← Back to list

There Is No Room for Air Anymore

NVIDIA’s 140 kW AI racks reveal a new reality: the future of artificial intelligence may be constrained less by algorithms than by heat…

Elijah Wickberg · 2026-06-12 13:11 · 72 claps · 2.9 min read
#semiconductors #technology #nvidia #infrastructure-as-code
Open on Medium ↗
Wiki topics: AI · AI · General 💻 · Programming

There Is No Room for Air Anymore

NVIDIA’s 140 kW AI racks reveal a new reality: the future of artificial intelligence may be constrained less by algorithms than by heat, power, and infrastructure.

There Is No Room for Air Anymore

AI seems to be getting faster and more compact. But the newest racks show something different: the hardest limit is no longer in the code, but in heat, cabling, and access to electricity.

The figure of 140 kW per rack looks like a technical footnote for engineers. But small lines like this often reveal where AI is heading better than big presentations do. Because 140 kW is no longer a “server rack” in the old sense. It is a load that forces the building itself to behave differently.

A data center used to be easy to imagine: rows of equipment, air conditioners, cables under the floor. Now one AI rack is starting to look like a small industrial node. NVIDIA Vera Rubin NVL72 combines 72 Rubin GPUs and 36 Vera CPUs in one rack-scale design, while NVLink 6 delivers up to 260 TB/s of rack-level bandwidth. NVIDIA has also described Rubin moving into full production for AI factories.

And this raises an uncomfortable question: if artificial intelligence is so “digital,” why are there more pipes, pumps, cables, and metal around it? Why does progress in chips so quickly become a cooling problem? Maybe because the real limit of AI is not where the model ends, but where the rack can no longer dump its heat into the air.

The most interesting part of this story is not whether a specific Rubin configuration consumes 140 kW, 180 kW, or more. Some estimates for Vera Rubin NVL72 already point to a higher range, and infrastructure notes mention roughly 190–230 kW for VR200-level systems. This is not really an argument about one number. It is a direction: AI hardware is quickly moving beyond the zone where older data centers could simply “add another rack.”

https://www.moduledge.com/blog/nvidia-vera-rubin

https://www.moduledge.com/blog/nvidia-vera-rubin

What is more expensive - buying new GPUs, or finding a place where they can actually be powered properly? For large buyers, this is no longer a rhetorical question. They need transformers, substations, copper busbars, cable routes, liquid cooling, backup power, water or another heat-transfer system. Supermicro, for example, has shown a Vera Rubin NVL72 solution with a new coolant, claiming 1,000x higher electrical impedance compared with standard mixtures. When equipment costs millions of dollars per rack, even a coolant leak becomes a financial event.

https://www.tomshardware.com/desktops/servers/supermicro-shows-off-vera-rubin-nvl72-rack-with-all-new-type-of-coolant-company-claims-coolant-offers-1-000-times-higher-electrical-impedance-over-standard-cooling

https://www.tomshardware.com/desktops/servers/supermicro-shows-off-vera-rubin-nvl72-rack-with-all-new-type-of-coolant-company-claims-coolant-offers-1-000-times-higher-electrical-impedance-over-standard-cooling

This is where copper stops being a background metal. It is not in the headline of the presentation, but it is everywhere in the body of the system: power delivery, load distribution, transformers, cables, interconnects, grid connections. If an AI rack asks for 100+ kW today and much more tomorrow, the question is no longer only who builds the better chip. The question is who can build the physical chain underneath that density fast enough.

One junior example in this early layer is NRED. Not as a ready-made supplier for data centers and not as the hero of this story, but as a small fragment of the broader copper map. At Wilmac, the company is working with early geological signals, while through MetalCore it has outlined an AI-assisted mineral evaluation approach, where geological, geochemical, and geophysical datasets are meant to be read not manually in isolation, but as a probability system. This is not a cable and not a substation yet. But before every cable, there has to be a material base somewhere.

What struck me in this news was not “140 kW.” It was the fact that even such a number already sounds almost like a warning. Air, which for years served as the invisible buffer of the server room, is no longer enough. Next come liquid, copper, land, permits, grid connections, and the waiting line for power.

AI can respond in milliseconds. But before that happens, someone has to find it a place where heat is not the first thing to say “enough.”


메타데이터
post_id
635f80aed863
slug
there-is-no-room-for-air-anymore-635f80aed863
url
https://medium.com/@elijah.wickberg/there-is-no-room-for-air-anymore-635f80aed863
canonical_url
https://medium.com/@elijah.wickberg/there-is-no-room-for-air-anymore-635f80aed863
author_url
https://medium.com/@elijah.wickberg
status
ok
fetched_at
2026-06-13 07:35:29