Scale is easy to announce. Capability is harder to measure.
The AI infrastructure boom is usually described through scale.
More data centers. More gigawatts. Larger campuses. Faster chips. Bigger capital commitments.
By almost any conventional measure, the United States appears far ahead. One early-2026 dataset counted 3,960 operational data centers in the U.S., compared with 498 in the United Kingdom, 470 in Germany, 365 in China and 275 in India. Visual Capitalist
That is an extraordinary difference.
But it also raises a more interesting question.
What exactly are we counting?
The dataset itself includes everything from relatively small cloud facilities to sprawling colocation campuses. Each is counted as one data center. The number therefore tells us a great deal about physical presence, but considerably less about the amount of compute those facilities can actually deliver. Visual Capitalist
And that distinction may become increasingly important.
Scale is easy to announce. Capability is harder to measure.
What Does Capacity Actually Mean?
A data center can exist without sufficient power.
It can have access to electricity without enough transmission capacity to deliver what is required.
It can be connected to the grid but lack the cooling infrastructure required for increasingly dense AI workloads.
It can contain advanced computing hardware without sufficient network connectivity, customer demand or utilization to make that hardware economically productive.
So perhaps we need to become more precise when we use the word capacity.
There is:
- Announced capacity — what someone intends to build.
- Constructed capacity — what physically exists.
- Connected capacity — what can actually be energized.
- Compute capacity — what the installed hardware can process.
- Usable capacity — what can reliably serve real workloads.
- Economic capacity — what can produce sufficient value to justify the investment.
Those are not necessarily the same number.
A gigawatt on a development plan is not automatically a gigawatt of productive compute.
The Current Demand Is Real
This is not an argument that today’s data-center demand is imaginary.
Quite the opposite.
Across the major North American markets, CBRE reported 10,903 MW of existing supply and another 7,481 MW under construction in the first half of 2026. Vacancy fell to only 1.4%, while more than 80% of capacity under construction had already been preleased. Northern Virginia alone reached a vacancy rate of just 0.2%. CBRE
That is evidence of an extraordinarily tight market today.
But today’s shortage does not automatically tell us what the market will require five or ten years from now.
And that is where infrastructure becomes complicated.
BloombergNEF recently increased its forecast for U.S. installed data-center capacity to 118 GW by 2030 and 194 GW by 2035. It also noted how rapidly the announced project pipeline has expanded and how much uncertainty remains around what will ultimately be built. Development timelines are already measured in years: BNEF estimates roughly 5.3 years for hyperscaler projects and 8.4 years for non-hyperscaler developments. BloombergNEF
Technology can change considerably during eight years.
So can demand.
We Have Seen This Before
The pattern is not unique to data centers.
We have seen versions of it in warehouses, office buildings, retail centers, telecommunications infrastructure and other capital-intensive markets.
A market identifies growing demand. Capital recognizes the opportunity. Development accelerates.
Eventually the question changes.
It is no longer:
How much can we build?
It becomes:
Did we build what the market actually needs?
A warehouse is not valuable simply because the building exists. Location, transportation access, loading configuration, automation capability, labor availability and proximity to customers determine what the building can actually do.
Office markets have demonstrated something similar. A city can simultaneously have significant vacancy and a shortage of the kind of modern office space certain tenants actually want.
That is not simply excess capacity.
It is misaligned capacity.
AI infrastructure could eventually experience the same phenomenon, but with many more interdependent systems.
Location Changes Everything
Consider Frankfurt.
It remains one of Europe’s largest and most important data-center markets. Its inventory reached approximately 1,222 MW in early 2026, with vacancy around 5%. Demand remains strong.
Yet grid capacity in Frankfurt’s traditional data-center area is constrained, and major upgrades are not expected until the 2030s. New developments are consequently pushing farther away from the established center. CBRE
That reveals an important infrastructure paradox:
A location can become constrained precisely because it was successful.
- Connectivity attracted development.
- Development attracted more demand.
- Demand consumed available infrastructure.
