacturer of gaming-focused graphic processing units (GPUs), it is now applying its technology to cloud, supercomputing, and AI chipsets.
Data center sales now account for a vast and growing majority of Nvidia's revenues—nearly 90%. Its market capitalization—perhaps the key indicator of its expected growth potential—hit $5.149 trillion as of June 2026. For context, Nvidia only breached $1 trillion three years earlier. This puts the company far above that of other chip manufacturers—and above every company in the world by market capitalization. This boom reinforces that networks are shifting rapidly and aggressively toward AI applications.
What does the growth of AI say about future data center deployments? AI work consists of two basic phases: training and inference (the work for which the model is trained). Each has a different impact on the data center market.
AI will require many changes within data centers. The most immediate concerns are provisioning higher density cooling and higher capacity interconnections. LLMs need far more compute, memory, and cooling than typical cloud computing loads. According to a recent University of Washington study, the hundreds of millions of queries ChatGPT handles each day may use as much as 1 GWh of energy. This is roughly enough power to support 33,000 homes.
Nvidia has been incrementally increasing the computational power of its chipsets. Its most current state-of-the art Blackwell B200 AI GPUs use 1.2 kW per chip. The GB200 Superchip with dual B200 GPS and a Grace CPU use a total of 2.7 kW. Nvidia is combining superchips into rack-scale solutions acting as one unit. Its GB200 NVL36 uses 36 GPUs. The GB200 NVL72 combines 36 Superchips into one unit operating at 97.2 kW per rack.
The data center market is not ready to support these staggering requirements at the site, utility generation, nor transmission levels. Standard air-cooling systems can't support GPU-based, power-intensive AI applications, so liquid cooling solutions are becoming more prevalent. Some are direct-to-chip, where liquid coolant on a plate is integrated directly onto the CPU or GPU. Other solutions involve immersion liquid cooling, where operators immerse servers in dielectric liquid.
Many colocation operators now advertise their AI-readiness and ability to accommodate advanced applications. But it's hard to meet rising demand for high-density deployments, especially as targets for AI readiness also grow. Critical challenges must be addressed to effectively deploy AI. These include boosting power, reconfiguring PDU block and rack designs, and providing space to accommodate liquid cooling infrastructure. And these changes need to be made while also providing proper support for existing customers in increasingly hybridized environments.
While the data center market is hopeful in meeting the massive new demands of AI, many data centers cannot accommodate such computational advances without difficult and expensive retrofits. So many new AI-ready facilities—sites that can provide the cooling, power distribution, and other demands of AI servers—will be needed at a time when the data center market is under increased regulatory scrutiny and power is already scarce. The full effects of AI on data center markets remain unknown. However, it is clear that these technologies, after years of development and impressive results, are here to stay.
...Plus industry-leading data on the global data center market with our Data Center Research Service platform. We collect and analyze the telecom industry's most comprehensive intelligence on data center capacity, service providers, pricing trends, and more. Get a look at the platform below.