The AI infrastructure market has hit two ceilings simultaneously, and most clients are only planning around one of them.
What Happened
A new report from the Center for a New American Security (CNAS), covered by Data Center Knowledge, identifies silicon and HBM (High-Bandwidth Memory, the memory architecture inside modern GPUs) scarcity as a structural constraint now co-equal with power in blocking hyperscale AI expansion. This is not a temporary supply blip. CNAS frames it as a durable bottleneck, one that compounds rather than replaces the grid access problem that has dominated data center headlines for two years.
On the GPU economics side, The Next Platform reports that accelerating compute and memory price inflation is forcing infrastructure clients to reforecast capex (capital expenditure, infrastructure spending) plans upward, with spending trajectories outpacing prior projections by meaningful margins. Budgets set twelve months ago are now structurally underfunded.
Meanwhile, the power problem is not resolving. It is evolving. Data Center Knowledge documents a 200,000 square foot Texas AI campus that abandoned grid interconnection entirely after a multi-year wait and moved to self-generated, behind-the-meter (on-site power generation that bypasses the public grid) power as its primary strategy. Separately, VoltaGrid raised $1 billion from Blackstone and Halliburton to scale on-site power systems, per Data Center Dynamics. Institutional capital is now explicitly betting that behind-the-meter is not a workaround. It is the primary model.
AWS, meanwhile, converted a leased Northern Virginia asset to owned freehold for $65 million, consistent with a broader hyperscaler pattern of locking in long-term control over critical compute real estate in NoVa (Northern Virginia, the largest US data center market). Microsoft has committed to doubling its AI infrastructure within two years, per The Next Platform. These are not modest bets.
Why It Matters
The mechanism here is a reinforcing loop, not a linear problem. Hyperscalers (the largest cloud providers: AWS, Azure, GCP, and Oracle) are absorbing available GPU supply, power capacity, and land simultaneously. Their scale creates a gravity well. Nvidia's commitment to IREN's 5 GW pipeline in Texas illustrates the next phase: GPU supply is increasingly pre-committed to vertically integrated compute campuses before it ever reaches the open market.
For clients who are not themselves hyperscalers, this creates a timing trap. Every quarter spent in procurement evaluation is a quarter of eroding optionality. H200 and B200 reserved capacity on AWS and Azure now carries wait times measured in quarters, not weeks. Price inflation compounds on top of that delay.
This is precisely where the neocloud (specialized GPU cloud providers, purpose-built alternatives to hyperscalers) market provides structural relief. Neocloud operators we work with are sourcing GPU capacity on shorter lead times, typically weeks versus the quarters required for hyperscaler reserved instances, at 30 to 50 percent lower pricing on comparable H100, H200, and B200 configurations. Their contract terms are also more flexible, a critical variable for sovereign AI programs and Fortune 500 enterprises still stress-testing their actual workload requirements before committing to multi-year hyperscaler contracts.
The CNAS silicon scarcity finding adds urgency to a specific dynamic: neocloud operators that secured GPU inventory earlier in the procurement cycle are sitting on capacity that hyperscalers cannot immediately replicate. That inventory gap is real and time-limited.
What Clients Should Do
If you are a frontier lab planning a 10,000-GPU training cluster, the question is not whether to use hyperscalers. It is how much of your cluster to route through neocloud operators to reduce blended cost and accelerate ramp (deployment timeline for capacity coming online). A portfolio approach, anchoring on one or two neocloud operators for reserved GPU capacity alongside a hyperscaler relationship for burst and managed services, is the optimal structure in the current market.
If you are a Fortune 500 enterprise rolling out AI infrastructure for the first time, avoid the reflex of defaulting entirely to your existing AWS or Azure relationship. The pricing differential on reserved GPU instances is large enough to materially affect your program's total cost of ownership over a three-year horizon. Neocloud alternatives should be evaluated in parallel before any commitment is made.
If you are a sovereign AI program or a government-adjacent initiative sourcing dedicated compute, colocation matters as much as the GPU layer. Tier III (a data center reliability classification representing 99.982% uptime) colocation operators including Equinix, Digital Realty, QTS, CyrusOne, and Aligned have capacity across NoVa, Dallas, Phoenix, Chicago, and Atlanta that can serve as the physical anchor for a GPU cluster procured through neocloud operators. The combination of owned or leased colo space plus reserved GPU contracts gives sovereign programs the control and auditability that hyperscaler shared environments cannot provide.
For AI scaleups ramping inference workloads, network fabric is now a first-order variable. The neocloud operators we work with are investing in high-bandwidth interconnect (the physical and logical links connecting GPU nodes within a cluster) specifically to meet inference latency requirements. This is not a secondary consideration. It belongs in your RFP.
Start conversations earlier than feels necessary. Capacity terms available today will not be available in sixty days.
XIRR Advisors
XIRR Advisors brokers reserved GPU capacity from neocloud operators and Tier III colocation space across the US. We do not broker hyperscalers. AWS, Azure, GCP, and Oracle sell direct. Our value is in the markets they cannot access for you.
Share your requirements: region, GPU type (H100, H200, B200, GB200, GB300), cluster size, timing, and MW requirements for colocation. We will canvas the neocloud and colocation markets on your behalf and return a shortlist within 48 hours. The service is free to clients. Providers pay our fee. Earlier conversations consistently produce better terms. Contact us at contact@xirradvisors.com or DM @XIRRAdvisors.
References
[1] Data Center Dynamics: VoltaGrid Raises $1B from Blackstone and Halliburton to Expand Power System Offering for Data Centers
[2] Data Center Knowledge: Unconventional Texas Data Center Explores Off-Grid Power
[3] Data Center Knowledge: After the Power Crunch, AI Infrastructure Hits a GPU Wall
[4] The Next Platform: Compute and Memory Price Hikes Drive IT Spending Way Higher
[5] Data Center Dynamics: Amazon Acquires Leased Data Center in Virginia for $65 Million
[6] The Next Platform: Microsoft Committed to Doubling AI Infrastructure in Two Years
[7] Data Center Knowledge: Nvidia Places Massive AI Infrastructure Bet on IREN's 5 GW Pipeline
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.