The bottleneck for AI infrastructure has shifted. GPU lead times are compressing. Power access is not.
What Happened
Three data points published this week crystallize the structural turn. First, Data Center Knowledge's neocloud earnings roundup confirms that power constraints now outpace GPU procurement as the primary growth limiter for specialized GPU cloud providers (neoclouds, meaning GPU-focused cloud operators distinct from hyperscalers like AWS, Azure, and GCP). CoreWeave and Nebius both flagged power availability, not silicon supply, as the ceiling on near-term expansion. Second, Moody's revised its hyperscaler capex forecast upward by $85 billion, pushing the 2026 figure toward $785 billion and approaching $1 trillion by 2027. That volume of concurrent buildout is stress-testing every node of the power supply chain simultaneously. Third, xAI's deployment of 19 natural gas turbines at its Colossus 2 campus in Southaven, Mississippi, as reported by Data Center Dynamics, illustrates the logical endpoint of grid scarcity: frontier labs are building behind-the-meter generation (on-site power that bypasses the public utility grid entirely) rather than waiting in interconnection queues that stretch years.
Layered on top of this is a quieter but equally material risk: municipal politics. A city near Minneapolis just approved a one-year data center moratorium, joining a growing list of localities imposing zoning freezes on new builds. Meanwhile in Europe, former Microsoft executives have launched Krios Infrastructure specifically to source pre-permitted, grid-connected land for hyperscale campuses, a business that would not exist if shovel-ready sites were plentiful.
Why It Matters
The implication is structural, not cyclical. AI workloads are power-dense by nature. A rack of H100s or B200s can consume 30 to 100 kilowatts depending on configuration, versus 5 to 10 kilowatts for a typical enterprise server rack. When you aggregate those racks into clusters of thousands of GPUs, the facility itself becomes a small power plant customer. Grid operators are noticing. ERCOT, which manages Texas's grid, is already warning that AI power demand projections may be overstated, a signal that speculative load filings are crowding the interconnection queue and slowing approvals for projects with genuine near-term demand.
For Fortune 500 enterprises planning their first large AI infrastructure deployment, this reframes the site selection conversation. Choosing a colocation facility is no longer primarily about price per kilowatt or proximity to headquarters. The threshold question is: does this facility have contracted, deliverable power at the density my AI workloads require, and can the operator prove it? Tier III (meaning a data center reliability tier guaranteeing 99.982% uptime through redundant power and cooling paths) certification matters, but so does the operator's actual power contract structure, their relationship with the local utility, and whether they hold any behind-the-meter generation capacity.
For sovereign AI programs in the US and EU, the calculus is sharper still. National AI initiatives require predictable, long-term power access at scale. The European site scarcity that prompted Krios's launch reflects a real constraint: permitted, powered land in major EU markets is functionally sold out in many locations. Programs that wait for fully built-out hyperscaler regions will face multi-year queues. Those that engage colo operators with existing power allocations gain optionality that simply is not available at any price later.
There is also a resiliency angle that deserves attention. Uptime Institute data cited by Data Center Knowledge shows AI-dense facilities introducing new failure vectors that erode a decade of hard-won resiliency progress. High-density liquid cooling systems, power distribution at the rack level, and thermal management in GPU clusters create failure modes that standard SLAs (Service Level Agreements defining uptime guarantees) were not written to address. Clients sourcing colo space should audit SLA language specifically against AI workload risk profiles, not just legacy enterprise server assumptions.
What Clients Should Do
If you are a frontier AI lab or a well-capitalized scaleup planning a large GPU cluster, the xAI playbook is instructive even if you cannot replicate it. Prioritize facilities with demonstrated high-density power delivery, preferably operators who have already built out liquid cooling infrastructure rather than retrofitting air-cooled halls. Lock in capacity commitments now. The facilities with genuine power headroom in primary US markets (Northern Virginia, Dallas, Phoenix, Chicago, Silicon Valley) are contracting that headroom quickly.
If you are a Fortune 500 enterprise or a government program approaching AI infrastructure for the first time, the portfolio approach applies to real estate as directly as it does to GPU procurement. Do not build your entire strategy around a single hyperscaler's availability zone. Engage two or three Tier III colocation operators across different markets. Operators like Equinix, Digital Realty, CyrusOne, QTS, Aligned, Iron Mountain, Compass, Stack, Vantage, and Switch each carry different power profiles, different utility relationships, and different flexibility on contract terms. A broker who knows those distinctions can save you months of discovery time.
If you are a system integrator sourcing infrastructure for an end client, municipal moratorium risk is now a diligence item, not an afterthought. Secondary markets that looked attractive six months ago may have changed political posture on data center permitting. Verify zoning status before shortlisting any site.
In every case, the conversation needs to start earlier than most clients expect. Power allocations, MSAs (Master Service Agreements, the parent contracts governing facility access), and construction slots are reserved months before contracts execute. The clients securing the best terms in today's market are the ones who began conversations in Q4 2025.
How XIRR Advisors Can Help
XIRR Advisors brokers two things: reserved GPU capacity from neocloud operators (not hyperscalers, who sell direct) and Tier III colocation space across US markets. Many clients need both simultaneously. A GPU cluster needs a home, and a colo facility needs to be sized around the power and cooling demands of the specific hardware generation you are deploying.
Share your requirements with us: GPU type and count, region, timing, and for colocation, your target megawatt range. We canvas the neocloud and colo markets on your behalf and return a shortlist within 48 hours. Earlier conversations get better terms. Our fee is paid by the provider. Clients pay nothing. Reach us at contact@xirradvisors.com or DM @XIRRAdvisors.
References
[1] Data Center Knowledge: Neocloud earnings roundup, power constraints now outpace GPU supply
[2] Data Center Dynamics: Moody's raises hyperscaler capex forecast $85B toward $1T by 2027
[3] Data Center Dynamics: xAI deploys 19 natural gas turbines at Colossus 2 in Southaven, Mississippi
[4] Data Center Dynamics: One-year data center moratorium approved in city near Minneapolis
[5] Data Center Dynamics: Krios Infrastructure launches to source powered land for European hyperscalers
[6] Data Center Knowledge: ERCOT warns Texas AI power boom may not materialize
[7] Data Center Knowledge: AI boom threatens years of data center resiliency gains
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.