The binding constraint on AI infrastructure in 2026 is no longer the GPU. It is the wall outlet behind it.
What Happened
Three stories this week illustrate the structural shift underway in AI infrastructure real estate.
First, Prime Data Centers broke ground on a 240MW campus in Phoenix, Arizona, with three of five buildings already pre-leased to a single leading hyperscaler before a shovel touched the ground. That is not a leasing market. That is a land grab. Pre-leasing at that scale, before construction completes, means available Tier III (data center reliability tier, 99.982% uptime) inventory in Sun Belt markets is being consumed faster than it is being built. Phoenix joins Northern Virginia (NoVa) and Dallas as markets where speculative capacity is a fiction.
Second, Data Center Dynamics published a pointed analysis on how AI growth is running directly into power and thermal engineering limits. Power Usage Effectiveness (PUE, the ratio of total facility power to IT equipment power, with lower being better), cooling density, and grid interconnection timelines are now hard ceilings on how fast any operator, hyperscaler or otherwise, can bring capacity online. Liquid cooling infrastructure, once optional, is now a baseline requirement for any facility running modern GPU racks.
Third, Data Center Knowledge laid out the nuclear power thesis for AI-scale campuses. Long-duration, carbon-low nuclear baseload is increasingly the only grid solution that pencils out for multi-gigawatt commitments. PPAs (Power Purchase Agreements, long-term electricity contracts) tied to nuclear capacity are becoming a competitive moat for operators who can secure them. Those who cannot are competing for constrained grid interconnection queues that stretch years.
Why It Matters
The GPU cost curve is not helping. Tom's Hardware Pro reported that Nvidia's next-generation Rubin AI systems are pricing at $7.8 million per unit, with HBM (High-Bandwidth Memory, the memory architecture used in modern GPUs) costs up 485% and now comprising 25% of total system cost. Separately, Data Center Knowledge noted that Nvidia's earnings show AI spending is spreading beyond GPUs into networking fabric and optical interconnect. Infrastructure budgets that were once dominated by a single GPU line item now require serious allocation across compute, memory, switching, and cabling.
The implication for sovereign AI programs and Fortune 500 enterprises planning first-generation AI infrastructure is significant. The cost of a wrong site decision, one with inadequate power capacity, inadequate cooling density, or a grid interconnection queue measured in years, now compounds across an entire hardware stack that is appreciating, not depreciating, in price. Getting the real estate right is upstream of every other decision.
For frontier labs planning large training clusters, the pre-leasing dynamics at Phoenix and comparable NoVa and Dallas campuses mean that available Tier III space with power already contracted is genuinely scarce. Hyperscalers (the largest cloud providers: AWS, Azure, GCP, Oracle) are absorbing inventory at a pace that leaves little room for smaller but serious operators. Neoclouds (specialized GPU cloud providers, an alternative to hyperscalers) that own or have contractually secured their own data center footprints are increasingly differentiated from those that do not.
For AI scaleups ramping inference workloads, the Dell repatriation trend documented by The Next Platform adds another layer. Cloud repatriation, moving workloads from public cloud back to on-premises or colocation infrastructure, requires securing colocation space before the hardware arrives. Lead times for both are converging and competing.
What Clients Should Do
If you are a Fortune 500 enterprise beginning AI infrastructure planning, start with the power conversation, not the GPU conversation. Identify candidate colocation markets (NoVa, Dallas, Phoenix, Chicago, Atlanta, Silicon Valley, Seattle, NYC/NJ) and assess what operators, including Equinix, Digital Realty, CyrusOne, QTS, Aligned, Iron Mountain, Compass, Stack, Vantage, and Switch, can actually deliver in terms of contracted power and cooling density within your timeline. Many enterprise clients discover that the market they assumed had capacity does not, and the market they overlooked does.
If you are a frontier lab or sovereign AI program with a defined MW (megawatt) target, the time to negotiate colocation terms is before your hardware procurement is complete, not after. Facilities that can support liquid cooling, high-density racks, and behind-the-meter (on-site power generation that bypasses the public grid) generation options command a premium and fill first. Waiting until GPU delivery is imminent leaves you negotiating from weakness.
If you are a scaleup evaluating hyperscaler reserved instances against neocloud operators, the pricing gap remains real. Neocloud operators consistently come in 30-50% below hyperscaler reserved pricing, with faster ramp times measured in weeks rather than quarters, and more flexible contract terms. The smart portfolio approach is hyperscaler for existing workloads with existing commitments, plus one or two neocloud operators for net-new GPU capacity, plus direct colocation for teams with the operational capability to run their own stack.
Earlier conversations produce better terms on both GPU capacity and colocation. Operators price certainty. Clients who arrive with defined requirements and realistic timelines get access to inventory that never hits the open market.
XIRR Advisors brokers reserved GPU capacity from neocloud operators and Tier III colocation space across the USA. We do not broker hyperscalers. AWS, Azure, GCP, and Oracle sell direct. Our role is to canvas the neocloud and colocation markets on your behalf, run the comparison, and return a shortlist within 48 hours. Share your requirements: region, GPU type, capacity target, timing, and MW if colocation is in scope. Many clients need both GPU capacity and a facility to run it. We handle both sides of that equation.
The provider pays our fee. Clients pay nothing. Reach us at contact@xirradvisors.com or DM @XIRRAdvisors.
References
[1] Data Center Dynamics: Prime Data Centers breaks ground on 240MW Phoenix hyperscaler campus
[2] Data Center Dynamics: AI growth is running into a power and heat constraint
[3] Data Center Knowledge: The nuclear option: data centers and the responsible provision of power
[4] Tom's Hardware Pro: Nvidia's memory costs soar 485%; AI systems now cost $7.8M to build
[5] Data Center Knowledge: Nvidia earnings show AI spending moving beyond GPUs
[6] The Next Platform: Dell bulks up hardware as AI infrastructure shifts to on-premises
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.