Where does AI live? Ask most technologists, and the answer comes fast: the cloud. But that’s only half the story.

In reality, AI lives in real-world infrastructure — in colocation data centers, bare metal racks, fiber lines, power substations and edge facilities. Its footprint is concentrated in just a handful of cities. And that physical geography is about to change, fast.

New data from Synergy Research Group reveals that just 20 cities and metro areas now account for 60% of the global colocation market. The cloud, it turns out, has favorite cities. But how long can this highly centralized model survive in an AI-driven world that’s rapidly evolving?

At the top of the global stack sits Northern Virginia, home to Ashburn’s data center alley. With nearly 7% of the world’s colocation capacity, it is the undisputed capital of the cloud—housing hyperscale infrastructure for Equinix, Digital Realty, QTS, NTT and others. Right behind them are:

Globally, the top 20 metros include 8 in the U.S., 7 in Asia-Pacific, 4 in Europe and 1 in Latin America. These cities have come to dominate due to a powerful mix of factors:

They have been the natural targets for AI training workloads, cloud gaming infrastructure, financial algorithm engines and large-scale enterprise applications.

But success comes at a cost. These metros are now straining under the weight of their own success — facing power constraints, skyrocketing real estate prices and limited land availability. As demand accelerates, so does the urgency to find new locations that offer scale without the gridlock.

Beyond the top 20, a new generation of AI capitals is forming — and fast. Cities like Johor in Malaysia, Jakarta in Indonesia, Chennai in India, Lagos in Nigeria, Rio de Janeiro in Brazil and Queretaro in Mexico are emerging as critical nodes in the global AI infrastructure economy.

This isn’t speculative—it’s already underway:

So why are these cities gaining ground?

The result? The AI infrastructure map is expanding—and decentralizing. These cities are no longer fringe players. They are strategic footholds in a rebalanced, post-hyperscale world.

AI efficiency gains are not linear — they are exponential.

These breakthroughs are rewriting the rules. As model efficiency improves, the gravitational pull of hyperscale clusters weakens. AI no longer has to live in Ashburn, Frankfurt or Singapore — it can live wherever the economics and latency profiles make sense.

The infrastructure that powers AI is being rearchitected — following a tiered, decentralized model:

This core–periphery–edge architecture is fast becoming the new normal.

As John Dinsdale, chief analyst at Synergy Research, notes — hyperscalers are already distributing workloads across geographically and economically diverse markets. The reasons go beyond performance — they now include power availability, regulatory pressure and geopolitical risk.

There’s yet another undercurrent transforming where AI lives — the rise of liquid cooling. As AI workloads grow denser and more power-hungry, traditional air-cooled data center designs are hitting physical and thermal limits. Liquid-cooled racks, once niche, are becoming essential to support the heat profiles of modern AI chips. This shift allows operators to pack more compute into smaller footprints — but it also reshapes data center layouts, fiber distribution, and cooling infrastructure. Regions with better access to water reuse systems and efficient cooling loops may now leapfrog others in attracting next-generation AI clusters.

Another potential tectonic shift is forming in AI infrastructure: workloads may soon stop defaulting to NVIDIA-powered GPUs. As open-source models like LLaMA, DeepSeek and Mistral gain traction — and as inference becomes lighter, faster and more portable — the long-held belief that GPU infrastructure is the only game in town is breaking down.

In its place, a new model could be emerging — one that emphasizes flexibility, cost efficiency and workload-specific optimization:

These changes are not theoretical. Lightweight, quantized models are already running with fewer than 5 billion parameters — yet delivering performance comparable to much larger large language models, often without the cost or latency penalties of cloud-hosted inference.

And for steady-state workloads — those that do not fluctuate wildly or require elastic scaling — bare metal is making a serious comeback. But this time, it’s powered by cloud-like automation, orchestration and developer tooling that were once the exclusive domain of hyperscalers.

This trend is not confined to AI. Even outside machine learning, companies are reassessing their reliance on public cloud platforms. Take 37signals, the maker of Basecamp and HEY — which recently migrated off Amazon Web Services after citing the exorbitant long-term costs of Amazon S3 and EC2. While unrelated to NVIDIA, the message is clear: cloud bills are being questioned.

This raises two critical questions:

If model efficiency continues to outpace Moore’s Law — and if enterprises start treating cloud as one option rather than a default — then bare metal may not just be back. It may be the next great frontier in infrastructure.

The bottleneck is no longer just compute — it’s power, policy and proximity.

Infrastructure strategy is no longer just technical — it is geopolitical. AI may have started in the "cloud" — but its future will be more distributed, more efficient and more tactical. The next generation of AI infrastructure will span:

So where does AI physically live? Today, mostly in Ashburn, Beijing and London. But that map is quickly redrawing itself — to regional markets, secondary metros and eventually all the way down to your device.