The GPU capacity market is bifurcating: hyperscalers are locking up power and silicon at a scale that benefits their own roadmaps first, and everyone else is learning to source differently.
What Happened
Three stories this week illustrate the pressure building across the stack.
First, per Tom's Hardware Pro, SpaceX rented 220,000 Nvidia GPUs and 300MW of compute capacity to Anthropic. Read that again: a frontier AI lab went outside the hyperscaler ecosystem entirely, leasing from a third-party operator, to get the scale it needed. This is not a workaround. This is a signal about where capacity actually lives in 2026.
Second, Microsoft is reportedly considering scrapping its 24/7 renewable energy matching target while simultaneously committing to doubling its AI infrastructure within two years. The contradiction is instructive. Azure is growing so fast that its own sustainability commitments are becoming a casualty. The implication for clients on Azure wait lists: that capacity is being absorbed internally and by the largest enterprise contracts first.
Third, AEP's contracted power capacity has surged to 63GW, with 90% tied to data centers, and Denmark's grid operator just imposed a three-month moratorium on new connections driven by data center demand. Power is no longer a background variable in GPU procurement. It is the primary constraint in both US and EU markets.
Why It Matters
The structural pattern here is straightforward once you see it: hyperscalers (the largest cloud providers, AWS, Azure, GCP, Oracle) are consuming a disproportionate share of available power, land, and GPU supply to feed their own product roadmaps. Microsoft's doubling commitment and AWS's shift toward an OEM hardware model signal these companies are optimizing supply chains for themselves. Third-party clients sit downstream.
Meanwhile, the demand signal keeps intensifying. AMD posted $10.3B in data center revenue with 57% growth, confirming that inference workload expansion is broadening the compute supplier landscape. OpenAI is pushing a new open networking protocol to address fabric bottlenecks (the high-speed interconnect fabric that links GPUs within a cluster) at hundred-thousand-GPU scale. The era of off-the-shelf cluster deployment is over. Infrastructure decisions now have architectural consequences.
For sovereign AI programs in the EU, the Danish moratorium is a warning shot. Nordic power, long considered abundant and green, is no longer automatically accessible. Programs that assumed they could site facilities in Scandinavia on 12-month timelines need to revisit those assumptions now.
For Fortune 500 enterprises entering AI infrastructure for the first time, the temptation is to default to AWS or Azure. That is the path of least resistance, and also the path of longest lead times, least pricing flexibility, and weakest SLA (Service Level Agreement, defining uptime guarantees) negotiation leverage. Neocloud operators, specialized GPU cloud providers built around H100, H200, and B200 capacity, are routinely 30-50% cheaper on reserved instances and can deploy in weeks rather than quarters.
The IREN acquisition of Mirantis for $625M is worth noting here. GPU-as-a-Service operators are racing to solve utilization, not just raw capacity, bundling Kubernetes orchestration with GPU infrastructure. The neocloud operators we work with are on similar trajectories: the product is no longer just bare-metal GPUs, it is managed, orchestrated compute that can absorb production inference workloads with real SLA commitments.
What Clients Should Do
If you are a frontier lab planning a 10,000-GPU-plus training cluster, the SpaceX-Anthropic deal is your reference point. Large-scale third-party leasing is viable and, in many cases, faster to access than hyperscaler reserved capacity. Start sourcing neocloud alternatives in parallel with any hyperscaler conversations. The operators with H200 and B200 availability are not advertising it publicly. That is what brokers are for.
If you are a Fortune 500 enterprise standing up AI infrastructure for the first time, do not let your default cloud vendor be your only option. A portfolio approach, anchoring on a hyperscaler for general workloads while reserving GPU-intensive training and inference capacity through a neocloud operator, consistently produces better unit economics and faster time to production. Pair that with Tier III colocation (data center reliability tier, targeting 99.982% uptime) in markets like Northern Virginia, Dallas, or Phoenix through operators such as Equinix, Digital Realty, or QTS, and you have a stack you control.
If you are a sovereign AI program in the EU navigating the Danish moratorium or similar grid constraints, the window for securing power-backed colocation in Western European markets is narrowing. Frankfurt, Amsterdam, and Dublin are tightening. Conversations that happen now, before the next wave of hyperscaler reservation cycles, will access better terms than conversations happening in Q4.
Earlier engagement always wins. The clients who locked in H100 reservations in mid-2024 are the ones running production workloads today. The pattern will repeat with B200 and GB200.
---
XIRR Advisors sources reserved GPU capacity from neocloud operators and Tier III colocation space across the USA on behalf of clients. We do not broker hyperscalers. AWS, Azure, GCP, and Oracle sell direct. Where we add value is in the neocloud and colocation markets, where pricing, terms, and availability are negotiated, not listed.
Share your requirements: region, GPU type (H100, H200, B200, GB200, GB300), cluster size, timing, and for colocation, your MW target. We will canvas the market and return a shortlist within 48 hours. Our fee is paid by the provider. Clients pay nothing. Email contact@xirradvisors.com or DM @XIRRAdvisors.
References
[1] Data Center Dynamics: Danish grid operator introduces three-month moratorium for new grid connections
[2] Data Center Dynamics: Microsoft considering scrapping 24/7 renewable energy matching target
[3] Data Center Dynamics: AEP contracted capacity surges, 90% data center-tied
[4] The Next Platform: Microsoft committed to doubling AI infrastructure in two years
[5] Tom's Hardware Pro: SpaceX rents 220,000 Nvidia GPUs and 300MW to Anthropic
[6] Data Center Knowledge: AMD data center revenue surges 57% on EPYC and Instinct demand
[7] Data Center Knowledge: IREN acquires Mirantis for $625M to monetize deployed GPU capacity
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.