[rdl]

Almost all of the larger commercial miners (especially public companies) are looking at this. There are a bunch of issues (I have a conference talk last week at one of the big mining conferences on the topic, and have met with a bunch of miners on this.)

Probably long term the solution will be hybrid — mining gets done using any spare power. AI training generally requires protected power even beyond firm power and few miners have this for their mining operations, and also most of the mining facilities aren’t in the ISO 5 or 6 facilities we (I insure miners and AI) want to see for $500mm worth of mining hardware. Usually the mining companies don’t want to make the upfront capital outlay for these, so a lot of the time they do partnerships where part of their PPA is shared.

[hastamelo]

aren't the new huge ai data centers being built without power redundancy/backups, because speed to market is critical?

https://www.mckinsey.com/industries/technology-media-and-tel...

[rdl]

No one I've seen does without at least battery backup/conditioning/etc. Some may have less than 100% redundancy on generators, but that's mainly because some of them are running with e.g. a 200MW grid feed and 20x10MW gensets and ALSO running about 10-15 of the "backup" gensets as prime power. It's possible corners are cut during commissioning (like they use modular structures during buildout of the tilt-up ISO 5 permanent structure), but substanially no one would do this as their long-term design goal.

[mhb]

People have spent $500M for hardware to mine crypto?

[sshine]

More. In 2024, Bitcoin mining companies collectively spent over $3.6 billion on hardware, data centers, and infrastructure to maintain and expand their operations.

https://arb365.net/en/articles/majnery-vlozhili-36-mlrd-v-in...

[rdl]

I insure like >$3B worth actively I think (which includes some additional power equipment, hvac, etc. below the scale of main site transformers, etc.). If you count former equipment which has been removed from service over the years it is probably about $5-10B for Bitcoin in total? (Not pulling the BDX to look at specific breakdown of miners vs AI)

[trollbridge]

Couldn't foundational model training (one of the most expensive aspects of AI) be done with spare power?

[rdl]

One $4mm cabinet of GB300 can generate maybe 10-20x the power/hvac cost per hour so it doesn't make sense to ever have it idle. Vs. $100K worth of miners for the same 40KW conditioned or 100KW raw power.

[l1n]

Was the talk recorded? Couldn't find a link.