# Where AI lives

Author: Daniel Concepcion
Published: 2026-08-22
Canonical: https://danielconcepcion.com/thinking/where-ai-lives/

> AI rack-scale density changes the facility around the compute and turns AI infrastructure into an industrial, capital and sovereignty decision.

Trying to understand where AI “lives” today, I looked properly at what is being built for AI datacentres.

In 2000, a rack in the datacentres I worked with drew about 1.5 kilowatts, while its servers connected at 1 gb/s. Over the next two decades, those figures climbed to 25 kilowatts per rack and 25 gb/s server connections. Each step felt enormous at the time.

Then came the architecture built for AI:

• One Vera Rubin NVL72 reference design is rated at up to 227 kilowatts, about 10 times the densest thing I ever installed. At an 85% load, one rack would use about as much electricity per year as 480 Spanish homes.

• A single GPU is provisioned with 1.6 terabits per second through two 800 gb/s ports. Espanix reported around 1.3 terabits per second of average daily traffic across Spain’s national internet exchange in late 2025. The provisioned capacity of one GPU is already in the same range as the average traffic crossing a national IX.

NVIDIA designed the 72 GPUs in the rack to operate as one rack-scale accelerator. At that density, the facility has to be planned from the silicon outwards. Power, liquid cooling, network, storage and even the choice of site arrange themselves around the compute.

That is when the word gigafactory started to make sense to me. Not our standard datacentres any more.

The European Union wants to establish up to seven of them, each with more than 100,000 advanced AI processors. At that scale, AI infrastructure becomes industrial and sovereignty policy. At the same time, NVIDIA’s agreements with six investment firms aim to establish full-stack AI infrastructure as an investable asset class for global capital.

For enterprise IT, this splits the architecture map. Centralised gigafactories handle massive model training, while inference lands in-house, in a neocloud or with the current hyperscaler, based on latency, privacy, cost and sovereignty.

Where does your company draw that line?

#AIInfrastructure #DataCenters #DigitalSovereignty

## Sources and links

- [NVIDIA Vera Rubin NVL72](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/)
- [European Commission: AI Gigafactories](https://commission.europa.eu/topics/competitiveness/competitiveness-coordination-tool-projects/ai-gigafactories_en)
- [NVIDIA AI infrastructure financing platforms](https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital)


## Companies, products and research

- [NVIDIA](https://www.nvidia.com/)
- [EuroHPC](https://www.eurohpc-ju.europa.eu/)


## More Thinking

- Newer: [Different models, different roles](https://danielconcepcion.com/thinking/different-models-different-roles/)
- Older: [What changes when the executor is AI?](https://danielconcepcion.com/thinking/when-the-executor-is-ai/)