# @stem_antics on Instagram

- **Type:** Video
- **Original URL:** https://www.instagram.com/p/DbYZIe7gbfK
- **Gondola URL:** https://gondola.cc/posts/68532556-stem-antics-instagram
- **Thumbnail:** https://img.gondola.cc/tr:w-,h-,fo-auto/postThumbnails/1b85a6663f.jpg
- **Posted:** 2026-07-29T15:04:53.000+00:00
- **Account Owner:** Stem Antics (@stem_antics) — https://gondola.cc/stem_antics

## Caption

Your phone draws roughly one watt to play a video. Hold onto that number. It is the anchor for everything below.

That single watt is the product of decades of ruthless optimisation. The display backlight and the fixed-function video decode block do almost all the work. Phone makers built dedicated silicon for H.265 and AV1 precisely because a general-purpose processor would burn ten times the energy for the same frames. Specialise, then power-gate everything you are not using.

Now scale up.

- Laptop under load: 45 to 90 W
- Single Nvidia H100 accelerator: 700 W
- Single Blackwell B200: up to 1,200 W
- One GB200 NVL72 rack: around 120 kW

That last figure is the fracture point. A conventional enterprise rack drew 7 to 12 kW. AI racks now pull ten times that in the same floor footprint, and heat flux at the die reaches 500 W per square centimetre. Air cooling collapses somewhere near 30 to 40 kW per rack. Removing 120 kW by forced air would demand roughly 85,000 cubic feet per minute through a single cabinet. Direct-to-chip liquid cooling is no longer an upgrade. It is the only physics that works.

Multiply by thousands of racks and the unit of measurement changes again. AI campuses are now specified in gigawatts, not megawatts. One gigawatt is the output of a large nuclear reactor. It is also a billion phones playing video at once.

Then add the overhead. Power usage effectiveness compares total facility draw to IT draw. At a PUE of 1.1, ten percent of every watt disappears into cooling, conversion and distribution before a single tensor operation happens.

Globally, data centres consumed about 415 TWh in 2024, near 1.5 percent of world electricity. The IEA projects 945 TWh by 2030, slightly more than Japan uses today.

The constraint has moved. It is not silicon anymore. It is the grid.

#stemantics #datacenter #electricalengineering #thermalmanagement #artificialintelligence

## Stats

- **Views:** 6,477
- **Likes:** 155
- **Shares:** 0
- **Comments:** 2

## Tags

thermalmanagement, electricalengineering, datacenter, artificialintelligence, stemantics

---
Copyright (c) Gondola