The bottleneck in AI infrastructure is no longer GPUs. It is the physical real estate, power, and cooling infrastructure required to run them.
This week's deal flow and policy news make that case clearly. The constraint picture has shifted from silicon scarcity toward site scarcity, and the organizations that recognize this earliest will lock in capacity at terms that simply will not be available eighteen months from now.
What Happened
Three stories from the past 72 hours define the new terrain.
First, regulatory friction is closing off markets faster than expected. Minneapolis imposed a six-month moratorium on data centers larger than 350,000 square feet, effectively freezing hyperscale development in the Upper Midwest near-term. Meanwhile, North Carolina's proposed SB 730 legislation would mandate minimum billing periods and closed-loop cooling requirements for AI data centers in the Southeast, a region that has been a go-to alternative to congested Northern Virginia (NoVa) and Dallas. Regulatory risk is no longer a footnote in site selection models. It is a primary variable.
Second, the infrastructure stack itself is getting more capital-intensive and pre-committed. Modine locked in $4 billion in cooling capacity reservations through 2029, treating thermal management supply as a strategic asset rather than a spot purchase. Separately, TeraWulf signed a $290 million deal with Schneider Electric that bundles direct-to-chip liquid cooling (a system that circulates coolant directly through server components rather than relying on room-level air) with lithium-ion UPS (uninterruptible power supply, the battery system that keeps servers running during grid fluctuations). Both deals signal that AI-grade facilities require long-lead infrastructure commitments that bear little resemblance to traditional enterprise data center procurement.
Third, sovereign capital is actively competing for the same Western sites. UAE-backed Core42 secured $550 million in HSBC financing to accelerate its buildout across Europe and the US. This is not a niche player. It is sovereign-backed infrastructure capital hunting the same Tier III (data center reliability tier, targeting 99.982% uptime) sites that Fortune 500 enterprises and frontier labs are evaluating. The competitive set for quality colocation space now includes government-aligned entities with patient capital and strategic mandates.
Why It Matters
The structural dynamic here is a supply-demand mismatch that is tightening from both ends simultaneously.
On the demand side, Nvidia's most recent earnings confirmed that AI infrastructure spending is expanding beyond GPU clusters into networking, optics, and full-stack interconnect. That means the physical footprint required per AI workload is growing, not shrinking. More networking gear means more rack space, more power draw, more cooling load.
On the supply side, site options are narrowing. The most promising near-term relief valve is Texas, where wind-built CREZ (Competitive Renewable Energy Zone, Texas's legacy transmission network built for wind power) grid corridors are creating rare alignment of cheap renewable power, available land, and transmission capacity. Dallas and the broader Texas corridor are real options. But they are not unlimited, and the region is already drawing aggressive interest.
For sovereign AI programs in the EU and US, the Core42 story is a signal, not background noise. Sovereign-backed operators are willing to pre-commit capital at scale and sign long-term PPAs (Power Purchase Agreements, multi-year electricity contracts) that lock in pricing. Independent clients without that balance sheet are at a structural disadvantage in open-market negotiations unless they move early and with clear requirements.
What Clients Should Do
If you are a frontier AI lab planning a multi-thousand GPU training cluster, site selection and colocation contracting should already be in motion. The lead time (deployment timeline from signed contract to operational capacity) for a purpose-built AI colo deployment in a constrained market like NoVa or the Chicago-to-Atlanta corridor is twelve to eighteen months. Waiting for GPU delivery before initiating the colo conversation is a sequencing error that will cost quarters.
If you are a Fortune 500 enterprise rolling out AI infrastructure for the first time, the instinct to default to AWS, Azure, or GCP (the hyperscalers, the largest cloud providers) for everything is understandable but increasingly expensive. Hyperscaler reserved instance pricing does not reflect the capital intensity of the underlying real estate and cooling infrastructure described above. Neocloud operators (specialized GPU cloud providers, an alternative to hyperscalers) running their own Tier III colocation footprints frequently offer 30 to 50 percent lower all-in pricing on H100 and H200 reserved capacity, with contract flexibility that hyperscalers do not match. The operators we work with are actively signing new reservations today.
If you are a scaleup ramping inference workloads across multiple regions, the distributed inference trend flagged by utility grid analysts is a structural opportunity. Inference workloads do not require the same geographic concentration as training. Placing inference capacity in secondary markets like Phoenix, Atlanta, or Dallas, where Tier III operators including Equinix, Digital Realty, CyrusOne, QTS, and Aligned have available inventory, can cut latency and power costs simultaneously.
The portfolio principle applies across all client types. The most resilient AI infrastructure strategies run a combination of hyperscaler capacity for burst and managed services, one or two neocloud operators for reserved GPU capacity at better economics, and direct Tier III colocation for workloads requiring dedicated infrastructure control. None of these in isolation is optimal.
How XIRR Advisors Can Help
XIRR Advisors brokers reserved GPU capacity from neocloud operators and Tier III colocation space across the US. We represent the client. The provider pays our fee. You pay nothing.
Share your requirements, including region, GPU type (H100, H200, B200, or GB200), capacity needed, timing, and for colocation, your megawatt (MW) requirement, and we will canvas the neocloud and colocation markets on your behalf and return a shortlist of qualified providers within 48 hours. Earlier conversations get materially better terms. Power constraints and site scarcity are not improving in the near term. Reach out at contact@xirradvisors.com or DM @XIRRAdvisors.
References
[1] Data Center Dynamics: Minneapolis city council approves six-month moratorium for data centers larger than 350,000 sq ft
[2] Data Center Knowledge: North Carolina SB 730 would force AI data centers into long-term power contracts
[3] Data Center Knowledge: Modine locks in $4B cooling capacity reservation through 2029
[4] Data Center Dynamics: TeraWulf taps Schneider in $290 million infrastructure deal
[5] Data Center Dynamics: Core42 secures $550 million financing to fuel AI build-out in Europe and US
[6] Data Center Knowledge: Nvidia earnings show AI spending moving beyond GPUs
[7] Data Center Knowledge: Texas CREZ grid corridors emerge as prime hyperscale AI infrastructure corridor
[8] Data Center Knowledge: Distributed AI inference could upend utility grid planning for hyperscalers
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.