The constraint that will define AI infrastructure in 2026 is not silicon. It is power.
What Happened
Three signals landed simultaneously this week, and together they tell a coherent story about where the data center market is heading.
First, AWS's US-EAST-1 region in Northern Virginia (NoVa, the largest data center market in the world) suffered a power-induced thermal outage that disrupted the anchor US cloud region. This is not a one-off. NoVa is the most capacity-dense corridor on earth, and the grid there is operating with shrinking headroom. When the most battle-hardened hyperscaler (hyperscaler: the largest cloud providers, AWS, Azure, GCP, Oracle) goes down due to power stress, the lesson is structural, not operational.
Second, utilities are warning regulators that unpredictable AI workload power swings are breaking conventional grid modeling. AI training clusters do not draw power like traditional enterprise workloads. They surge, throttle, and surge again, creating instability that distribution engineers did not design for. This is forcing grid operators to rethink interconnection queues and siting approvals, adding months or years to large campus timelines.
Third, 69 US jurisdictions have now enacted bans or moratoriums on new data center construction, four of them permanent. Greenfield site selection is materially harder than it was 18 months ago. The jurisdictions that remain open are pricing that scarcity into land, permitting fees, and power agreements. Meanwhile, the Danish grid operator just imposed a three-month moratorium on new connections driven by data center demand, a reminder that this is not exclusively a US problem. EU clients planning sovereign AI programs or regional inference deployments face the same dynamic in key Nordic and Western European markets.
Layered on top of all this: Microsoft has committed to doubling its AI infrastructure footprint within two years. That demand wave has to land somewhere. When the largest infrastructure buyer on earth is 2x-ing, every available megawatt in constrained markets gets absorbed faster.
Why It Matters
The AWS outage is the clearest illustration of a core problem with defaulting to hyperscalers for AI infrastructure: you inherit their risk profile, their geography, and their grid exposure. US-EAST-1 is the most relied-upon cloud region in the world, which is precisely why its power vulnerabilities are so consequential. Concentration risk is real.
For frontier labs running continuous training at scale, an outage is not a billing inconvenience. It is a training run interrupted, with cascading costs that dwarf the instance charges. For Fortune 500 enterprises standing up production AI pipelines in financial services or pharma, a regional cloud outage under a shared SLA (Service Level Agreement, defining uptime guarantees) is a compliance event, not just an ops ticket.
The alternative worth examining is purpose-built Tier III (data center reliability tier, delivering 99.982% uptime with fully redundant power and cooling paths) colocation, combined with reserved GPU capacity from neocloud operators (specialized GPU cloud providers, an alternative to hyperscalers). Colo operators like Equinix, Digital Realty, CyrusOne, QTS, Aligned, and Compass hold their own utility relationships, operate behind-the-meter (on-site power generation that bypasses the public grid) generation in many campuses, and spread exposure across multiple feeds. Their power infrastructure is not shared with the workload demand spikes of a hundred other tenants running AI training simultaneously.
On the GPU supply side, neocloud operators are capitalizing on exactly this moment. They are acquiring capacity aggressively, often at 30 to 50% lower cost than hyperscaler reserved instances, with ramp times (deployment timelines for capacity coming online) measured in weeks rather than the quarters-long wait lists that H200 and B200 reservations through AWS, Azure, or GCP currently carry.
What Clients Should Do
If you are a frontier lab planning a 10,000-GPU training cluster, the risk calculus has shifted. Parking that cluster entirely in a single hyperscaler region, especially a power-stressed one like NoVa, is a single point of failure at a scale you cannot afford. The right architecture is a portfolio: one or two neocloud operators for the bulk of reserved GPU capacity, a Tier III colo footprint for bare-metal workloads where you control the stack, and hyperscaler burst capacity as a secondary layer. Lock in neocloud reservations now. Availability for H200 and B200 clusters is tightening as operators place their own supply bets.
If you are a Fortune 500 enterprise in financial services, pharma, or manufacturing rolling out AI infrastructure for the first time, do not default to hyperscaler dedicated instances because it feels familiar. The pricing gap is too large and the lead times too long. A Tier III colo facility in Dallas, Phoenix, Atlanta, or Chicago with a direct connection to a neocloud GPU pool often delivers better economics, better SLAs, and faster deployment than a hyperscaler private cloud buildout. Start with a colocation shortlist and a GPU reservation in parallel. They are not mutually exclusive and often reinforce each other.
If you are a sovereign AI program or a government-adjacent initiative sourcing EU capacity, the Danish moratorium is an early warning. Nordic power advantages are real, but grid access is tightening. Western European colo markets in Frankfurt, Amsterdam, and Paris still have viable capacity windows, but they are closing. Start site selection conversations immediately, because permitting and power procurement timelines in the EU are long by default, and political risk is compressing them further.
For system integrators and consultancies sourcing on behalf of enterprise clients, the most common mistake right now is serializing the process: find the GPU, then find the facility. Run them in parallel. GPU operators can often place capacity into client-designated colo facilities, and the best terms come when both sides of the deal are structured together.
The broader pattern: power scarcity, local bans, and grid instability are permanently re-rating location quality for AI infrastructure. Markets that have available power, permitting runway, and fiber diversity are worth a premium. Markets that lack any one of those three should be avoided, regardless of how familiar they feel.
---
XIRR Advisors sources reserved GPU capacity from neocloud operators and Tier III colocation space across the USA and EU. We represent the client exclusively. Providers pay our fee. Share your requirements, region, GPU type, cluster size, timing, or megawatts for colocation, and we will canvas the neocloud and colocation markets on your behalf and return a shortlist within 48 hours. Earlier conversations consistently secure better pricing and faster ramp. Reach us at contact@xirradvisors.com or DM @XIRRAdvisors.
References
[1] Data Center Dynamics: AWS experiences power issues at Northern Virginia cloud region
[2] Data Center Knowledge: From capacity to chaos: how AI data centers challenge the grid
[3] Tom's Hardware Pro: AI data center bans are rapidly multiplying across the US, 69 jurisdictions block new builds with four moves noted as permanent
[4] Data Center Dynamics: Danish grid operator introduces three-month moratorium for new grid connections
[5] The Next Platform: Microsoft committed to doubling AI infrastructure in two years
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.