# Single-data-center training caps out in two years; distributed training is the only way out

- **Speaker:** Greg Osuri
- **Attribution:** attributed
- **Said:** 2025-01-31
- **Theme:** Energy & AI — https://www.gregosuri.com/predictions/energy-ai/
- **Source:** DeepSeek and the Dominance of Open Source AI \| Akash & Prime Intellect: Mined with CoinFund Ep. 19 (CoinFund)
- **Watch at:** 00:19:12 — https://www.youtube.com/watch?v=Eb-XhauNLNE&t=1152s
- **Canonical:** https://www.gregosuri.com/predictions/energy-ai/2025-01-31-single-data-center-training-caps-out-in-two-years-vcy7/
- **Cite as:** Greg Osuri, "Single-data-center training caps out in two years; distributed training is the only way out," DeepSeek and the Dominance of Open Source AI \| Akash & Prime Intellect: Mined with CoinFund Ep. 19 (CoinFund), 31 Jan 2025, 00:19:12. https://www.youtube.com/watch?v=Eb-XhauNLNE&t=1152s

## Quote — verbatim

> You look at what technologies can enable model training advancement, and if you look around, only decentralization — distributed training — is the only way out. That's why, in two years I think we're going to cap out as to how much you can do in a single data center.

## Context — site annotation, not spoken

After citing the Memphis 250MW data center, planned 1.2GW sites, ~96 US nuclear reactors already at full use, the last reactor taking 14 years/$34B, and the ~8-10K H100 per-data-center ceiling in the US.