Training, Inference, and the Geography of AI Compute: Not Every GPU Cluster Needs Northern Virginia

For most of the cloud era, data center geography followed a relatively predictable formula.

Build near major population centers. Build near dense fiber. Build near cloud ecosystems. Build where enterprises and network traffic already concentrate.

That formula created enormous data center markets in Northern Virginia, Silicon Valley, Dallas, Chicago, New York and other established hubs.

AI is changing the equation.

Not because those markets suddenly matter less.

Northern Virginia remains the largest data center market in the world and one of the most strategically important pieces of digital infrastructure anywhere. CBRE reported more than 4 GW of inventory in the market at the end of 2025, with approximately 1.1 GW of net absorption during the year and vacancy of just 0.5%. Its combination of fiber connectivity, cloud presence and proximity to major users remains extraordinarily difficult to replicate.

But AI introduces a different question:

Does every unit of compute actually need to be there?

Increasingly, the answer is no.

Training and inference create different infrastructure requirements, and even inference itself contains multiple categories of workloads with different latency tolerances.

As those differences become better understood, we believe AI will materially expand the geography of data center development.

The result will not be the disappearance of major data center markets.

It will be a more distributed compute map in which the workload increasingly determines the location.

Training Changed the Location Equation First

Large-scale AI training is unusual because enormous amounts of computing power can operate without continuously interacting with an end user.

The training cluster itself is extremely sensitive to latency.

Thousands or tens of thousands of accelerators may need to communicate with each other at extraordinary speeds through high-performance networking fabrics.

But that is primarily latency within the cluster.

The physical distance between the training campus and the person eventually using the model matters much less.

A model-training job does not generally need to sit twenty miles outside Washington, D.C. simply because users of that model eventually will.

That distinction changes site selection.

If the workload can tolerate geographic distance from the end user, infrastructure can move toward the resources it needs most: power, land, cooling capability, fiber, water where required, permitting, and increasingly, speed to energization.

McKinsey has noted that training and inference are beginning to drive different hyperscaler infrastructure strategies, with training supporting very large, high-density campuses while inference increasingly requires more regional infrastructure optimized around connectivity and responsiveness.

CBRE similarly reported in March 2026 that improved long-distance networks and growing AI-training demand are opening markets that historically may have been considered too remote for major data center development. It specifically identified states including Nevada, Pennsylvania and Michigan as increasingly attractive because of land and power availability.

That represents a fundamental change.

For training infrastructure, compute can increasingly move toward the electrons rather than requiring the electrons to move toward the compute.

Power-Rich Markets Suddenly Have a Digital Export Product

This creates an interesting economic dynamic for regions that historically were not major technology markets.

A location may have abundant generation, transmission infrastructure and inexpensive land while having relatively little local demand for hyperscale computing.

Historically, that could limit its attractiveness as a major data center market.

AI training changes that.

The product being created inside the data center can travel digitally.

A training campus in a power-rich secondary market does not need millions of consumers living immediately around it.

Its output is a trained model.

That means locations with favorable energy characteristics can effectively turn electricity into a globally valuable digital product.

From a power-market perspective, that is significant.

Regions with abundant generation, favorable electricity economics, existing transmission, available industrial land, supportive permitting, and strong long-haul fiber can increasingly compete for infrastructure that once would have been concentrated almost exclusively in established data center hubs.

The economic development opportunity follows.

AI infrastructure can potentially bring billions of dollars of capital investment into markets that would never have supported comparable traditional cloud demand.

Northern Virginia Still Shows Why Geography Matters

None of this means geography becomes irrelevant.

Northern Virginia is actually a useful illustration of both sides of the argument.

Its enormous scale did not happen by accident.

The market developed around exceptional fiber connectivity, proximity to government and enterprise customers, access to cloud ecosystems, and one of the world's densest concentrations of interconnected digital infrastructure.

Those advantages remain extremely valuable.

But the market also demonstrates what happens when demand begins colliding with physical infrastructure constraints.

CBRE reported in March 2026 that 96% of Northern Virginia's scheduled 2026 data center supply was already committed and that Dominion Energy's power-allocation process had extended delivery timelines for some new developments. Powered and entitled land remained exceptionally scarce.

That creates an economic trade-off.

If a workload needs Northern Virginia's network ecosystem, paying the premium may make complete sense.

If it does not, then competing for scarce power in the world's largest data center market may be unnecessary.

AI makes that distinction increasingly important.

Inference Changes the Map Again

Training is only part of AI compute.

