The GPU capacity market just reshuffled again, and clients who locked procurement strategies six months ago are already working from stale assumptions.
What Happened
Three stories published this week tell a coherent and urgent story about where compute infrastructure is heading.
First, xAI is now actively soliciting third-party AI compute customers after signing a landmark deal with Anthropic, per Data Center Dynamics. The same entity that competes with OpenAI at the model layer is now selling GPU capacity to Anthropic, one of OpenAI's primary rivals. That is not a niche footnote. It signals that vertically integrated AI infrastructure players are becoming serious neocloud operators (specialized GPU cloud providers, an alternative to hyperscalers such as AWS, Azure, GCP, and Oracle), and that the capacity market is fragmenting faster than most procurement teams have absorbed.
Second, Data Center Knowledge reports that Texas has overtaken Northern Virginia (NoVa, the largest US data center market) in global data center rankings, driven by AI infrastructure investment. Dallas and West Texas are absorbing demand that Virginia simply cannot accommodate given its power and land constraints. This reshapes site selection for colocation clients across every segment, from sovereign AI programs building national compute reserves to Fortune 500 enterprises standing up their first GPU clusters.
Third, Nvidia's latest earnings reveal that AI spending is accelerating into networking and optics, not just raw GPU silicon, per Data Center Knowledge. Simultaneously, a separate analysis highlights that HBM (High-Bandwidth Memory, the memory architecture used in modern GPUs) and CXL (Compute Express Link, a high-speed interconnect standard linking CPUs and accelerators) are now the binding constraints in large-scale cluster design. You can reserve all the H200 or B200 nodes you want. If the interconnect fabric and memory subsystem are not co-planned, cluster utilization collapses.
Why It Matters
The structural pattern here is compression on multiple fronts at once: geography, supply chain, and architecture.
On geography, Virginia's saturation is not a temporary dip. It is a structural ceiling. Clients who anchored colocation strategies in NoVa are now competing for a shrinking pool of available capacity. Texas markets, Phoenix, and Chicago are the pressure valves, and they are filling fast. Meanwhile, inference workloads are pulling GPU deployments back into metro colocation facilities, because latency-sensitive production AI applications cannot tolerate the round-trip times of exurban hyperscale campuses. That dynamic is reopening urban colo markets that many assumed AI had permanently sidelined.
On supply chain, extended hardware lead times are compressing procurement windows across the board. Frontier labs with multi-thousand-GPU training requirements and Fortune 500 enterprises deploying their first AI infrastructure face the same arithmetic: the time between "we need capacity" and "capacity is live" has lengthened, while the competitive cost of delay has shortened. Clients who initiate conversations six to nine months ahead of need consistently secure better pricing and contract terms than those who arrive with urgent timelines.
On architecture, the shift in Nvidia revenue toward networking is a warning signal. H100, H200, B200, and GB200 procurement decisions made without a parallel plan for the interconnect fabric and memory hierarchy are likely to underdeliver on projected utilization. Neocloud operators who have pre-built dense, high-bandwidth fabrics offer a structural advantage here over DIY colocation builds for teams without dedicated infrastructure engineering.
The xAI development adds another layer. A new, well-capitalized neocloud entrant with its own energy and compute vertical entering the third-party market increases supply and competitive pricing pressure. That is good for clients. But it also means the provider landscape is shifting, and brokers who do not track these entrants in real time leave pricing leverage on the table.
What Clients Should Do
If you are a frontier lab or large-scale AI application company planning a training cluster above 4,000 GPUs, the architecture conversation must happen before the procurement conversation. Locking H200 or B200 reservations without specifying the interconnect layer and memory configuration is a budget risk, not just a performance risk. Neocloud operators with pre-integrated cluster designs can shortcut this significantly.
If you are a Fortune 500 enterprise or sovereign AI program evaluating your first dedicated GPU infrastructure, Texas is now a primary site-selection market, not a fallback. Dallas, in particular, offers available Tier III (data center reliability tier, targeting 99.982% uptime) capacity from operators including CyrusOne, QTS, and Aligned at competitive power rates. NoVa remains viable but commands a premium and carries availability risk.
If you are a scaleup ramping inference for a production application, the metro colocation thesis deserves serious evaluation. Latency to end users matters. Hyperscaler on-demand pricing for inference workloads is expensive at scale. Neocloud operators with metro GPU capacity, combined with a Tier III colocation anchor, often deliver a significantly lower total cost at production volumes, frequently 30 to 50 percent below comparable hyperscaler reserved instances.
Across all client types, the most important tactical shift is lead time. Capacity conversations that begin today close at better terms than those that begin under deadline pressure in Q3.
Work With XIRR Advisors
XIRR Advisors brokers reserved GPU capacity from neocloud operators and Tier III colocation space across the USA and EU. We represent clients exclusively. Providers pay our fee. You pay nothing.
Share your requirements, including region, GPU type (H100, H200, B200, GB200, GB300), cluster size, and timing, or megawatts if you are evaluating colocation. We will canvas the neocloud and colocation markets on your behalf and return a shortlist within 48 hours. The earlier the conversation, the better the terms we can secure. Email contact@xirradvisors.com or DM @XIRRAdvisors.
References
[1] Data Center Dynamics: xAI is actively seeking more AI compute customers following Anthropic deal
[2] Data Center Knowledge: Texas Powers Past Virginia in Global Data Center Rankings
[3] Data Center Knowledge: Nvidia Earnings Show AI Spending Moving Beyond GPUs
[4] Data Center Knowledge: Scaling the Memory Wall: HBM, CXL, and the New GPU Playbook
[5] Data Center Knowledge: AI Inference Pulls Infrastructure Back Into Metro Data Centers
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.