- **The goal is to maximize absolute compute ASAP, which means accessing huge amounts of energy ASAP.** While all else equal they’d prefer cheap and clean, they first and foremost care about scale and speed.
- **Datacenters for inference vs training have different energy requirements.** Inference requires proximity to end-users and a high degree of reliability and redundancy in energy supply. Training datacenters are more geographically flexible and can tolerate less redundancy, though high GPU costs still penalize any significant downtime.
- **The main focus in this paper is training datacenters** since the requirements are less stringent and allow for flexibility in locating compute in areas with high solar potential and cheap land. Estimates vary, but training datacenters likely represent about half of new capacity expected by 2030.
- **Our understanding of key requirements and rough prioritization is summarized below**. Of course, they may differ in type or priority by user.
|**Criteria**|**Priority**|**Requirement**|
|---|---|---|
|Scale|P0|**~30-300 GW total by 2030, ~50% training/~50% inference.** Triangulating from both hyperscaler interviews and:<br><br>- [Jesse Jenkins](https://www.datacenterknowledge.com/ai-data-centers/ai-s-power-needs-not-as-bad-as-feared-princeton-professor-says) (Princeton): ~20 GW<br>- [Goldman Sachs](https://www.goldmansachs.com/insights/articles/AI-poised-to-drive-160-increase-in-power-demand): ~30 GW<br>- [International Energy Agency](https://www.iea.org/reports/world-energy-outlook-2024): ~40 GW (global growth, not US specific)<br>- [McKinsey](https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/ai-power-expanding-data-center-capacity-to-meet-growing-demand): ~50 GW<br>- [Situational Awareness](https://situational-awareness.ai/racing-to-the-trillion-dollar-cluster/): ~400 GW is the upper end of anything we’ve seen<br><br>**Minimum site size: 100 MW capacity on contiguous land.**<br><br>**Minimum cluster: 500 MW total capacity within 10 miles.** This could be composed of several smaller sites in the same geographic cluster or one very large site.|
|Speed|P0|**ASAP.** Speed is relative to whatever the best alternative is. The perceived fastest option right now is colocated natural gas when grid connections/capacity are not readily available.|
|Reliability|P1|**Approximately 125% of capacity** (optimize total cost where you assume downtime translates 1:1 into chips)|
|Cost|P1|**Generally aim for <$100 per MWh.** Energy costs are only a portion of total datacenter costs (generally less than 50%), and energy costs have been rising in recent years as demand has increased. A good reference point is the average price for industrial electricity in the US—[a bit over $80/MWh](https://www.eia.gov/electricity/data/browser/#/topic/7?agg=0,1&geo=g&endsec=vg&linechart=~~~ELEC.PRICE.US-IND.A~~&columnchart=ELEC.PRICE.US-ALL.A~ELEC.PRICE.US-RES.A~ELEC.PRICE.US-COM.A~ELEC.PRICE.US-IND.A&map=ELEC.PRICE.US-ALL.A&freq=A&ctype=linechart<ype=pin&rtype=s&maptype=0&rse=0&pin=).|
|Risk|P1|**Time risk and cost risk.** In evaluating options, buyers will consider outcome predictability vs the next best alternative.|
|Effort/complexity|P1|**AI is complicated enough as is, adding on additional complexity is not desirable.** That said, many of these companies already have teams that can deal with additional effort/complexity so long as they can still be confident the total time to power is at least as good as the next best alternative.|
|Emissions intensity|P2|**Nice to have, but won’t sacrifice speed or scale for it**. If they can do both at the same time, great, but everything is secondary to speed. Still, most major tech companies have strong climate commitments, and there is some risk to the long-term social licence of AI if emissions concerns remain unaddressed.|