Once a model has been trained, inference is what happens every time someone actually uses it.

A user asks a chatbot a question. A coding assistant generates software. An enterprise agent analyzes a document. An autonomous system makes a decision. A search engine generates an AI response. A model processes audio, video or images.

Those are inference workloads.

Unlike training, many inference applications interact directly with users or other systems in real time.

That means network latency can matter significantly more.

CBRE expects increasing inference demand to create a need for more regional and distributed data centers as AI expands across search, customer interaction and other real-time applications.

This is one reason the AI infrastructure map will probably not consolidate entirely into a handful of giant rural campuses.

Training pushes compute toward power.

Many forms of inference pull compute back toward users, data and networks.

The future likely requires both.

But "Inference Must Be at the Edge" Is Too Simple

There is another misconception developing around AI infrastructure.

If training can move away from cities, the argument goes, inference must move directly into them.

Sometimes that will be true.

But inference is not one workload.

Consider the difference between several applications.

A voice assistant carrying on a natural conversation may be highly sensitive to response time. An autonomous machine may require extremely fast decision-making. An interactive gaming or augmented-reality application may have very tight latency requirements.

Those workloads can benefit substantially from geographic proximity.

Now consider overnight document processing, large-scale summarization, synthetic-data generation, background enterprise analysis, batch inference, non-interactive media processing, or an AI agent completing a task that already takes several seconds or minutes.

Moving that workload another few hundred miles may have little meaningful effect on the end-user experience.

Uptime Institute's 2025 AI Infrastructure Survey illustrates this complexity. For inference location decisions, respondents ranked use of existing infrastructure, data sovereignty and power availability among the most important factors. Proximity to data and integrated applications mattered, but it was only one component of a broader infrastructure decision.

That is an important distinction.

Inference creates a latency budget.

Different workloads can spend that budget differently.

And the more latency a workload can tolerate, the wider the geographic area in which its compute can economically operate.

Geography Becomes an Optimization Problem

This creates a more sophisticated way to think about AI infrastructure.

Instead of asking "Where should AI data centers be located?" the better question is "Where should this particular workload be executed?"

An interactive consumer application may prioritize proximity to population centers. A regulated financial workload may prioritize data jurisdiction and security. An enterprise application may need proximity to existing cloud environments and proprietary datasets. A giant model-training run may prioritize hundreds of megawatts of inexpensive electricity. A background inference workload may be able to choose between several regions depending on available compute capacity and power economics.

That means geography increasingly becomes another variable in the economics of compute.

The AI stack can optimize across power price, available capacity, network latency, hardware availability, utilization, data location, and application performance.

This is a fundamentally different infrastructure model from assuming every workload belongs in the same handful of traditional cloud markets.

Power Cost Matters More When Geography Becomes Flexible

The ability to move compute geographically also makes electricity economics more important.

If a workload absolutely must operate in a specific market, the operator has limited ability to arbitrage energy cost.

Power becomes a cost of doing business in that location.

But if the workload can operate across multiple regions, electricity becomes part of the location decision.

A difference of several cents per kilowatt-hour can become meaningful when applied to hundreds of megawatts operating continuously.

The same is true of access to available generation.

A market offering slightly cheaper electricity but a five-year power-delivery timeline may be less attractive than a somewhat more expensive market capable of energizing the project much sooner.

The relevant variable is therefore not simply cheap power.

It is the combination of cost, availability, scalability and timing.

This is why AI development increasingly intersects with power-market strategy.

The developer is no longer merely asking whether a utility serves a particular property.

The broader question is whether the energy system surrounding that location can support the economic profile of the compute being deployed.

Network Improvements Expand the Radius

The other enabling factor is networking.

Long-distance digital connectivity continues to improve.

That matters because geographic distance is only problematic to the extent that it creates unacceptable performance.

CBRE specifically identified advances in long-distance networks as one reason AI-training infrastructure is expanding into markets previously considered too remote for hyperscale development.

This creates a powerful interaction between fiber and power.

A region does not necessarily need to become a traditional carrier-hotel market to become important to AI.

It needs sufficient connectivity to move enormous volumes of data reliably between the compute campus and the rest of the digital ecosystem.

As that capability expands, power-rich secondary markets become viable over increasingly large distances.

In that sense, long-haul fiber allows electricity-rich regions to participate in the AI economy without physically moving their electricity across hundreds of miles of transmission infrastructure.

The electricity is consumed locally.

The computational output moves over fiber.

Data Gravity Still Matters

There are limits to how far compute can move.

