The GPU is no longer the only line item that matters.
What Happened
Nvidia's latest earnings revealed something procurement teams should internalize immediately: AI infrastructure spending is expanding well beyond GPU clusters into networking and optical interconnect. Per Data Center Knowledge, explosive growth in those segments signals that clients budgeting only for compute are systematically underestimating total buildout costs. The interconnect (the high-speed fabric linking GPUs within and across servers) is no longer a rounding error.
Meanwhile, the geography of where AI infrastructure lands is shifting fast. Texas has overtaken Northern Virginia (NoVa) in global data center rankings, driven by power availability and land access, two inputs that NoVa has increasingly struggled to supply. Reinforcing that regional shift: DC Blox just added $600 million in debt financing to accelerate colocation (shared data center space leased to multiple tenants) expansion across the Southeast US, introducing meaningful new supply in markets like Atlanta and beyond. For clients that have been competing for constrained Virginia capacity, these are actionable alternatives.
On the inference (running deployed AI models in production) side, Data Center Knowledge reports that latency-sensitive workloads are pulling GPU deployments back into metro-area colocation facilities, reversing the campus-only logic that dominated training infrastructure planning. Tier III (99.982% uptime standard) urban colos in cities like Dallas, Chicago, and Atlanta are seeing renewed demand from exactly this use case.
Why It Matters
The structural pattern across these stories is the same: full-stack infrastructure cost is rising faster than GPU prices are falling, and the clients who optimize only on GPU price-per-hour are walking into budget overruns.
Hyperscalers (the largest cloud providers, AWS, Azure, GCP, Oracle) bundle compute, networking, and storage into a single invoice, which obscures how much of that spend is going where. That opacity is part of their business model. Neocloud operators (specialized GPU cloud providers, an alternative to hyperscalers) typically separate those layers, making cost attribution cleaner and giving clients more negotiating leverage on each component. The gap in all-in cost between hyperscalers and neoclouds is frequently 30 to 50 percent on the compute layer alone, before networking configuration is even optimized.
For frontier labs planning large training runs, the networking cost revelation from Nvidia's earnings is particularly relevant. A 10,000-GPU cluster at hyperscaler list pricing, including interconnect provisioning, can run materially higher than a comparable neocloud deployment where the client negotiates the fabric configuration directly. For Fortune 500 enterprises entering AI infrastructure for the first time, the lesson is different: don't anchor your total cost of ownership (TCO) model on GPU unit pricing alone. Power, cooling, and interconnect collectively represent a significant share of operating costs, and advances in power electronics are beginning to alter how efficiently dense GPU deployments consume watts.
For sovereign AI programs, the geographic diversification story matters enormously. Concentrating national AI infrastructure in a single constrained market creates both cost and resilience risk. Texas, the Southeast, and select EU markets offer credible alternatives.
What Clients Should Do
If you are a frontier lab or large-scale AI application company planning a training cluster, benchmark the full stack, not just GPU pricing. Ask any provider to separate GPU, networking, storage, and power costs. Then compare that itemized quote against neocloud operators who can often deliver equivalent configurations faster (weeks versus quarters on hyperscaler wait lists) and at lower total cost. The neocloud market has matured enough that the largest operators now support the networking fabrics and SLAs (Service Level Agreements, which define uptime guarantees) that serious training workloads require.
If you are a Fortune 500 enterprise or systems integrator deploying inference infrastructure for production workloads, the metro colocation story is directly relevant. Urban Tier III colos from operators like Equinix, Digital Realty, CyrusOne, and QTS in markets including Dallas, Chicago, Atlanta, and the New York/New Jersey corridor can now support GPU-dense racks with modern power and cooling specs. Pairing colocation space with reserved neocloud GPU capacity is an increasingly common portfolio approach among sophisticated clients.
If you are a scaleup ramping inference, do not wait until you are capacity-constrained to start conversations. The supply chain turbulence documented by The Next Platform is real. Extended lead times mean the clients who begin procurement discussions now are the ones who secure favorable pricing and earlier delivery windows.
The smart portfolio is a combination: hyperscalers for certain workloads where their ecosystems justify the premium, neocloud operators for reserved GPU capacity where price and speed matter, and owned or leased colocation for workloads requiring physical control. Most clients are underweight on the latter two.
XIRR Advisors sources reserved GPU capacity from neocloud operators and Tier III colocation space across the USA on behalf of clients. Share your requirements, including region, GPU type (H100, H200, B200, GB200, GB300), cluster size, timing, or megawatt needs for colocation, and we will canvas the neocloud and colocation markets and return a shortlist within 48 hours. Earlier conversations reliably produce 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: Nvidia Earnings Show AI Spending Moving Beyond GPUs
[2] Data Center Knowledge: Texas Powers Past Virginia in Global Data Center Rankings
[3] Data Center Dynamics: DC Blox Adds $600M to Existing Debt Facility
[4] Data Center Knowledge: AI Inference Pulls Infrastructure Back Into Metro Data Centers
[5] Data Center Knowledge: How Power Electronics Cut Generator Run Hours in AI-Scale Data Centers
[6] The Next Platform / The Register: Supply Chain Turbulence Forces New Playbook on Infrastructure Teams
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.