For the past several years, the data center conversation has centered on one thing: how to get enough power. And for good reason. AI training workloads are among the most energy-intensive computing tasks ever deployed commercially, demanding 100 to 200+ kW per rack and requiring vast campuses engineered around power delivery, cooling, and resilient infrastructure. The race to megawatts has been relentless.

But a second race is forming, one that will play out differently. As AI moves from being built to being used at scale, the infrastructure question is shifting. Inference is coming. And when it arrives as the dominant AI workload, location will matter in ways it simply has not for training.

Training does not care where it lives

This is not a criticism. It is an engineering reality. AI training is typically latency insensitive. According to McKinsey's December 2025 analysis of hyperscaler strategies, training workloads can tolerate delays of up to 100 milliseconds between adjacent regions, which allows operators to site them in remote, power-rich areas where grid capacity, land, and water are more readily available.

That is why so many large training campuses have been built in places optimized entirely for power availability and cost, geography aside.

The trade-off made sense. There is no customer waiting on the other end in real time. You can run it far from a metro if the power is there and the economics work.

Inference changes the equation

Inference is different in almost every dimension that matters to location strategy. It powers the real-time applications that end users actually interact with: search, chatbots, recommendation engines, autonomous systems, agents, and real-time fraud detection. Every one of those interactions is a latency event. And latency is a function of distance.

McKinsey projects that by 2030, inference will surpass training to become the dominant AI workload, representing more than half of all AI compute and roughly 30 to 40 percent of total data center demand. Inference is forecast to grow at a compounded annual rate of 35 percent over the next five years, reaching more than 90 gigawatts of demand globally. That is not a niche or secondary market. It is the future shape of AI infrastructure.

What makes this consequential for site selection is that inference workloads are increasingly co-located with the applications and storage they serve. Unlike training clusters designed for synchronized, high-density GPU parallelism, inference tasks are highly atomizable.

They can be distributed across smaller, closer nodes. And because inference costs are recurring and directly tied to revenue generation rather than being major focused capital events, operators are far more sensitive to performance variability. A degraded inference experience is not an internal problem. It is a customer-facing one.

The geography of inference is not the geography of training

The same McKinsey research notes that Tier 1 hubs like northern Virginia and Santa Clara, which together account for roughly 30 percent of US data center capacity, are now severely constrained by grid congestion, multiyear permitting timelines, and land costs exceeding two million dollars per acre.

Lead times for new capacity in these markets are now frequently hitting the five-to-seven-year mark. That makes them increasingly unattractive for inference-heavy builds, where speed to deployment and network proximity both matter.

As a result, hyperscalers are actively pivoting toward Tier 2 and 3 markets where power can be delivered 12 to 24 months faster and land costs run up to 70 percent lower. The criteria driving those decisions are no longer just about raw megawatts. They include access to dense fiber networks, proximity to user populations and application workloads, and the ability to build mixed-use campuses that combine inference and general compute in the same physical footprint.

McKinsey notes that already about 70 percent of new core campuses combine general compute and inference, often separated by buildings or data halls, with inference racks positioned closer to access points, storage, and networking zones.

Fiber connectivity is not a secondary consideration in this world. It is foundational. Inference performance depends on low round-trip time between the data center and the end application. That requires multiple diverse fiber paths, high-speed interconnects, and proximity to the network aggregation points where traffic actually flows. A site that is power-rich but fiber-poor is not an inference platform.

Central Texas is positioned for this shift

Greater Austin, one of the fastest-growing metros in the country, is emerging as one of the most compelling Tier 2 markets for exactly these reasons. The region combines near-term power deliverability, a rapidly growing AI and semiconductor ecosystem, access to dense regional fiber corridors, and a cost and permitting environment that Tier 1 markets can no longer match. Those attributes matter for AI training today. They become even more important for inference as buyers begin designing for latency-sensitive, revenue-critical workloads.

At Blueprint Data Centers, we structured our Austin area dual-campus in Taylor and Georgetown with this trajectory in mind. These sites are not just about delivering megawatts on a near-term schedule, though that remains a critical advantage right now. They are about giving buyers a platform that works across the full arc of AI workloads: training today, inference at scale as that market matures.

Both sites sit on dense regional fiber corridors with multiple Tier 1 and regional network operators providing diverse paths into each campus, enabling the low round-trip times that inference-sensitive applications will require. The dual-campus design further allows customers to architect resilient, distributed deployments across both locations without leaving the greater Austin metro.

The buyers thinking carefully right now are not just asking whether they can get power. They are beginning to ask a second question: when inference becomes their dominant workload, will this site still perform? The answer depends on location, fiber, and the ability to build flexibly across multiple nodes in a coherent market. That is the brief we have been building to.

The inference era is not here yet at full scale. But the infrastructure decisions being made today, including which sites to commit to, which markets to anchor in, and which partners to work with, will determine who is well-positioned when it arrives. Training bought time. Inference will reward location.