One is data.

AI workloads can involve enormous datasets.

Moving large quantities of information between environments can create network cost, latency and operational complexity.

Some enterprise datasets cannot easily leave a particular cloud. Some regulated data cannot leave a particular jurisdiction. Some applications depend heavily on databases or services already operating in another region.

This creates what the technology industry often describes as data gravity.

Compute may theoretically be able to run anywhere.

The information required by that compute may not.

That is another reason the future will probably consist of multiple infrastructure tiers rather than one universally optimal AI location.

Certain workloads will follow cheap power. Others will follow users. Others will follow data. And increasingly sophisticated AI platforms will determine how to distribute computing across all three.

Training and Inference May Eventually Share Infrastructure

The distinction between training and inference should also not be taken too literally.

A data center does not necessarily need to perform only one function forever.

Modern GPU infrastructure can support multiple workloads.

A large campus built initially around training could eventually serve inference. Hardware can be repurposed. Workloads can shift as customer demand changes.

And the boundary between training and inference itself is becoming less clean as techniques such as reinforcement learning, fine-tuning and continuous model improvement require recurring computation after initial model development.

That makes flexibility valuable.

A power-rich campus capable of hosting very large GPU clusters may remain useful even as the mix of workloads changes.

The more inference that can tolerate regional rather than metro-level latency, the larger that addressable market becomes.

This Is Good News for Secondary Data Center Markets

The implications for real estate and infrastructure development are substantial.

Historically, secondary markets often struggled with a chicken-and-egg problem.

Hyperscale customers wanted established infrastructure.

Infrastructure developers wanted committed hyperscale customers.

AI demand can break that cycle because certain workloads have less dependence on the surrounding local digital ecosystem.

CBRE has already documented development shifting toward emerging markets with faster access to power as established markets encounter constraints.

That does not mean every rural site with a transmission line becomes an AI campus.

Power remains only one requirement.

Large-scale AI infrastructure still needs high-quality fiber, appropriate land, permitting, cooling, equipment access, construction capability, and an energy pathway that is both technically and commercially executable.

But the geographic opportunity set is clearly expanding.

The Long-Term AI Map Will Look Different From the Cloud Map

The first generation of cloud infrastructure created massive concentrations of computing around a relatively small number of markets.

AI is likely to create something more distributed.

There will still be enormous network-centric hubs.

Northern Virginia will remain one.

So will other established cloud markets.

But alongside them we expect increasingly large power-centric compute campuses to emerge in markets where energy, land and infrastructure can support scale.

Regional inference facilities will occupy another layer.

Highly latency-sensitive applications may eventually create additional capacity closer to users.

The result could resemble a hierarchy: large centralized AI factories where power economics dominate, regional infrastructure balancing power and network proximity, and more localized infrastructure where extremely low latency is essential.

The workload determines which layer makes sense.

This Changes How We Think About Data Center Markets

For developers, investors and power providers, this has an important implication.

The historic question was often: "Is this an established data center market?"

AI makes that question less useful.

A better question may increasingly be: "What class of AI workload can this market support competitively?"

A location does not need to beat Northern Virginia at being Northern Virginia.

It may offer something entirely different.

A region with abundant power, large parcels and excellent long-haul connectivity may be a superior location for a massive training environment even if it has relatively little existing enterprise data center demand.

A metro market with expensive power but extraordinary connectivity may remain superior for latency-sensitive inference.

Neither market is inherently better.

They are serving different parts of the computing stack.

Bottom Line

AI is beginning to separate the geography of compute from the geography of users.

Training started that transition.

Large GPU clusters can increasingly locate where the combination of power, land, infrastructure and time to market is most attractive because the end user does not need to sit beside the training campus.

Inference introduces geography again—but in a much more nuanced way.

Some workloads need to be close to users. Some need to be close to data. Some need to remain within a particular jurisdiction. And many others may have enough latency flexibility to move toward cheaper or more readily available power.

That creates a new infrastructure landscape.

The future of AI will not be built entirely in Northern Virginia.

Nor will it be built entirely in remote gigawatt campuses.

It will be distributed according to the economics and performance requirements of the workloads themselves.

For power markets, that creates an extraordinary opportunity.

Regions with underutilized energy resources, available land and strong connectivity can increasingly convert those advantages into digital infrastructure.

For the data center industry, it changes what constitutes a viable market.

And for AI platforms, geography becomes another tool for reducing the cost of compute.

Not every GPU needs to be close to the user.

Sometimes the better strategy is to put the GPU close to the power.