Power is now the binding constraint on AI infrastructure. Not chips. Not capital. Not engineering headcount. The organizations that internalize this first will secure capacity. Everyone else waits in a queue that is measured in years.
What Happened
Three signals this week crystallized the structural shift.
First, OpenAI disclosed it has secured 10GW of AI infrastructure capacity, with 3GW contracted in just 90 days. That is not a procurement event. That is a land grab. When a single organization reserves 10GW, it is not just buying compute. It is consuming power interconnects, grid commitments, and colocation real estate across multiple markets simultaneously.
Second, Microsoft's AI surge has exposed a critical capacity gap inside Azure. A $627 billion backlog does not clear overnight. Azure is constrained by power delivery timelines, cooling retrofit schedules, and physical construction cycles. Hyperscaler (the largest cloud providers including AWS, Azure, GCP, and Oracle) reservation queues for H200 and B200 capacity now stretch multiple quarters for most clients.
Third, Oracle's Project Jupiter has abandoned gas turbines in favor of Bloom Energy fuel cells at its New Mexico AI campus. Behind-the-meter (on-site power generation that bypasses the public grid) solutions are no longer a niche workaround. They are becoming standard practice because grid interconnection queues and utility approval timelines, now documented as the primary deployment ceiling by Data Center Knowledge's power analysis, are simply too slow for AI build schedules.
And the density problem compounds everything. Cooling infrastructure is now an AI deployment bottleneck at high-density GPU deployments. Older Tier III (data center reliability tier, 99.982% uptime) facilities that were perfectly adequate for traditional enterprise workloads are being rejected outright for modern AI cluster deployments because their cooling architecture cannot handle the per-rack density that H100 and B200 clusters demand.
Why It Matters
The mechanism here is straightforward: hyperscalers and frontier labs are consuming the inputs, specifically power commitments, colocation square footage, and cooling-ready white space, that everyone else needs.
Meta approaching $140 billion in 2026 capex (capital expenditure, infrastructure spending), restructuring its workforce to fund it, is not a curiosity. It is a signal that demand compression for everyone downstream is accelerating. When organizations of that scale are contracting power and colocation at speed, smaller clients get displaced from prime sites and preferred utility queues.
For Fortune 500 enterprises entering AI infrastructure for the first time, this creates a specific and underappreciated risk. The assumption that cloud capacity is infinitely elastic is wrong. Azure and AWS sell direct. They prioritize their largest committed spenders. A manufacturing firm or financial services company that defers its infrastructure decision by two quarters may find that the specific GPU configurations, regions, and contract terms it needed are no longer available at any price.
For sovereign AI programs in the US and EU building national compute reserves, the power constraint is existential. You cannot replicate a government-class AI cluster in a facility that was designed for web servers. Cooling readiness, power redundancy, and physical security requirements all converge on a short list of qualified Tier III operators: Equinix, Digital Realty, CyrusOne, QTS, Aligned, Iron Mountain, Compass, Stack, Vantage, and Switch. That list does not grow quickly. Sites in Northern Virginia (NoVa, the largest US data center market), Dallas, Phoenix, and Chicago that meet the spec are being absorbed faster than new supply is being commissioned.
For scaleups and AI application companies productionizing inference workloads, the neocloud (specialized GPU cloud providers, an alternative to hyperscalers) market remains the most practical path. The GPU cloud operators we work with are typically 30 to 50 percent cheaper than hyperscaler reserved instances, can deploy capacity in weeks rather than quarters, and offer more flexible contract terms for teams whose capacity requirements are still evolving.
What Clients Should Do
If you are a frontier lab or large enterprise planning a multi-thousand-GPU training cluster, the window to negotiate favorable colocation terms with operators like Equinix or Digital Realty is now. Cooling-ready, high-density white space with sufficient power capacity is the scarcest input in this market. Locking a Master Service Agreement (MSA, the parent contract) with a qualified Tier III operator before your deployment timeline forces the issue gives you negotiating leverage you will not have later.
If you are a Fortune 500 organization rolling out AI infrastructure for the first time, do not default to hyperscalers as your only option. Run a structured evaluation: hyperscaler reserved capacity versus neocloud operators versus owned colocation. A portfolio approach, anchored on one or two neocloud operators for GPU capacity plus a Tier III colocation relationship for dedicated infrastructure, typically delivers better economics and faster ramp time than a single-vendor hyperscaler commitment.
If you are a scaleup ramping inference capacity, prioritize contract flexibility over headline pricing. The neocloud operators with the most competitive terms right now are the ones absorbing capacity before hyperscaler queues clear. Get into discovery conversations immediately. Terms available today will not survive another quarter of demand compression.
For any client evaluating colocation, treat cooling readiness as a first-order site-selection criterion, not a facilities afterthought. Ask operators specifically about their liquid cooling infrastructure (direct liquid cooling that delivers coolant to the rack or chip level) and per-rack power density limits before signing anything.
XIRR Advisors runs custom-sourcing engagements for reserved GPU capacity from neocloud operators and Tier III colocation space across the US. We canvas the market on your behalf and return a shortlist within 48 hours of receiving your requirements. Share your region, GPU type, capacity volume, and timing, or megawatts for colocation, and we get to work. Earlier conversations get better terms. The fee is paid by the provider. Clients pay nothing.
Contact us at contact@xirradvisors.com or DM @XIRRAdvisors.
References
[1] Data Center Dynamics: OpenAI claims to have secured 10GW of AI infrastructure capacity ahead of 2029 target
[2] Data Center Knowledge: Microsoft AI surge exposes data center capacity gap
[3] Data Center Knowledge: Oracle's Project Jupiter ditches gas turbines for Bloom fuel cells
[4] Data Center Knowledge: The breaking points: power emerges as AI's defining limit
[5] Data Center Knowledge: The breaking points: cooling struggles to keep pace with AI density
[6] Tom's Hardware Pro: Mark Zuckerberg says Meta is cutting 8,000 jobs to pay for AI infrastructure
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.