Why immersion cooling is no longer optional for AI training clusters

NVIDIA H100 GPUs draw 700W. B200 modules draw over 1,000W. A single AI training rack now exceeds 100kW — eight times what a traditional data center was built to dissipate. We walk through the thermodynamics, the operating economics, and the design choices that made dielectric immersion the only viable path for serious AI compute.

The full body of this article is part of the editorial pipeline. We publish on a fortnightly cadence, with each piece authored by the named leader, edited internally, and reviewed against our editorial standards before publication. To request advance access to a draft, or to discuss a specific topic with the author, contact our communications team.

The themes covered in this piece — the technical, economic, and operational realities of building hyperscale AI infrastructure in the Kingdom — are subjects we cover continuously across the Insights stream. For the most current thinking, browse the full Insights index, or subscribe via the contact form to receive new pieces directly.