The bottleneck in AI infrastructure has quietly migrated. It is no longer the GPU waitlist. It is the power connection.
What Happened
Three signals this week crystallize the shift. First, Data Center Knowledge reports that distribution complexity and interconnection queue depth, not raw generation capacity, have become the primary gating factor for new AI deployments. Utilities simply cannot get electrons from the substation to the raised floor fast enough to match demand. Second, Core Scientific secured 300MW of capacity at its Pecos, Texas site and is actively pursuing behind-the-meter generation to sidestep grid interconnection queues entirely. Behind-the-meter means on-site power generation, typically gas turbines or fuel cells, that bypasses the public utility grid. Third, a similar fuel-cell strategy is playing out in New Mexico, where LS Electric is supplying power distribution systems for a Bloom Energy and Oracle project, Project Jupiter, that pairs solid-oxide fuel cells with colocation infrastructure precisely to avoid constrained grid tie-in timelines.
On the regulatory front, the landscape is getting choppy. Wisconsin's Public Service Commission just tightened hyperscale tariff rules, setting a precedent that state regulators are willing to ring-fence large-load customers from shifting costs onto residential ratepayers. Maine briefly considered a data center moratorium before the governor vetoed it, but legislative pressure continues. Both states signal that Northeast and Midwest site-selection risk models need a regulatory variable that did not exist two years ago.
Why It Matters
The power constraint is a structural realignment, not a temporary pinch. Interconnection queues at major US utilities run 3 to 5 years in dense markets like Northern Virginia (NoVa, the largest US data center market), Northern Illinois, and the Phoenix metro. That timeline is incompatible with AI program cycles, which are measured in quarters, not years.
This is why the behind-the-meter trend matters so much. Operators willing to build or contract for on-site generation, via a PPA (Power Purchase Agreement, a long-term electricity contract) with a fuel-cell provider or a natural gas installation, can compress that 3-to-5-year queue to 12 to 18 months. The Pecos and New Mexico deals are not outliers. They are the template.
For hyperscalers, this creates a competitive moat. AWS, Azure, GCP, and Oracle Cloud Infrastructure have the balance sheets and land positions to execute behind-the-meter strategies at scale. They are also bifurcating their internal network fabrics for AI workloads. Google's Virgo network architecture, detailed this week by The Next Platform, represents a dedicated inference and training fabric that leaves conventional spine-leaf network topology (the standard hierarchical data center switching architecture) behind. This is purpose-built infrastructure that a shared colocation tenant cannot replicate out of the box.
But here is the counterintuitive implication. Google's custom silicon path, including the TPU 8 (Tensor Processing Unit, Google's custom AI chip), is diverging sharply from merchant GPU architectures. That divergence is actually good news for clients who need H100, H200, B200, or GB200 capacity. Google is not competing with NVIDIA GPU cloud operators for that segment. The neocloud operators (specialized GPU cloud providers, an alternative to hyperscalers) that XIRR works with are filling exactly this gap, with H200 and B200 clusters available in weeks, not quarters, at 30 to 50 percent lower cost than comparable hyperscaler reserved instances.
What Clients Should Do
If you are a Fortune 500 enterprise standing up AI infrastructure for the first time, power should be the first question in your site-selection process, not an afterthought. Ask your prospective colocation operator, whether Equinix, Digital Realty, CyrusOne, QTS, Aligned, Iron Mountain, or anyone else, for the committed power delivery date, not the contracted capacity. Those two numbers are often different. If the committed date is more than 12 months out, you need a parallel track.
If you are a frontier AI lab planning a large training cluster, the choice is not hyperscaler versus colo. It is a portfolio. Hyperscalers give you flexibility and managed networking. Dedicated neocloud GPU capacity gives you cost efficiency and faster ramp. A colocation deployment in Dallas, Phoenix, or Chicago, with either behind-the-meter power or a firm utility SLA (Service Level Agreement, the contract defining uptime guarantees), gives you control. Run all three.
If you are a sovereign AI program in the US or EU procuring capacity for a national initiative, the Wisconsin and Maine regulatory signals are directly relevant. State-level utility policy is now a site-selection variable with real teeth. Build regulatory risk scoring into your colocation RFP (Request for Proposal) process. NoVa is constrained. The Southeast and Texas markets are moving faster, with better power economics today.
For scaleups ramping inference workloads, the neocloud market has meaningfully more available capacity right now than hyperscaler order books suggest. Lead times (the deployment timeline for capacity coming online) of 4 to 8 weeks are achievable for 512-GPU to 2,048-GPU inference deployments. That is the conversation to have before you default to the AWS or Azure waitlist.
XIRR Advisors: How We Help
XIRR Advisors brokers reserved GPU capacity from neocloud operators across the US and EU, and Tier III (99.982% uptime) colocation space in every major US market. We represent clients exclusively. The provider pays our fee. You pay nothing.
If you need GPU capacity, a colocation deployment, or both, share your requirements: region, GPU type, cluster size, timing, and MW target for colocation. We will canvas the neocloud and colocation markets on your behalf and return a shortlist within 48 hours. Earlier conversations get materially better terms, especially as power-constrained markets tighten further through 2026. Reach us at contact@xirradvisors.com or DM @XIRRAdvisors.
References
[1] Data Center Knowledge: Power Emerges as AI's Defining Limit
[2] Data Center Dynamics: Core Scientific Secures 300MW at Pecos, Eyes Behind-the-Meter Power
[3] Data Center Dynamics: LS Electric to Supply Power Distribution for Bloom Energy, Oracle New Mexico Project
[4] Data Center Knowledge: Wisconsin PSC Tightens Data Center Tariff Rules
[5] Data Center Knowledge: Maine Vetoes Data Center Moratorium but Pressure Continues
[6] The Next Platform: Google's New Networks Tuned for GenAI Inference and Training
[7] The Next Platform: With TPU 8, Google Makes GenAI Systems Much Better, Not Just Bigger
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.