OpenAI and Anthropic reportedly seek 20-30MW data center deals, likely for inference — distributed demand alongside giant training sites

Reported (unverified) 1 sources 2026-09-23 (weekly, ±3-day window) /en/ai/signal/capex-dc-2026-09-23-43

🔺 Triangulation (claims vs. facts)

Facts (verified)

CNBC has reported that OpenAI and Anthropic are negotiating data center deals in the 20-30MW range. This signal is rated as reported and unverified; what can be stated as fact is that the report exists, and neither company has confirmed it within the scope of this signal.

Announcements / observations

According to the report, these deals are separate from large training clusters and are thought to be for inference. OpenAI is said to be negotiating in the US and the Nordics, Anthropic in the UK and the Nordics. Some read it as a move to add mid-sized sites closer to demand on top of gigawatt-scale training campuses.

Unverified / reserved

Three points remain unverified: (1) whether talks have produced signed contracts, and with which providers; (2) whether use is limited to inference, since "likely for inference" is the report's own inference; (3) total megawatts and go-live timing by site. None is backed by official disclosure from either company.

▲ Primary sources (official IR / press / expert)

Press

xAI's Michael Roberson joins OpenAI infrastructure team as commissioning program manager

Data Center Dynamics ・ 2026-10-09

xAI's Michael Roberson joins OpenAI infrastructure team

Source ↗ Press

OpenAI and Anthropic seek 20-30MW data center deployments

Data Center Dynamics ・ 2026-09-23

OpenAI and Anthropic seek 20-30MW data centers for inference

Source ↗

Primary sources aggregated by structural_signals(066). Each item links out to its original source.

✎ Analysis

The read as a leading indicator is that AI labs' data center demand is splitting into two tiers: gigawatt-scale training campuses on one side, and many mid-sized sites on the other. By hyperscale standards 20-30MW is small, but if inference demand grows close to users, spreading sites of this size across regions makes sense for both power access and latency. It is a signal of a change in the shape of demand that big-campus headlines tend to hide.

The fork is whether (a) these sites are bridge capacity secured until large campuses come online, or (b) inference takes a permanently larger share and mid-sized, distributed sites become the main line of demand. Under (b), the effect shows up first in utilization and rents for colocation providers and in power-rich regions such as the Nordics.

What to watch next: (1) announcements of signed contracts by either lab or by colocation providers; (2) new mid-sized data center leasing and rents in the Nordics and the UK; (3) shifts in how both labs talk about inference cost and inference capacity.

🔗 Related stories

Inference› OpenAI›

← Leading Indicators Archive