I Squared pays $225 million for Cogent inference capacity
Infrastructure capital is moving from training to inference as the hyperscalers run climate PR in parallel.

I Squared Capital bought $225 million of data center assets from Cogent Communications, Reuters reports. The deal targets inference workloads specifically. That matters because inference is where the margin compression happens at scale—every token served costs money, and the unit economics tighten as throughput climbs.
The hyperscalers are moving in the other direction publicly. Microsoft, Google, Amazon and Meta announced a joint initiative with Elemental Impact, a nonprofit investor, to pilot advanced cooling, energy storage and low-carbon building materials in their data centers. Axios frames it as an attempt to use the AI buildout as a proving ground for decarbonization tech rather than a new emissions source.
The climate optics and the inference capital are not contradictory. The hyperscalers are spending record capex on training infrastructure while their served-token volumes climb faster than per-token costs fall. I Squared is betting that someone has to own the racks where those tokens actually run, and that the hyperscalers will lease rather than build all of it.
Inference infrastructure trades at different multiples than training clusters. Training capacity depreciates as the next architecture ships. Inference depreciates as the model weights get cheaper to serve. I Squared is pricing in a world where inference demand grows faster than per-watt efficiency gains, at least through 2026.
The Elemental partnership reads as alignment theater—joint press, nonprofit branding, no binding commitments. The I Squared deal is a $225 million call on inference volume. Watch which one the hyperscalers fund with actual capex next quarter.
Sources · 2
Tech giants back new data center climate initiative
Axios Business
I Squared bets on AI inference with $225 million data center buy from Cogent - Reuters
Reuters Business
Matched signals
Lattice signals Numen pinned to this story at publish time.
Unlock the analytical widgets on every article — signal matches, Trends snapshots, X overlays, agent reasoning — with a Member account.
Upgrade →Search interest · 30 days
Google Trends snapshot captured at publish time.
Search interest for “AI inference data centers”
0% · 30d
Snapshot · captured 5/27/2026· Google Trends · scaled 0–100 to peak in window.
Unlock the analytical widgets on every article — signal matches, Trends snapshots, X overlays, agent reasoning — with a Member account.
Upgrade →On X right now
Top engagement posts about this topic, ranked by likes + retweets + quotes.
HealthRanger @HealthRanger
621 eng45dVideo clip: Aaron Day joins me to talk about what's really behind AI data centers, given that AI inference fees from customers cannot even begin to cover the data center expansion costs... https://t.co/O3Bpx02rtp
View on X →ekram @ekrahm
44 eng45dI don’t think people fully realize how big of a deal Darkbloom is. It turns every Mac in the world into part of a new type of AI compute grid. Anyone can plug in their idle compute. makes for dramatically cheaper inference than relying solely on centralized data centers. https://t.co/Po7C3iCHTu
View on X →Gimlet Labs @gimletlabs
9 eng44dGimlet has been named to the @Redpoint InfraRed 100, which honors companies building at the cutting edge of infrastructure. As data centers transform to meet the immense needs of agentic inference, it is an incredibly exciting time to be building in this space. Check out our https://t.co/Dv8JbelHYC
View on X →syed mohib @syedmohib3
1 eng44dData centers every week "Just add more GPUs" 😣 Just published my latest AI article research paper covering LLM optimization, efficient AI systems, and scalable inference. Medium: https://t.co/SEzYLZXBhQ OSF: https://t.co/LQNhIJV7BR #AI #LLM #MachineLearning #KV #OpenSourceAI
View on X →GigaChadRizzGod @GigaChadRizzGod
1 eng44dStarcloud just dropped 50+ Starlink Mini Laser terminals on 25+ satellites. Orbital AI nodes running full inference & training on constant solar + vacuum cooling + light-speed mesh. Earth data centers about to look like flip phones in 2026. SpaceX isn’t just building https://t.co/m3ZmwGWzFD
View on X →
Unlock the analytical widgets on every article — signal matches, Trends snapshots, X overlays, agent reasoning — with a Member account.
Upgrade →Your read
How did this article land?
Three sliders. Optional comment. Anonymous is fine.
Open to anyone. One response per reader.