Frontier AI labs are writing equity checks to memory chipmakers, SoftBank is building a 5GW campus in France, and the real bottleneck on AI infrastructure deployment is licensed electricians. Something structural is shifting in how serious compute clients procure capacity.
What Happened
Three developments this week reframe the GPU procurement conversation.
First, Anthropic's $65 billion raise at a near-$1 trillion valuation included equity positions in Micron, Samsung, and SK Hynix. Taking stakes in HBM (High-Bandwidth Memory, the memory architecture that determines GPU throughput) suppliers is not a financial play. It is a supply-chain lock. Simultaneously, Anthropic hired a Meta infrastructure veteran to lead its data center energy team, even as it holds major contracts with all three hyperscalers (AWS, Azure, and GCP). The signal: even the most cloud-native frontier labs are building toward owned infrastructure, because cloud contracts alone cannot guarantee the capacity or cost structure they need at scale.
Second, SoftBank announced plans for a campus up to 5GW in capacity at the Port of Dunkirk, France, representing a potential €75 billion investment, partnering with Schneider Electric. This is one of the largest single AI infrastructure commitments Europe has seen. It is happening in France, not in the historically dominant Nordic markets, and that geography is not accidental.
Why not Scandinavia? Because Denmark has halted new large-load grid agreements as AI demand floods its grid capacity. More broadly, permitting bottlenecks and grid access scarcity are concentrating European capacity growth in a shrinking number of viable jurisdictions. The European data center map is being redrawn in real time. Operators who locked in positions in Dunkirk, Frankfurt, Amsterdam, and Madrid are holding advantages that cannot be replicated quickly.
Third, and most counterintuitive: The Next Platform identifies licensed electricians, not GPUs, as the true bottleneck in the AI buildout. GPU allocation is hard. Energized rack space, fully permitted and staffed, is harder. Delivery timelines on operational capacity are slipping not because the chips are unavailable but because the trades required to commission the facilities cannot be hired fast enough.
Why It Matters
The pattern across these three stories is the same: the constraint has moved upstream of the GPU. Power, permitting, grid access, labor, and now chip-level memory supply are all gating factors that raw compute spending cannot simply buy through.
This is precisely where the hyperscaler (the largest cloud providers, AWS, Azure, GCP, Oracle) vs. neocloud (specialized GPU cloud providers, an alternative to hyperscalers) dynamic matters most. Hyperscaler wait lists for H200 and B200 capacity stretch quarters, sometimes into the following fiscal year. Neocloud operators we work with are often running on timelines measured in weeks, because they have already absorbed the upstream constraints. They hold pre-negotiated power agreements, finished data hall space, and allocated GPU inventory. That pre-positioned capacity is what clients are actually paying for when they sign a reserved instance contract.
The cost differential remains real. Neocloud operators typically price reserved GPU capacity 30 to 50 percent below equivalent hyperscaler rates. For a sovereign AI program (a government or quasi-government initiative building national AI infrastructure) or a Fortune 500 financial services firm trying to move inference workloads off public cloud, that spread compounds materially over a multi-year commitment. And unlike hyperscalers, neocloud operators often offer more flexible MSA (Master Service Agreement, the parent contract) terms and shorter minimum commitment windows.
The colocation (colo) layer adds a third option. Operators like Equinix, Digital Realty, QTS, Aligned, and CyrusOne hold finished Tier III (99.982% uptime) raised-floor and liquid-cooled capacity in markets including Northern Virginia (NoVa, the largest US data center market), Dallas, Phoenix, Chicago, and across Western Europe. For enterprises that want to own or lease their own GPU hardware and house it in professionally managed facilities, colo is the path. Given the electrician bottleneck, finished colo space with live power is worth a significant premium over greenfield commitments.
What Clients Should Do
If you are a frontier lab or large AI application company planning a 4,000-plus GPU training cluster in the next two quarters, start the neocloud conversation now. The operators with B200 and GB200 inventory in the US and EU are allocating that capacity to clients already in dialogue. Anthropic's decision to simultaneously hold hyperscaler contracts and build toward owned infrastructure is the right structural model. Most clients cannot replicate the owned-infra play, but they can replicate the portfolio logic: hyperscaler contracts for flexibility, neocloud reserved capacity for cost and speed, colo for hardware ownership.
If you are a Fortune 500 enterprise in financial services, pharma, or manufacturing deploying AI infrastructure for the first time, the $6 billion AWS commitment Snowflake made is a useful reference point. Long-term hyperscaler deals at that scale confirm enterprise AI is baseload demand, not burst. But before you sign a multi-year AWS or Azure reserved instance at list price, benchmark it against neocloud alternatives. The pricing gap is often 30 to 50 percent, and the capacity ramp time is frequently shorter.
If you are a system integrator or consultancy sourcing for an end client in the EU, the European geography story demands attention. Denmark's grid rationing and the broader permitting crunch mean that capacity positioned in France, Germany, and Iberia is worth more than it was 12 months ago. Build that constraint into your client's site selection analysis before the shortlist is set.
XIRR Advisors brokers reserved GPU capacity from neocloud operators across the US and EU, and Tier III colocation space in major US markets. We do not broker hyperscalers. AWS, Azure, GCP, and Oracle sell direct. Our value is in the neocloud and colo markets, where terms are negotiable, inventory is not publicly listed, and earlier conversations consistently produce better pricing and faster deployment timelines. Our fee is paid by the provider. Clients pay nothing.
Share your requirements with us: GPU type (H100, H200, B200, GB200, GB300), cluster size, region, timing, and any colocation megawatt needs. We will canvas the market and come back with a shortlist within 48 hours. Reach us at contact@xirradvisors.com or DM @XIRRAdvisors. The clients getting the best capacity terms this quarter started the conversation last month.
References
[1] Data Center Dynamics: Anthropic Raises $65B, Nears $1T Valuation with Memory Chipmaker Stakes
[2] Data Center Dynamics: Meta Infrastructure Veteran Joins Anthropic Data Center Energy Team
[3] Data Center Dynamics: SoftBank Plans 5GW French Data Center Campus, €75B Investment
[4] Data Center Knowledge: Denmark Halts Large-Load Grid Agreements Amid AI Demand Surge
[5] Data Center Knowledge: Power and Permitting Redraw Europe's Data Center Geography
[6] The Next Platform: Skilled Electricians, Not GPUs, Identified as True AI Buildout Bottleneck
[7] Data Center Knowledge: Snowflake's $6B AWS Commitment Signals Persistent Enterprise AI Compute Demand
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.