The GPU procurement market is being reshaped by three simultaneous forces: memory costs exploding, new compute suppliers entering the market, and geography shifting in ways that change where capacity should physically land.
What Happened
Start with the hardware economics. Per Tom's Hardware Pro, Nvidia's Rubin AI systems now price out at $7.8 million to build, with HBM (High-Bandwidth Memory, the memory architecture used in modern AI GPUs) costs up 485% and memory now comprising 25% of total system cost. Separately, Data Center Knowledge reports that HBM constraints and CXL (Compute Express Link, a high-speed interconnect standard for pooling memory across servers) bottlenecks are now a primary architectural risk in infrastructure design. These are not independent data points. They describe the same structural problem: the GPU itself is no longer the primary cost or constraint. The memory subsystem is.
On the supply side, the competitive map is expanding in a meaningful way. Data Center Dynamics reports that xAI is actively soliciting third-party AI compute clients following its Anthropic deal, positioning the Colossus cluster in Memphis as an external GPU cloud. The SpaceX IPO filing reinforces this: the company is explicitly repositioning itself as a vertically integrated AI infrastructure platform, spanning compute, networking, energy, and orbital connectivity. Whether or not xAI/SpaceX matures into a reliable multi-tenant operator is an open question. But the directional signal is clear: the universe of GPU capacity suppliers is expanding beyond the hyperscalers (the largest cloud providers, AWS, Azure, GCP, and Oracle) and beyond the established neocloud operators (specialized GPU cloud providers that compete on price and flexibility).
Meanwhile, geography is shifting. Data Center Knowledge's analysis shows Texas overtaking Virginia in global data center market rankings, driven by available power, land, and AI-specific demand. At the same time, a separate report documents inference workloads pulling GPU capacity back into metro colocation (data center space leased in or near major cities) facilities, reversing the trend toward remote hyperscale campuses. These two trends pull in different directions and both are real. Training clusters want cheap power and land in Texas. Production inference wants low latency and metro proximity in Chicago, Dallas, Frankfurt, or Amsterdam.
Why It Matters
The memory cost escalation changes the procurement calculus for everyone. Frontier labs planning next-generation training runs cannot treat GPU count as the only variable. A cluster optimized on raw H200 or B200 node count, without accounting for HBM bandwidth per GPU and memory pooling architecture, will underperform on memory-bound workloads and cost more than modeled. Nvidia's own earnings confirm this: explosive networking and optics revenue growth means the full system cost, including interconnect fabric, is now meaningfully higher than the GPU line item alone suggests.
For Fortune 500 enterprises evaluating AI infrastructure for the first time, this complexity is a reason to avoid locking into a single provider or a single architecture. The hyperscaler default feels safe, but H200 and B200 wait lists stretch quarters, pricing reflects that scarcity, and the hyperscalers have no particular incentive to help clients optimize memory architecture.
Neocloud operators, including those we work with, have been faster to provision latest-generation capacity, often 30-50% cheaper on reserved compute, and more willing to customize deployment configurations. With supply constraints and memory costs both rising, the delta between hyperscaler list pricing and neocloud reserved pricing is widening, not narrowing.
Sovereign AI programs in the US and EU face a version of this problem at scale. Large public-sector GPU commitments made without geographic optionality or memory architecture review are now potentially misaligned with where power and latency economics actually favor deployment.
What Clients Should Do
If you are a frontier lab planning a 10,000-GPU-plus training cluster, the key questions are no longer just GPU type and count. What is the HBM configuration per node? What is the interconnect fabric (the high-speed network linking GPUs within a cluster)? Is the colocation site in a power market, such as West Texas or the Midwest, where you have negotiating leverage on a PPA (Power Purchase Agreement, a long-term electricity supply contract)? These questions should drive provider selection, not the other way around.
If you are a scaleup or AI application company ramping inference in production, metro colocation is back on the table. Facilities operated by Equinix, Digital Realty, and CyrusOne in Dallas, Chicago, Frankfurt, and Amsterdam offer the latency profile that hyperscale campuses in remote markets cannot. Pairing metro colo with reserved GPU capacity from a neocloud operator, rather than running inference entirely on a hyperscaler, typically yields material cost savings at scale.
If you are a Fortune 500 enterprise or system integrator sourcing for a client, the right structure in 2026 is a portfolio: hyperscaler for flexibility and integration simplicity on lower-priority workloads, one or two neocloud operators for reserved GPU capacity on cost-sensitive training and inference, and Tier III (data center reliability tier, defined by 99.982% uptime) colocation for any on-premises or co-located deployments. Single-vendor concentration is a procurement risk, not a simplification.
Across all client types, the clients getting the best terms right now are the ones having conversations earliest. Reserved GPU capacity from the neocloud operators we work with is still available for H100, H200, and B200 deployments, but window size and pricing flexibility are both shrinking as the year progresses.
XIRR Advisors brokers reserved GPU capacity from neocloud operators and Tier III colocation space across the USA. We do not broker hyperscalers. AWS, Azure, GCP, and Oracle sell direct. Our value is in the neocloud and colo markets, where pricing is negotiable, terms are flexible, and the providers pay our fee. Clients pay nothing.
Share your requirements, region, GPU type, capacity size, timing, or megawatts for colocation, and we will canvas the market and return a shortlist within 48 hours. Many clients need both GPU capacity and physical colocation space. We handle both. Reach us at contact@xirradvisors.com or DM @XIRRAdvisors. Earlier conversations get better terms.
References
[1] Tom's Hardware Pro: Nvidia's Memory Costs Soar 485%; Rubin AI Systems Hit $7.8M
[2] Data Center Knowledge: Scaling the Memory Wall: HBM, CXL, and the New GPU Playbook
[3] Data Center Dynamics: Musk: SpaceX/xAI Is Actively Seeking More AI Compute Customers Following Anthropic Deal
[4] Data Center Knowledge: SpaceX IPO Filing Recasts Company as AI Infrastructure Giant
[5] Data Center Knowledge: Texas Powers Past Virginia in Global Data Center Rankings
[6] Data Center Knowledge: AI Inference Pulls Infrastructure Back Into Metro Data Centers
[7] Data Center Knowledge: Nvidia Earnings Show AI Spending Moving Beyond GPUs
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.