The geography of AI infrastructure is shifting faster than most procurement teams realize, and the clients who map those moves earliest will lock in the best terms.
What Happened
Three signals this week define the new terrain. First, Data Center Knowledge reports that Texas has overtaken Northern Virginia (NoVa, the largest US data center market) in global data center rankings, driven by available power capacity and land that the saturated NoVa corridor simply cannot match. Dallas and West Texas are now serious alternatives for large-footprint deployments, not backup options.
Second, a counterforce is pulling in the opposite direction. Data Center Knowledge also reports that production AI inference workloads, the revenue-generating layer where latency directly affects user experience, are dragging GPU deployments back into urban metro colocation facilities. The build-to-the-edge logic is straightforward: a 50ms round-trip to a remote hyperscale campus is unusable for real-time applications. A metro Tier III (data center reliability tier, 99.982% uptime) colo in Chicago or Northern New Jersey is not.
Third, Data Center Dynamics flags a constraint that almost nobody is pricing into their site selection: fiber scarcity at AI factory locations. As cluster interconnect (the high-speed fabric linking GPUs within a rack and across racks) scales, external bandwidth becomes equally critical. Sites that look attractive on power and land may be stranded assets if dark fiber (unused fiber-optic cable available for lease) inventory is thin. This is an early due-diligence item, not a closing checklist item.
Why It Matters
The geographic bifurcation reflects two fundamentally different workload profiles. Training and fine-tuning are power-hungry, latency-tolerant, and benefit from cheap land and abundant grid capacity, making Texas and emerging West Texas markets increasingly logical. Inference, by contrast, is latency-sensitive and demands proximity to end users, which pushes demand back toward established metro colo operators: Equinix, Digital Realty, CyrusOne, QTS, and Aligned all operate dense metro footprints across NoVa, Dallas, Phoenix, Chicago, Atlanta, NYC/NJ, Silicon Valley, and Seattle.
At the same time, power resilience is becoming a non-negotiable evaluation criterion. Data Center Dynamics notes that AI workloads are forcing operators to rethink UPS (Uninterruptible Power Supply) chemistry and duty cycles across their resilience stacks. Transient load spikes from large GPU clusters stress battery infrastructure in ways traditional enterprise workloads never did. Clients leasing colo space need to ask specific questions about UPS topology and battery chemistry, not just headline uptime SLAs (Service Level Agreements, which define uptime guarantees).
For sovereign AI programs and Fortune 500 enterprises building out dedicated infrastructure for the first time, the hyperscaler default is increasingly expensive and slow. Hyperscaler wait lists for H200 and B200 capacity stretch quarters. Neocloud operators (specialized GPU cloud providers, an alternative to hyperscalers) frequently offer the same generation hardware in weeks, at 30-50% lower cost on reserved commitments. The operators we work with hold capacity across both large-campus and metro-adjacent facilities, which matters when a client needs to split training and inference across different physical footprints.
One more structural force: Nvidia's latest earnings show that AI spending is rapidly expanding into networking and optics. Clients budgeting only for GPU line items are under-provisioning. A 1,000-GPU cluster requires a serious investment in the interconnect fabric (the switching and cabling layer connecting GPUs), and those costs are no longer rounding errors.
What Clients Should Do
If you are a frontier lab or large scaleup planning a training cluster above 4,000 GPUs, Texas and West Texas deserve a serious look on power availability and cost per megawatt. But stress-test fiber connectivity before signing an MSA (Master Service Agreement, the parent contract). Fiber gaps at the site level can be fatal to cluster performance regardless of how good the power story looks.
If you are a Fortune 500 enterprise or a sovereign AI program deploying inference infrastructure, metro colo is not a compromise. It is the architecturally correct answer for latency-sensitive workloads. Equinix and Digital Realty both offer dense metro footprints with carrier-neutral interconnect. Pair metro colo space with reserved GPU capacity from a neocloud operator and you have a production-grade inference stack at a fraction of hyperscaler cost.
If you are a system integrator sourcing for multiple end clients, the portfolio model is the only defensible approach: hyperscalers for burst and compliance-driven workloads, neocloud operators for reserved training and inference capacity, and owned or leased colo for latency-critical production tiers. Single-vendor strategies are leaving money and performance on the table.
Start these conversations earlier than feels necessary. Colo operators are quoting longer lead times on power-dense deployments, and neocloud operators with the best pricing fill their near-term inventory first.
Work With XIRR Advisors
XIRR Advisors brokers reserved GPU capacity from neocloud operators and Tier III colocation space across the USA. We do not resell hyperscaler products. AWS, Azure, GCP, and Oracle sell direct. Our value is in the neocloud and colo markets where pricing, speed, and contract flexibility are all negotiable, and where having a specialist in your corner materially changes outcomes.
Share your requirements: region, GPU type (H100, H200, B200, GB200, GB300), cluster size, and timing, or megawatts if you are evaluating colocation. We will canvas the neocloud and colo markets on your behalf and return a shortlist within 48 hours. The earlier you engage, the better the terms available. Our fee is paid by the provider. Clients pay nothing. Reach us at contact@xirradvisors.com or DM @XIRRAdvisors.
References
[1] Data Center Knowledge: Texas Powers Past Virginia in Global Data Center Rankings
[2] Data Center Knowledge: AI Inference Pulls Infrastructure Back Into Metro Data Centers
[3] Data Center Dynamics: AI Factories Have a Fiber Problem
[4] Data Center Dynamics: Water, Grid Volatility and the Growing Burden on the UPS Battery Layer
[5] Data Center Knowledge: Nvidia Earnings Show AI Spending Moving Beyond GPUs
Share your requirements. We'll canvas the market.
Tell us your needs (region, GPU type, capacity, timing — or MW for colocation) and we'll canvas the neocloud and colocation markets on your behalf. Shortlist in 48 hours.
Earlier conversations get better terms. When you engage early, we have time to negotiate with vendors before you need to commit. You pay nothing. Provider-paid model.