Eventually, the advantage that created the cluster becomes one of the constraints on its continued expansion.
BloombergNEF sees a similar shift across Europe. Traditional hubs such as Frankfurt, London, Amsterdam, Paris and Dublin face combinations of grid congestion, power prices, interconnection delays and limited land, while developers increasingly examine markets such as the Nordics and Spain where energy and land conditions may be more favorable. BloombergNEF
The infrastructure map moves.
Then Look at India
India illustrates another part of the story.
Its data-center footprint remains much smaller than America’s by facility count, but its development trajectory is significant.
CBRE says India reached approximately 1.75 GW of operational data-center capacity by the end of the first half of 2026, with cumulative announced investment commitments of about $173 billion since 2021. Mumbai and Chennai remain dominant, supported by connectivity, cable landing stations and established commercial ecosystems. Power availability, renewable energy access, network connectivity and market access are increasingly important site-selection considerations. CBRE
Again, counting facilities tells only part of the story.
Different markets are building for different demands, different populations, different regulatory environments and different resource constraints.
There Is No Single Data-Center Race
This may be where much of the discussion becomes too simple.
AI training does not necessarily require the same infrastructure geography as inference.
Low-latency applications may need compute closer to users.
Large training workloads may place greater weight on inexpensive energy, enormous power availability and cooling.
Financial, government and regulated workloads may place greater weight on jurisdiction, data sovereignty and security.
Cloud services may prioritize another combination of connectivity, redundancy and market proximity.
So asking which country has the most data centers may eventually become about as informative as asking which country has the most warehouses.
Interesting. But incomplete.
The better questions are:
- What kind of compute is being built?
- Where is it being built?
- What demand will it serve?
- What resources will sustain it?
- And will all of those pieces arrive at the same time?
The Bottleneck Is the System
AI may look digital from the outside.
Its infrastructure is remarkably physical.
Generation → Transmission → Grid Connection → Transformers → Site → Cooling → Compute → Utilization.
Weakness anywhere in that chain can constrain the value of everything that follows.
That is why focusing only on the number of facilities, chips or announced gigawatts can become misleading.
It reminds me of judging transportation by the biggest and fastest car sitting in the driveway.
The specifications may be impressive.
But the meaningful questions are whether it reliably gets us where we need to go, how efficiently it does so, what it costs to operate and whether it is appropriate for the journey.
Infrastructure should be evaluated the same way.
!! Not biggest. Not fastest. Not most.
What matters is the capability delivered. The world will need massive investment in infrastructure
From Overcapacity to Misaligned Capacity
This is also why I am increasingly cautious with the word overcapacity.
The AI infrastructure risk may not simply be that we build too many data centers.
It may be that we build too much of the wrong capacity, in the wrong places, on timelines the surrounding infrastructure cannot support.
A country could theoretically have enormous, announced capacity while still experiencing shortages of economically usable compute.
A region could simultaneously have stranded facilities and insufficient capacity somewhere else.
Power could exist without transmission.
Buildings could exist without power.
Compute could exist without sufficient utilization.
And technology could change before long-lived infrastructure earns its expected return.
That is not merely overcapacity.
It is a failure of alignment.
What Are We Really Building?
The AI infrastructure boom is real.
So is the demand.
But perhaps the next phase of the discussion should move beyond measuring how much infrastructure is being announced and start examining how intelligently it is being assembled.
Because ultimately:
Infrastructure becomes valuable when capacity, location, resources, demand and timing align. Building more does not guarantee that alignment.
The global AI infrastructure race may therefore be less about who builds the most capacity than about who creates the most reliable, efficient and economically usable compute from the resources available.
Capital can finance a building.
Technology can fill it with servers.
But neither can manufacture transmission capacity, electricity, cooling, skilled labor, permits, connectivity or customer demand overnight.
Scale is easy to announce. Capability is harder to measure.
And perhaps the most important infrastructure bottleneck is not simply how much we can build.
It is whether we are planning the whole system well enough for everything to work when it arrives.