Anthropic and OpenAI hunt for smaller data center deals, sources tell CNBC, in race to deploy AI capacity

Anthropic and OpenAI are pursuing smaller 20-30MW AI data-center capacity agreements in the U.K., Nordics and potentially the U.S., supplementing their multihundred-megawatt and gigawatt-scale commitments. The shift favors faster deployment of inference workloads as land, power availability, construction delays and community opposition constrain large data-center projects. JLL expects inference capacity to rise from 9% of global data-center workloads in 2025 to 37% by 2030, versus training falling from 14% to 13%; neocloud provider Crusoe also raised $3.9B at a $30.9B post-money valuation.

Analysis

The important change is not aggregate GPU demand but the buyer’s procurement mix: faster-to-energize, geographically distributed capacity shifts bargaining power toward operators with immediately available power, interconnection and modular deployment capability. EQIX and VRT should have higher incremental exposure than greenfield-heavy developers, because each additional site requires duplicated electrical, cooling, networking and deployment services. This is a 6-18 month utilization and pricing-power theme, while any near-term benefit depends on disclosed bookings rather than reported discussions.

For NVDA, distributed inference is modestly positive for unit demand but less clearly positive for revenue per workload than centralized training. Inference customers optimize cost per token and latency, creating a larger opening for AMD, custom ASICs and lower-spec GPU configurations once software portability improves; the market should not mechanically capitalize every incremental MW at training-like economics. A more fragmented customer footprint may also reduce neocloud concentration risk, but it weakens the scarcity premium of any one provider’s large dedicated campus.

The contrarian read is that smaller deployments may signal schedule arbitrage rather than an incremental demand inflection: capacity can be secured before applications and monetization are proven. If AI-service revenue growth fails to sustain utilization, short-tenor distributed leases become a margin problem for private neoclouds and equipment orders can be deferred rapidly. JLL may see advisory activity, but the likely fee pool is too small relative to its diversified earnings base to support a standalone position.

Key Decisions for Investors

• Initiate a 3-6 month relative long EQIX / short DLR position only if EQIX discloses improving AI-related cabinet or power bookings; target 10-15% relative upside. Thesis is that existing interconnected capacity earns a premium versus larger, slower wholesale delivery. Exit if EQIX utilization or pricing guidance weakens, or if DLR demonstrates equivalent rapid powered-capacity absorption.
• Accumulate VRT on 8-12% pullbacks over the next 1-3 months, sized modestly given valuation risk. Modular power and thermal equipment should capture duplicated site-level capex; seek a 15-20% upside over 6-12 months, with the thesis invalidated by backlog conversion slowing or a material cut to organic growth guidance.
• Maintain NVDA exposure but do not add solely on this development; use any AI-infrastructure rally to fund a small NVDA/AMD relative-value watch position rather than a directional inference bet. Reassess after next earnings for inference revenue disclosure, gross-margin trajectory and evidence that customers are adopting lower-cost accelerators or ASICs.
• Avoid treating private-neocloud funding and announced capacity discussions as confirmation of public-equity demand. Set an alert for disclosed lease tenor, customer prepayment and utilization metrics from listed data-center operators; shorter commitments or elevated cancellation provisions would favor reducing data-center REIT exposure before 2027 inference demand is proven.