Data center capital expenditure is surging at a pace few infrastructure sectors have seen in decades, and the reason is not generic cloud growth but one specific workload: artificial intelligence. McKinsey's June 2026 analysis says global demand could almost triple between 2025 and 2030, from about 82 gigawatts to roughly 220 gigawatts. Of that jump, AI-related demand is expected to grow 3.5 times, from about 44 gigawatts to 155 gigawatts, accounting for about 70 percent of total data center demand. The physical backbone behind cloud computing has become the bottleneck everyone is fighting to own.
It's not a uniform race. The United States holds the largest installed hyperscaler capacity at 36 gigawatts, with Northern Virginia as the single biggest cluster globally. China follows at 20 gigawatts, concentrated in Beijing, Shanghai, and the eastern manufacturing belt, though the government is pushing new capacity into western regions under the Eastern Data, Western Computing strategy. The UK has 2.2 gigawatts, Sweden 0.8, Singapore 1.4, and the global total is roughly 83 gigawatts today. Those numbers don't include the new campuses being planned at hundreds of megawatts or even gigawatt scale.
According to McKinsey, global investment in data centers excluding IT hardware may exceed $1.7 trillion cumulatively through 2030. That's not just about bigger buildings; it's about liquid-cooled facilities, faster delivery timelines, and a scramble for advanced accelerators. The distinction between training and inference is also sharpening. Training involves concentrated bursts of high consumption over shorter periods. Inference is what happens when the model answers prompts in real time. By 2030, training demand could rise to 62 gigawatts while inference climbs to 93 gigawatts, meaning the network of smaller, closer facilities may eventually outstrip the massive training campuses.
The Edge Is No Longer a Side Project
While hyperscalers race for gigawatt training campuses, a different layer is growing just as quickly. Azura Consultancy reports that by 2025 the edge data center has evolved into a vital layer of the global digital ecosystem. These facilities handle AI inference for manufacturing, real-time traffic optimization in smart cities, and low-latency streaming that centralized campuses simply can't deliver at acceptable speeds. The global edge data center market is projected to surpass $25 billion in value this year, with a compound annual growth rate exceeding 20 percent through 2030.
That growth is not abstract. Telecom operators began placing micro facilities near cell towers years ago to handle content delivery and reduce network strain. By the early 2020s, 5G and advances in containerization let edge sites host full applications, not just caching. The facilities now combine compact physical footprints with direct fiber integration, often within a 10- to 20-millisecond latency range from the end user. They're modular by design, built to deploy in weeks rather than years.
The picture is clear: centralization handled storage and enterprise applications well, but latency-sensitive workloads have outgrown it. Smart factories using computer vision for quality control, autonomous vehicles processing sensor streams, and healthcare systems requiring sub-second diagnostics all push compute toward the source. It's not a preference; it's a physics problem.
Sovereign AI Demands Local Infrastructure
One development that gets less press but matters just as much is the sovereign AI push. On October 15, 2025, DDN and Yotta Data Services announced a multi-million-dollar deployment of DDN EXAScaler AI400X3 systems to power 8,000 NVIDIA B200 GPUs within Yotta's Petascale AI cloud infrastructure, branded Shakti Cloud. The deployment is part of India's sovereign AI initiative under the IndiaAI Mission, and it is designed to establish one of the largest sovereign AI platforms in Asia.
The implications extend beyond India. Sovereign AI isn't only about data residency; it's about controlling the supply chain for training and inference. When governments demand that regulated industries, from banking to healthcare, keep data inside national borders, they also need domestic GPU clusters, storage architectures that scale linearly, and multi-tenant security policies that meet local compliance. DDN's EXAScaler runs parallel file system architecture delivering high bandwidth and high IOPS, with QoS policies critical for regulated sectors. It's not a generic cloud stack; it's an infrastructure designed for control.
Sunil Gupta, co-founder and CEO of Yotta, said the partnership reinforces a commitment to advancing India's AI journey in alignment with the IndiaAI Mission. The platform provides government, public sector, and enterprises with advanced AI infrastructure without moving data offshore. That model is replicating elsewhere. Nations that once relied on global hyperscaler regions are now building local campuses, not to replace them but to protect the most sensitive workloads.
Why This Matters Right Now
The stakes are immediate and practical. If you are running AI inference today, you are either renting space in a hyperscale campus with growing wait times, or you are looking at edge sites that can deploy faster but at smaller scale. The gap between those options is where the next wave of infrastructure investment is landing. McKinsey notes that facilities once planned for tens of megawatts are now targeting hundreds of megawatts or even gigawatts, especially in the US and China, while inference capacity is expected to spread closer to users.
It's also not clear this will hold at current velocity. Electricity constraints have already pushed Singapore's incremental growth into nearby Johor, Malaysia, and Malaysia recently restricted non-AI data centers citing energy concerns. Grid limits, cooling requirements, and the time to connect power are becoming real constraints, not footnotes. Liquid cooling and modular design are responses, but they add cost and complexity.
For the broader cloud and edge computing market, the message is that growth is no longer just about adding capacity. It's about where compute lives relative to the data it serves. Training camps will stay large and centralized. Inference will spread. Sovereign initiatives will build separately. And the firms that can deliver both scale and proximity, without compromising security or compliance, will define the next phase.
The direction is set. The numbers are in. The only open question is whether construction, power, and talent can keep pace with demand that is already outpacing every forecast from three years ago.
Read more from our desk coverage of cloud and edge computing. For a deeper analysis of colocation economics behind AI, see the original McKinsey report.