Anthropic, OpenAI quietly chase smaller AI data centre deals: report

Anthropic and OpenAI are expanding their AI infrastructure procurement beyond multi-hundred-megawatt and gigawatt-scale data-center agreements, pursuing smaller 20-30MW deployments. The shift highlights continued urgency to secure compute capacity and could broaden demand opportunities for data-center developers, power providers and AI infrastructure suppliers.

Analysis

The investable read-through is not incremental AI demand per se, but a shift toward modular deployment that raises the value of speed-to-power and repeatable electrical/mechanical integration. VRT, ETN and POWL should capture this through more frequent equipment orders and service content; smaller projects can shorten procurement cycles and smooth revenue recognition versus a handful of campus-scale awards. The benefit is strongest over the next 1-3 quarters if book-to-bill and backlog conversion accelerate without a corresponding increase in inventory or working-capital drag.

Powered-shell owners with dispersed, immediately usable capacity—DLR, CORZ and APLD—could gain leasing velocity because capacity need not be assembled in a single ultra-large location. This could improve utilization and pricing for previously subscale sites, although the economic value depends on contract duration, tenant-funded capex, and the ability to secure utility interconnection; headline megawatts without these terms are not equivalent to contracted cash flow. GE Vernova and regional utilities are a more delayed, 6-18 month beneficiary, since distributed loads can broaden substation and gas-generation demand beyond a small number of hyperscale markets.

The contrarian interpretation is that fragmented procurement may reflect difficulty bringing large campuses online rather than an upward revision to aggregate compute demand. If smaller deployments substitute for, rather than supplement, large commitments, AI infrastructure suppliers could see lower average order values and weaker operating leverage despite healthy unit volumes. The thesis is falsified if VRT/ETN report slowing backlog growth, lower data-center order conversion, or if DLR/CORZ disclose materially shorter lease terms or weaker returns on incremental capacity.

Key Decisions for Investors

• Initiate a 1-3 month long VRT / short SPY pair, sized for an 8% pair-loss limit and roughly 16% upside target. VRT has the clearest exposure to rapid-fit-out intensity; exit if its next reported backlog growth decelerates materially or management signals delayed customer acceptance.
• Accumulate ETN on broad technology-risk-off weakness over the next 4-8 weeks rather than chasing AI beta. Its power-distribution exposure offers a lower-valuation, less single-customer-sensitive route to the same deployment cycle; reassess if Electrical Americas margins or order backlog fall sequentially.
• Place DLR and CORZ on an earnings watch list rather than entering solely on this signal. Upgrade to long positions only if management quantifies incremental contracted MW, multi-year lease duration, and tenant-funded buildout economics; absent those disclosures, smaller-site demand is not yet a reliable FFO catalyst.
• Avoid treating the development as a blanket bullish signal for merchant data-center developers such as APLD. Require evidence of signed leases, financing availability, and delivered power before underwriting value, since smaller deployments may reduce the need for large speculative campuses.